UNIVERSIDAD POLITÉCNICA DE MADRID

ESCUELA TÉCNICA SUPERIOR DE INGENIERÍA AGRONÓMICA,

ALIMENTARIA Y DE BIOSISTEMAS

Characterization of the climate-vegetation-soil

complexity using vegetation indices

Tesis Doctoral

Por
Andrés Felipe Almeida-Ñauñay
Ingeniero Agrónomo

Noviembre 2022
 Departamento de Matemática Aplicada
Escuela Técnica Superior de Ingeniería Agronómica Alimentaria y de
Biosistemas
Universidad Politécnica de Madrid

Tesis Doctoral

Characterization of the climate-vegetation-soil

complexity using vegetation indices

Por
Andrés Felipe Almeida-Ñauñay
Ingeniero Agrónomo

Directora:
Ana María Tarquis Alfonso
Dra. Ingeniero Agrónomo (Universidad Politécnica de Madrid).

Codirector:
Miguel Quemada Sáenz-Badillos
Dr. Ingeniero Agrónomo (Universidad Politécnica de Madrid).

Madrid, Noviembre 2022

Acknowledgements

I would like to use these first lines to thank all my family, my mother (Segunda), my
father (Luis) and my brothers (Sebastián and Edison) and my pet (Lucero †) who have
helped me so much in the achievement of this work. Without their help it would have
been impossible to achieve this important goal. In addition, I would like to thank all the
people in Ecuador, especially my paternal grandparents, Luz and Rafael †, my maternal
grandparents, María † and Segundo †, the Rodríguez family and Graciela who have been
watching my progress, offering me advice when I needed it.

I would like to dedicate the following lines to my thesis supervisor, Prof. Ana María
Tarquis Alfonso. She is undoubtedly the key person who has made it possible for me to
complete this work. All the time, ideas, and advice she has given me are invaluable and
have allowed me to develop both, personally and professionally.

I would like to thank my thesis co-director, Prof. Miguel Quemada Sáenz-Badillos. His
involvement and support have been flawless. I am very grateful for all the time he has
dedicated to help me improve the quality of my work, which has made everything easier.

I would also like to thank all the people who are part of the Complex Systems research
group. Especially Rosa Benito and Juan Carlos Losada who have always had time to help
me. Special mention to my colleagues: Andrea, David and Ernesto who have made
everything easier and more enjoyable.

In another way. I would also like to thank my colleagues from the AgSystems research
group, Jose Luis and Maria Dolores, who have helped me throughout my PhD whenever
I have needed it, and who I have had the pleasure of collaborating with during my last
year.

I would also like to thank all the research and management staff at the Centre for
Agricultural and Environmental Risk Studies and Research (CEIGRAM). To the
management team (Esperanza, Katerina and Hamid) and technical support (Begoña) who
have facilitated my work and provided the necessary equipment to carry out my tasks.

Finally, I would like to thank all my colleagues inside and outside CEIGRAM, with
special mention to Daniele, Maritza, Afsané, Mabel, Diana, Luz, Jesús and María who
have helped me to make this stage enjoyable.

i
ii
iii
Agradecimientos

Me gustaría aprovechar estas primeras líneas para agradecer a toda mi familia, a mi madre
(Segunda), a mi padre (Luis) y a mis hermanos (Sebastián y Edison) y a mi mascota
(Lucero †) que tanto me han ayudado en la consecución de este trabajo. Sin su ayuda
hubiera sido imposible alcanzar esta meta tan importante. Además, me gustaría agradecer
a todas las personas de Ecuador, especialmente a mis abuelos paternos Luz y Rafael †, a
mis abuelos maternos, María† y Segundo†, a la familia Rodriguez y Graciela que han
estado pendientes de mi progreso, ofreciéndome consejos cuando los he necesitado.

Las siguientes líneas me gustaría dedicarlas a mi directora de tesis, la Prof. Ana María
Tarquis Alfonso. Ella es sin duda, la persona clave por la que he conseguido terminar este
trabajo. Todo el tiempo, ideas y consejos que me ha aportado son muy valiosos y me han
permitido desarrollarme tanto personal, como profesionalmente.

Me gustaría agradecer a mi codirector de tesis, el Prof. Miguel Quemada Sáenz-Badillos.

Su implicación y ayuda han sido intachables. Agradezco mucho todo el tiempo que ha
dedicado para ayudarme a mejorar la calidad de mi trabajo, y que ha permitido que todo
sea más sencillo.

Agradecer también a todas las personas que forman parte del grupo de investigación de
Sistemas Complejos. En especial a Rosa Benito y Juan Carlos Losada que siempre han
tenido tiempo para ayudarme. Mención especial a mis compañeros: Andrea, David y
Ernesto que han hecho que todo sea más fácil y ameno.

Por otro lado, también quiero agradecer a mis compañeros del grupo de investigación
AgSystems, José Luis y María Dolores, que a lo largo de todo doctorado me han ayudado
siempre que lo he necesitado, y que además he tenido el placer de colaborar con ellos mi
último año.

Agradezco también a todo el personal investigador y de gestión que forman parte del
Centro de estudios e investigación de riesgos agrarios y medioambientales (CEIGRAM).
En especial al equipo de gestión (Esperanza, Katerina y Hamid), soporte técnico (Begoña)
que han facilitado mi labor y han dispuesto el equipo necesario para realizar mis tareas.

Por último, me gustaría agradecer a todos mis compañeros de dentro y fuera del
CEIGRAM, con mención especial a Daniele, Maritza, Afsané, Mabel, Diana, Luz, Jesús
y María que han ayudado a que esta etapa haya sido entretenida.

iv
v
Summary

Multiple studies describe vegetation as a spatial-time-varying complex ecological system,

varying in time and space, with non-linear relationships between the different variables.
Therefore, developing and implementing methodologies that allow us to broaden our
knowledge of the system is essential. Throughout this thesis, the complexity of the
climate-vegetation-soil system is analysed from two points of view. First is the temporal
component through the time series analysis in natural grasslands. Secondly, the spatial
complexity is analysed by studying a high intra-variability system created by different
doses of Nitrogen in wheat crops (Triticum aestivum L.).
Satellites have shown the potential for monitoring vegetation and climate dynamics
across larger areas. At the same time, unmanned aerial vehicles (UAV) are more
frequently used for monitoring several crops to improve the management, knowing the
variability that many fields present. For these purposes, vegetation indices (VIs) are used
to obtain information through the relationship between the measurements of vegetation
reflectance at different bandwidths of the electromagnetic spectrum.
The most critical drivers that regulate vegetation growth are precipitation and
temperature. The excess or deficit of these key factors involves high potential risks, such
as drought, a significant limitation in the vegetation growth in semiarid Mediterranean
areas. The VIs can be combined with meteorological indices to characterise climate-
vegetation-soil scenarios better.
This work is divided into four main chapters. The first chapter's main objective was to
reveal the relationship between agricultural and meteorological drought. Two
Mediterranean grasslands vulnerable to drought were selected. Then, several correlations
were calculated between climate and agricultural drought indices. The former was
evaluated by the Standardised Precipitation Index (SPI) and Standardised Precipitation-
Evapotranspiration Index (SPEI)). The latter was represented by Vegetation Health Index
(VHI) and Standardised VHI (SVHI).
Our results revealed that SPEI was better correlated with VHI compared to SPI. In
addition, SVHI obtained better results in the critical vegetation phases than VHI. Overall,
SPEI and SVHI were the best-correlated indices. The quarterly scale showed stronger
relationships than the monthly scale; the most correlated time frames were Mediterranean
spring and autumn.

vi
The main objective of the second chapter was to assess the response of VIs to the temporal
dynamics of temperature and precipitation in two semiarid Mediterranean grasslands. The
Normalized Difference Vegetation Index (NDVI) and the Modified Soil Adjusted
Vegetation Index (MSAVI), which included a soil factor adjustment, were calculated
from the MODIS product (MOD09A1), presenting a spatial resolution of 500 x 500 m.
Correlations and cross-correlations between VIs and each climatic variable were
computed. Different lagged responses of each VIs series were detected, varying in zones,
phenological seasons, and climatic variables. These differences indicated that a linear
correlation was insufficient to characterise their relationships.
Recurrence and Cross Recurrence Plots (RPs and CRPs) were applied to reveal the
characterization of such complex systems. Furthermore, the recurrence quantification
analysis (RQA) measured the system's complexity. RPs pointed out that short-term
predictability and high dimensionality of VIs series, as well as precipitation, characterised
this dynamic. Meanwhile, temperature showed a more regular pattern and lower
dimensionality. Comparing both VIs, the MSAVI showed a more evident pattern than
NDVI. CRPs revealed that precipitation was a critical variable in distinguishing between
zones due to their complex pattern and influence on the soil’s water balance that the VI
replicates.
The third chapter's main objective was to visualize the dynamic response of the series
anomalies of the NDVI and MSAVI and climate series. For this purpose, two natural
grassland areas with different agro-climatic characteristics were selected in central Spain,
and recurrence techniques were employed on images of a higher spatial resolution
(250x250 m).
RPs showed different VIs responses in both areas. Once more, the MSAVI patterns
showed a clear structure, while NDVI presented a noisy pattern showing a higher
stochastic behaviour. When CRPs were applied over anomalies series, we could observe
that precipitation is more synchronised with VIs anomalies, most likely because of the
influence of soil moisture in vegetation growth. Overall, our results suggest that the VI
time series show a more evident pattern and differentiate both sites better when a variable
soil adjustment factor is included.
In the fourth part of the work, the main objective was to assess the influence of the bare
soil in high-resolution images obtained by a multispectral sensor on board an unmanned
aerial vehicle (UAV) and whether decreasing its effect may contribute to a better
assessment of plant traits.

vii
The experimental area of winter wheat crop, located at Aranjuez (Madrid, Spain), was
divided into four sectors with 133 plots. Different N doses were applied to the plots to
create high experimental intra-variability. Each plot measured yield, protein content, and
nitrogen (N) output at harvest. An UAV was flown at growth stages (GS) – GS32, GS39,
and GS65 – to ensure accurate crop monitoring. Four VIs, NDVI, MSAVI, Normalized
Difference Red Edge Index (NDRE) and Blue-Red Index (BRI1), were selected based on
the sensor spectral information.
A sequential cutting method, named threshold value optimization (TVO), was
implemented to remove soil background pixels. Based on a percentile approach, threshold
values were computed from the VIs distributions. Then, the predictive performance of the
VIs in each segmentation was evaluated.
The results indicated that NDVI, MSAVI, and NDRE could predict wheat traits.
However, the optimal thresholds were not constant and ranged from 0.1 to 0.3 depending
on the VI and the wheat trait. The TVO improved yield and N output estimation at the
stem elongation growth stage (GS32). On the other hand, the TVO achieved a limited
improvement in estimating protein content at anthesis (GS65).
In summary, this thesis has demonstrated that remote sensing is a vital tool to help
characterise the high complexity of agroecological systems. The merging of this tool with
non-linear and optimisation methods improves the usefulness of vegetation and
precipitation indices.

viii
ix
Resumen

Múltiples estudios describen la vegetación como un sistema ecológico complejo, que

varía en el tiempo y en el espacio, y que presentan relaciones no lineares entre las distintas
variables que lo conforman. Por tanto, es fundamental desarrollar y aplicar nuevas
metodologías que permitan ampliar nuestro conocimiento del sistema. A lo largo de este
trabajo, la complejidad del sistema clima-vegetación suelo se analiza desde dos puntos de
vista. En primer lugar, la componente temporal, a través del análisis de series temporales
en pastizales naturales. En segundo lugar, se analiza la complejidad espacial, a través del
estudio de un sistema de alta intravariabilidad, como es el caso de un cultivo de alto
rendimiento con distintas dosis de fertilización nitrogenada (Triticum aestivum L.).
Los satélites han demostrado un gran potencial para monitorizar la dinámica del clima y
la vegetación en extensas áreas. Al mismo tiempo, los vehículos aéreos no tripulados
(UAV) son frecuentemente utilizados para ayudar en la monitorización y el manejo de
cultivos en los que la variabilidad es muy alta. Por este motivo, los índices de vegetación
(IVs) se utilizan para obtener información mediante la relación entre la reflectancia
medidas en la vegetación a distintas longitudes de onda del espectro electromagnético.
Por lo que sabemos, las variables climáticas regulan el crecimiento de la vegetación,
siendo la precipitación y la temperatura los factores más importantes. El exceso o déficit
de estos factores, posibilita riesgos importantes, como es el caso de la sequía, elemento
clave en el desarrollo de la vegetación en climas mediterráneos semiáridos. Los índices
de vegetación combinados con índices climáticos permite caracterizar adecuadamente el
sistema clima-vegetación-suelo.
Este trabajo se divide en cuatro capítulos principales. El objetivo del primer capítulo, era
hallar la relación entre sequía agrícola y sequía meteorológica. Para el análisis, se
seleccionaron dos zonas de pastos mediterráneos vulnerables a la sequía. Después, se
evaluó la correlación entre los índices de sequía meteorológica: Índice estandarizado de
precipitación (SPI) e índice estandarizado de la precipitación-evapotranspiración (SPEI);
y los índices de sequía agrícola; Índice de la salud de la vegetación (VHI) e Índice
estandarizado de la salud de la vegetación (SVHI).
Nuestros resultados revelaron que el SPEI se correlacionó mejor con el VHI en
comparación con el SPI. Además, el SVHI presentó mejores resultados durante las fases
críticas de desarrollo de vegetación, comparado con el VHI. En general, SPEI y SVHI
fueron los índices mejor correlacionados y la escala que mostró mejores resultados fue la

x
trimestral donde las fases más correlacionadas fueron la primavera y el otoño
mediterráneos.
El objetivo principal del segundo capítulo del trabajo era evaluar la respuesta de los IVs
a la dinámica temporal de la temperatura y la precipitación de dos zonas de pastos
mediterráneas. Para ello, se calcularon el Índice de Vegetación de Diferencia
Normalizada (NDVI) y el Índice de Vegetación Ajustado al Suelo (MSAVI), el cual
incluye un factor de corrección del suelo a partir del producto (MOD09A1) de la
colección de imágenes MODIS con una resolución de 500 x 500m. Como paso preliminar,
se realizó un análisis de correlaciones y correlaciones cruzadas entre los distintos IVs con
cada una de las variables climáticas. Las diferencias indicaban que un análisis de
correlación linear es insuficiente para caracterizar estas relaciones.
Posteriormente, se caracterizó y cuantificó la complejidad del sistema, para ello se
utilizaron los mapas recurrencia (RP), recurrencia cruzada (CRP) y el análisis de
cuantificación de la recurrencia (RQA). Los RPs mostraron una predictibilidad a corto
plazo y una alta dimensionalidad de las series de IVs, siendo la precipitación, el factor
clave de esta dinámica. Mientras, la temperatura mostró un patrón más regular y de una
menor dimensionalidad. Comparando ambos índices, el MSAVI mostró un patrón mas
claro que el NDVI. Por otro lado, los CRPs revelaron que la precipitación es una variable
crítica para distinguir entre zonas, debido a su patrón complejo y a su influencia en el
balance hídrico del suelo, cuyo comportamiento reflejan los IVs.
En el tercer capítulo, el objetivo principal es la visualización de la respuesta dinámica de
las series de anomalías de NDVI, MSAVI, y clima. En este caso, se seleccionaron dos
zonas de pastizal natural con diferentes características agroclimáticas en el centro de
España y se utilizaron técnicas de recurrencia en imágenes de mayor resolución espacial
250x250 m.
Los RPs mostraron que la respuesta de los IVs era distinta en cada una de las zonas. El
MSAVI presentaba una estructura clara, mientras que el NDVI mostraba un patrón
ruidoso, asociado a un comportamiento más estocástico. En los CRPs de las series de
anomalías, se pudo observar que la precipitación estaba más sincronizada con las
anomalías de IVs, que las series originales, probablemente debido a la influencia de la
humedad del suelo. En general, los resultados sugieren que las series de IVs ajustadas con
un factor de corrección del suelo muestran un patrón más evidente y diferencian mejor
entre ambos sitios.

xi
En la cuarta parte del trabajo, el objetivo principal era verificar la influencia de los pixeles
de suelo en las imágenes de alta resolución (UAVs) y comprobar si su disminución
contribuye a mejorar la evaluación de las características productivas de la vegetación.
El experimento de trigo, estaba situado en Aranjuez (Madrid, España). El área se dividió
en cuatro sectores con 133 parcelas en total y se aplicaron distintas dosis de N a las
parcelas para crear una alta intravariabilidad espacial. Durante la cosecha, en cada una de
las parcelas se midió el rendimiento, el contenido en proteína y el nitrógeno total en grano.
EL UAV capturó las imágenes en distintas etapas de crecimiento del cultivo (GS) - GS32,
GS39, y GS65 – para asegurar una correcta monitorización del cultivo. A partir de la
información espectral se seleccionaron cuatro IVs: NDVI, MSAVI, el Índice de
Diferencia Normalizada de Borde Rojo (NDRE) y el Índice Rojo-Azul (BRI1).
Se desarrolló un método de corte secuencial basado en el cálculo de valores umbral a
partir de los percentiles de las distribuciones de cada uno de los IVs. Este método se
denominó optimización del valor umbral (TVO) y su objetivo principal es eliminar los
píxeles que se corresponden al suelo. Tras aplicar la metodología a las imágenes de alta
resolución, se evaluó el rendimiento predictivo de los IVs en cada una de las
segmentaciones.
Nuestros resultados sugieren que NDVI, MSAVI y NDRE son una herramienta adecuada
para analizar las características productivas del trigo. Sin embargo, los valores umbrales
óptimos no eran constantes y variaban de 0,1 a 0,3 dependiendo del VI y de la
característica productiva del trigo a analizar. El TVO mejoró la estimación del
rendimiento y el contenido de N en la fase de elongación del tallo (GS32). Por otro lado,
el TVO no logró una mejora considerable en la estimación del contenido de proteína en
la antesis (GS65).
En resumen, esta tesis ha demostrado que las herramientas de teledetección son
escenciales para caracterizar la gran complejidad de los sistemas agroecológicos. Y que,
junto con las metodologías no lineales y los métodos de optimización, es posible mejorar
la utilidad de los índices de vegetación y precipitación.

xii
xiii
INDEX

1 INTRODUCTION .................................................................................................... 1

1.1 Research context ................................................................................................ 2

1.2 Climate-vegetation-soil system .......................................................................... 2

1.2.1 Climate factors and activity ........................................................................ 2

1.2.2 Vegetation as a complex system ................................................................. 2

1.2.3 Soil characterization ................................................................................... 3

1.3 Climate-vegetation-soil monitoring. .................................................................. 4

1.3.1 Precipitation indices ................................................................................... 5

1.3.2 Vegetation indices ...................................................................................... 5

1.3.3 Bare soil role in vegetation monitoring ...................................................... 7

1.4 Climate-Vegetation-Soil complexity ................................................................. 7

1.4.1 Temporal complexity.................................................................................. 8

1.4.2 Spatial complexity .................................................................................... 10

1.5 Research objectives .......................................................................................... 11

1.6 Related publications ......................................................................................... 12

2 MATERIALS AND METHODOLOGY ............................................................... 15

2.1 Complexity of vegetation and drought indices at different time Scales.

Standardization and drought episodes. ....................................................................... 16

2.1.1 Area of study ............................................................................................ 16

2.1.2 Satellite imagery and vegetation indices calculations .............................. 19

2.1.3 Climate data and precipitation indices...................................................... 20

2.1.4 Standarised indices ................................................................................... 21

2.1.5 Correlation, classification and quantification analysis of drought Episodes

22

2.2 The Vegetation–Climate system complexity ................................................... 23

2.2.1 Area of study ............................................................................................ 23

2.2.2 Satellite imagery and meteorological data................................................ 24

xiv
 2.2.3 Inter-Annual and Intra-Annual Analysis and time-Series and VIs Phase
Cross-Correlations .................................................................................................. 25

2.3 Vegetation and climate anomalies time series. ................................................ 27

2.3.1 Area of study, satellite data and vegetation index estimation .................. 27

2.3.2 Climate data .............................................................................................. 27

2.4 Recurrence Plots and Recurrence Quantification Analysis ............................. 28

2.4.1 Recurrence analysis for vegetation complexity ........................................ 28

2.4.2 Recurrence analysis for anomalies time series. ........................................ 34

2.5 Threshold values optimization (TVO) to reduce soil spatial variability in a wheat
(Triticum aestivum L.) crop with UAV imagery ........................................................ 34

2.5.1 Study area and experimental design ......................................................... 34

2.5.2 Wheat measurements, imagery acquisition and postprocessing ............... 36

2.5.3 Threshold value optimization (TVO) ....................................................... 38

3 ASSESMENT OF DROUGHT INDICES ON DIFFERENT TIME SCALES: A

CASE IN SEMIARID MEDITERRANEAN GRASSLANDS ...................................... 40

3.1 Objectives, area of study and materials ........................................................... 41

3.2 Descriptive Statistics........................................................................................ 42

3.3 Standardised Indices Time Series .................................................................... 44

3.4 Correlation Analysis ........................................................................................ 46

3.5 Drought Classification ..................................................................................... 47

3.6 Discussion ........................................................................................................ 48

4 THE VEGETATION–CLIMATE SISTEM COMPLEXITY THROUGH

RECURRENCE ANALYSIS ......................................................................................... 51

4.1 Objectives, area of study and materials ........................................................... 52

4.2 Soil Line Acquisition ....................................................................................... 53

4.3 Inter-Annual Analysis and intra-annual analysis ............................................. 53

4.4 Time-Series Correlation and correlation by phases ......................................... 59

4.5 Recurrence plots and recurrence diagonal profile characterisation ................. 63

xv
 4.6 Cross recurrence plots and recurrence diagonal profile characterisation ........ 66

4.7 Recurrence Quantification Analysis of RPs and CRPs ................................... 69

4.8 Discussion ........................................................................................................ 70

5 RECURRENCE PLOTS FOR QUANTIFYING THE VEGETATION INDICES

DYNAMICS IN A SEMIARID GRASSLAND ............................................................ 76

5.1 Objective, area of study and materials ............................................................. 77

5.2 Soil slope and soil adjustment factor ............................................................... 77

5.3 Descriptive statistics ........................................................................................ 80

5.4 Recurrence and Cross-recurrence analysis of original series .......................... 84

5.5 Recurrence and Cross-recurrence analysis of the anomaly series ................... 90

5.6 Discussion ........................................................................................................ 94

6 OPTIMIZATION OF SOIL BACKGROUND REMOVAL TO IMPROVE THE

PREDICTION OF WHEAT TRAITS WITH UAV IMAGERY ................................. 100

6.1 Objectives, study area and materials .............................................................. 101

6.2 Raw data and histogram analysis ................................................................... 102

6.3 TVO performance .......................................................................................... 103

6.4 Pre- and post-segmentation analysis .............................................................. 106

6.5 Sector analysis ............................................................................................... 108

6.6 Discussion ...................................................................................................... 109

7 CONCLUSIONS .................................................................................................. 113

7.1 Specific and general conclusions ................................................................... 114

7.2 Limitations and further research .................................................................... 117

7.2.1 Limitations .............................................................................................. 117

7.2.2 Further research ...................................................................................... 118

8 REFERENCES ..................................................................................................... 119

9 ANNEXES ........................................................................................................... 144

Original series and Anomalies series time series ...................................... 145

xvi
 Threshold values of each sequential segmentation between VIs and wheat
traits. 147

Recurrence plot (RP) and Cross Recurrence plots (CRP) parameters and
Recurrence Quantification Analysis (RQA) using general z-score NDVI series. .... 149

Median values of vegetation indices (Index Value) and average temperature

each 8 Days (TEMP) from 2002 to 2018, in each zone: ZGU = Guadalix de la Sierra,
ZSO =Soto del Real. A) NDVI and B) MSAVI. ...................................................... 150

Recurrence plots of NDVI and MSAVI for each zone (ZGU and ZSO).
Vegetation indices series are normalized by the z-score method. ............................ 151

xvii
List of tables

Table 1 Soil horizons and zone characteristics of both study areas. ZGU = Guadalix de
la Sierra, ZSO = Soto del Real ....................................................................................... 17
Table 2. Selected vegetation-drought indices ................................................................ 20
Table 3. Classification of meteorological and agricultural drought indicators using the
index value (IV). The threshold values of IV are adapted from (Ali et al., 2017; Gidey et
al., 2018b; Ye et al., 2015). ............................................................................................ 23
Table 4. Annual climate and soil characteristics of both study zones. ZAV: Tornadizos
de Ávila (Avila), ZMA: Soto Del Real (Madrid). .......................................................... 24
Table 5. Summary of the sectors. Number and size of plots, fertilization and irrigation
treatments. ...................................................................................................................... 36
Table 6. Pearson series correlation values between VHI (Standardised Vegetation Index)
and SVHI with SPI (Standardised Precipitation Index) and SPEI (Standardised
Precipitation-Evapotranspiration Index) at different time scales. M is the monthly time
scale; Q is the quarterly time scale. ................................................................................ 46
Table 7. Correlation values between SVHI (Standardised Vegetation Health Index), SPI
(Standardised Precipitation Index), and SPEI (Standardised Precipitation
Evapotranspiration at different time frames in both zones: ZGU (Guadalix de la Sierra)
and ZSO (Soto del Real). M is the monthly scale; Q is the quarterly scale. .................. 47
Table 8. Drought classified data frequency of drought index (SPI, SPEI), non-
standardised vegetation health index (VHI), and standardised vegetation health index
(SVHI) in both zones, Guadalix de la Sierra (ZGU) and Soto del Real (ZSO). M is the
monthly scale. Q is the quarterly scale. .......................................................................... 48
Table 9. Statistical trends significance of vegetation indices ((Normalised Difference
Vegetation Index—NDVI, modified soil-adjusted vegetation index—MSAVI)) and
climatic variables (TEMP: Temperature; PCP: Precipitation) in ZAV: Tornadizos de
Ávila and ZMA: Soto del Real. ...................................................................................... 54
Table 10. Chow’s test of annual pasture phases. ZAV: Tornadizos de Ávila, ZMA: Soto
Del Real. REG_1 (linear regression of the first phase data), REG_2 (linear regression of
the second phase data, REG_T (both data regression.) B0 is the intercept, and B1 is the
slope of the linear regression. ......................................................................................... 56
Table 11. Annual pasture phases based on vegetation index (VIs) trend at the studied
zones (Tornadizos de Ávila and Soto del Real). ............................................................ 56

xviii
Table 12 Time-series Pearson correlation coefficients (CR) and partial correlation
coefficients (PCR) between NDVI time-series and meteorological time-series for each
study zone (ZAV: Tornadizos de Ávila; ZMA: Soto del Real) during the period of 2002-
2018. TEMP is 8-day average air temperature (°C), and PCP is the accumulated
precipitation in 8-day (mm). ........................................................................................... 59
Table 13. Time-series Pearson correlation coefficients (CR) and partial correlation
coefficients (PCR) between MSAVI time series and meteorological time-series for each
study zone (ZAV: Tornadizos de Ávila; ZMA: Soto del Real) during the period of 2002–
2018. TEMP is 8-day average air temperature (°C), and PCP is the accumulated
precipitation in 8-day (mm). ........................................................................................... 59
Table 14 Pearson correlation coefficients (CR) and partial correlation coefficients (PCR)
between NDVI and meteorological parameters in the five phases. TEMP is 8-day average
air temperature (°C), and PCP is the accumulated precipitation in 8-day (mm). ZAV is
Tornadizos de Ávila, and ZMA is Soto del Real. ........................................................... 61
Table 15. Pearson correlation coefficients (CR) and partial correlation coefficients (PCR)
of MSAVI and meteorological parameters in five distinct phases. TEMP is 8-day average
air temperature (°C), and PCP is the accumulated precipitation in 8-day (mm). ZAV is
Tornadizos de Ávila, and ZMA is Soto del Real. ........................................................... 61
Table 16. Cross-correlation coefficients between NDVI and meteorological parameters
with different lags (ℓ) in different phases. ZAV: Tornadizos de Ávila and ZMA: Soto del
Real. TEMP is 8-day average air temperature (°C), and PCP is the accumulated
precipitation in 8-day (mm). Each time lag is of 8 days. The bold letter represents the
maximum correlation in each row. ................................................................................. 62
Table 17. Cross-correlation coefficients between MSAVI and meteorological parameters
with different lags (ℓ) in different phases. ZAV: Tornadizos de Ávila and ZMA: Soto del
Real. TEMP is 8-day average air temperature (°C), and PCP is the accumulated
precipitation in 8-day (mm). The bold number represents the maximum correlation in
each row. Each time lag is of 8 days. ............................................................................. 62
Table 18. Recurrence plot (RP) parameters in ZAV: Tornadizos de Ávila and ZMA: Soto
del Real. MSAVI: Modified Soil-Adjusted Vegetation Index, m: Embedding dimension,
τ: Delay, 𝜀: threshold. ..................................................................................................... 63
Table 19. Cross-Recurrence plot (RP) parameters in ZAV: Tornadizos de Ávila and
ZMA: Soto del Real. MSAVI-TEMP: Cross-Recurrence plot between MSAVI and

xix
Average temperature. MSAVI-PCP: Cross-Recurrence plot between MSAVI and
Accumulated precipitation.............................................................................................. 66
Table 20. Recurrence plot (RP) and Cross Recurrence plots (CRP) parameters and
Recurrence Quantification Analysis (RQA) using general z-score vegetation indices
series in ZAV: Tornadizos de Ávila and ZMA: Soto del Real. MSAVI: Modified Soil-
Adjusted Vegetation Index, m: Embedding dimension, τ: Delay, 𝑟: threshold, RR:
Recurrence rate, DET: Determinism, LT: Average length of diagonal structures, ENTR:
Shannon Entropy, LAM: Laminarity, TT: Trapping time. ............................................. 69
Table 21. Statistical trend significance between the annual variables studied and time in
each area (ZGU and ZSO): NDVI, MSAVI, Temperature (TEMP) and Precipitation
(PCP). ............................................................................................................................. 81
Table 22. Recurrence plot (RP) parameters and Recurrence Quantification Analysis
(RQA) of each study area using normalized VI (NDVI, MSAVI), minimum temperature
(TMIN) and accumulated precipitation (PCP). ZGU = Guadalix de la Sierra, ZSO = Soto
Del Real. NDVI = Normalized Difference Vegetation Index, MSAVI = Modified Soil-
Adjusted Vegetation Index, m= Embedding dimension, τ = Delay, 𝜀 = threshold, RR =
Recurrence rate. DET = Determinism, LT = Average length of diagonal structures, ENTR
= Shannon Entropy, LAM = Laminarity ........................................................................ 87
Table 23. Cross-Recurrence plot (CRP) parameters and Recurrence Quantification
Analysis (RQA) of VIs (NDVI and MSAVI), minimum temperature (TMIN) and
precipitation (PCP) in each study area. ZGU = Guadalix de la Sierra, ZSO = Soto Del
Real. NDVI = Normalized Difference Vegetation Index, MSAVI = Modified Soil-
Adjusted Vegetation Index, m = Embedding dimension, τ = Delay, 𝜀 = threshold, RR=
Recurrence rate, DET = Determinism, LT = Average length of diagonal structures, ENTR
= Shannon Entropy, LAM = Laminarity. ....................................................................... 90
Table 24. Recurrence plot (RP) parameters and Recurrence Quantification Analysis
(RQA) using VIs anomalies (A_NDVI, A_MSAVI), minimum temperature (A_TMIN)
and precipitation anomalies (A_PCP) in each study area. ZGU series = Guadalix de la
Sierra, ZSO = Soto Del Real. NDVI = Normalized Difference Vegetation Index, SAVI =
Soil-Adjusted Vegetation Index, MSAVI = Modified Soil-Adjusted Vegetation Index, m=
Embedding dimension, τ = Delay, 𝜀 = threshold, RR = Recurrence rate., DET =
Determinism, LT = Average length of diagonal structures, ENTR = Shannon Entropy,
LAM = Laminarity. ........................................................................................................ 92

xx
Table 25. Cross-Recurrence plot (CRP) parameters and Recurrence Quantification
Analysis (RQA) of VIs anomalies (A_NDVI, A_MSAVI). minimum temperature
anomalies (A_TMIN) and precipitation anomalies (A_PCP) in each study area. ZGU =
Guadalix de la Sierra, ZSO = Soto Del Real. NDVI = Normalized Difference Vegetation
Index, MSAVI = Modified Soil-Adjusted Vegetation Index, m= Embedding dimension,
τ = Delay, 𝜀 = threshold, RR = Recurrence rate, DET = Determinism, LT = Average
length of diagonal structures, ENTR = Shannon Entropy, LAM = Laminarity. ............ 94
Table 26. Coefficient of determination (R2) between vegetation indices (NDVI, MSAVI,
NDRE, and BRI1), protein content, and nitrogen output before (pre-segmentation) and
after (post-segmentation) the segmentation methodology. GS32, GS39, and GS65
represent the winter wheat growth stages. .................................................................... 107
Table 27. Root mean square error (RMSE) between vegetation indices (NDVI, MSAVI,
NDRE, and BRI1), protein content, and nitrogen output before (pre-thresholding) and
after (post-thresholding) the thresholding value optimization methodology. GS32, GS39,
and GS65 represent the winter wheat growth stages. ................................................... 107

xxi
List of figures

Figure 1. Site and climate station’s location, ZSO is on the left and ZGU is on the right.
Images were obtained from the Mosaic of current orthophotos corresponding to sheet
0509 of the MTN50 in Spain during the years 2017,2018. Pixel plot size: 250x250 m2.
Pixel orthophotographs size: 0.05 x 0.05 m2. RGB composition. Blue crosses are the
locations of the climate station locations and red pinpoints are the soil trench locations.
........................................................................................................................................ 17
Figure 2. Localisation of the study zones and the three selected plots in each one. Left:
ZAV, Tornadizos de Ávila, Ávila (ZAV_1, ZAV_2, and ZAV_3), Right: ZMA, Soto del
Real, Madrid (ZMA_1, ZMA_2 and ZMA_3). Mosaic of the most current
orthophotographs corresponding to sheet 0509 of the MTN50 in Spain, during the years
2017, 2018. Pixel size: 0.25 km. RGB composition. ..................................................... 23
Figure 3. Random series distance matrix (DM) and recurrence plot (RP) example. The
length of the series is of 25 points. Absolute distance is calculated and represented as a
coloured matrix (A). Different thresholds were applied to obtain RPs: 𝜀 < 3 (B) and 𝜀 <
1 (C). ............................................................................................................................... 30
Figure 4. Artificial series RPs adapted from (Zhao et al., 2015): A) White-noise series,
embedding dimension (m) = 1, delay (τ) = 0, recurrence rate (RR) = 5%, B) Sine series,
m = 2, τ = 6, RR = 5%. Both series have a length of 387 data points. ........................... 31
Figure 5. Location of the trial and sectors (S1, S2, S3, S4) over a normalized difference
vegetation index (NDVI) map retrieved from the UAV platform during stem elongation
in 2019. The 133 plots monitored in each flight are shown. Coordinate system: ETRS89
/ UTM zone 30N (EPSG: 25830) ................................................................................... 35
Figure 6. Vegetation Condition Index (VCI) (a) and Temperature Condition Index (TCI)
(b) statistics from 2001 to 2020, along the year in each zone: ZGU = Guadalix de la Sierra,
ZSO = Soto del Real. The Y-axis shows a box and whisker plot of VCI and TCI on each
date. The error bars indicate the maximum and minimum values; horizontal lines indicate
the median value, and the box indicates the 25% and 75% percentiles. Q1, Q2, Q3 and
Q4 are the quarters of the year. The blue and red line are the medians of each zone. ... 42
Figure 7. VHI vegetation index statistics, from 2001 to 2020, throughout the year in each
zone: ZGU = Guadalix de la Sierra, ZSO = Soto del Real. The Y-axis shows a box-and-
whisker plot of VHI (Vegetation Health Index) on each date. The error bars indicate the
maximum and minimum values; horizontal lines indicate the median values, and the box

xxii
indicates the 25% and 75% percentiles. Q1, Q2, Q3 and Q4 are the quarters of the year.
The blue and red lines are the medians of each zone. .................................................... 43
Figure 8. Accumulated 8-day precipitation (PCP) (a) and temperature (TEMP) (b)
statistics throughout the year, including from 2001 to 2020, in each zone: ZGU =
Guadalix de la Sierra, ZSO = Soto del Real. The Y-axis shows a box and whisker plot of
a) PCP_ZGU, PCP_ZSO, and b) TEMP_ZGU and TEMP_ZSO on each date. The error
bars indicate the maximum and minimum values, the horizontal lines indicate the median
values, and the box indicates the 25% and 75% percentiles. Q1, Q2, Q3 and Q4 are the
quarters of the year. ........................................................................................................ 44
Figure 9. SVHI (Standardised Vegetation Index), SPI (Standardised Precipitation Index),
and SPEI (Standardised Precipitation Evapotranspiration Index) time series at different
time scales in ZGU (Guadalix de la Sierra). The red dotted lines mark ±1 values. ....... 45
Figure 10. SVHI (Standardised Vegetation Index), SPI (Standardised Precipitation
Index), and SPEI (Standardised Precipitation Evapotranspiration Index) time series at
different time scales in ZSO (Soto del Real). The red dotted lines mark ±1 values. ..... 45
Figure 11. Correlation between meteorological drought indices (SPI and SPEI) and
standardised vegetation index (SVHI) at different time scales (M: Monthly, Q: Quarterly)
in ZGU (Guadalix de la Sierra) and ZSO (Soto del Real). The alpha corresponds to the α
value in Table 2.............................................................................................................. 46
Figure 12. Soil slope (s) estimation from the spectral bands NIR vs. RED plots adapting
the minimum NIR method (Xu and Guo, 2013). Inside each figure, the regression
equation was obtained from the Weighted Least Squares method. Each black point
represents the NIR vs. RED reflectance in an 8-day interval between 2002 and 2018. Red
points are the minimum value of NIR at 0.005 RED breaks. Panels (a–c) correspond
respectively to the soil slope of pixels 1, 2 and 3 for the Tornadizos de Ávila (ZAV) zone.
Panels (d–f) correspond respectively to the soil slope of pixels for the Soto del Real
(ZMA) zone. ................................................................................................................... 53
Figure 13. Inter-annual variations of vegetation indices (Normalised Difference
Vegetation Index—NDVI, modified soil-adjusted vegetation index—MSAVI) and
meteorological parameters of water (Annual accumulated precipitation) and energy
(Annual average Temperature) from 2002 to 2018 in Tornadizos de Ávila (ZAV) (a–c)
and Soto del Real (ZMA) (d–f). ..................................................................................... 54
Figure 14. Relationship between the standardised NDVI and MSAVI with min-max
standardised climate parameters of energy (Annual average Temperature) and water

xxiii
(Annual accumulated precipitation) in Tornadizos de Ávila (ZAV) (a-b) and Soto del
Real (ZMA) (c-d) at a yearly scale from 2002 to 2018. ................................................. 55
Figure 15. Boxplots of vegetation indices and meteorological parameters of energy
(Average temperature) and water (Accumulated precipitation) in 8 days periods from
2002 to 2018 in Tornadizos de Ávila (ZAV) (a–c) and Soto del Real (ZMA) (d–f). .... 57
Figure 16 Growing phases response of NDVI to water (Accumulated precipitation) and
energy (Average temperature) parameters at 8 days in Tornadizos de Ávila (ZAV) (a) and
Soto del Real (ZMA) (b). In each figure, the regression equation was obtained from the
Least Squares method. .................................................................................................... 58
Figure 17. Growing phases response of MSAVI and water (Accumulated precipitation)
and energy (Average temperature) parameters, in 8 days periodsat Tornadizos de Ávila
(ZAV) (a) and Soto del Real (ZMA) (b). In each figure, the regression equation was
obtained from the Least Squares method. ...................................................................... 58
Figure 18 Time series cross-correlation between NDVI and climatic variables time series,
Temperature (TEMP) and accumulated precipitation (PCP) in Tornadizos de Ávila
(ZAV) and Soto del Real (ZMA). Each lag is of an 8-days period. The blue dot line is the
confidence interval at 95%. ............................................................................................ 60
Figure 19. Time series cross-correlation between MSAVI and climatic variables time
series, Temperature (TEMP) and Accumulated precipitation (PCP) in Tornadizos de
Ávila (ZAV) and Soto del Real (ZMA). Each lag is of an 8-days period; the blue dot line
is the confidence interval at 95%. ................................................................................... 60
Figure 20. Optimised recurrence plots (RP) using normalised vegetation indices
(MSAVI), Average temperature (TEMP) and Accumulated precipitation (PCP) data and
rescaled distance matrix for Tornadizos de Ávila (ZAV) and Soto del Real (ZMA). Time
units are represented as the X and Y-axis. Each time-unit is 8-days, coincident with 8-day
composed MODIS images during the study period (2002–2018). Panels (a–d) correspond
respectively to RP of NDVI, MSAVI, Temperature and Accumulated Precipitation for
the ZAV zone. Panels (e–h) correspond respectively to NDVI, MSAVI, Temperature and
Accumulated precipitation for the ZMA zone. ............................................................... 64
Figure 21. Diagonal-wise recurrence profile of the RPs obtained from the MSAVI,
Average temperature (TEMP) and accumulated precipitation (PCP) in Tornadizos de
Ávila (ZAV) (a–c) and Soto del Real (ZMA) (d–f). ...................................................... 65
Figure 22. Optimized Cross-Recurrence Plots (CRPs) and diagonal-wise recurrence
profiles between vegetation indices data (NDVI) temperature data (TEMP) and

xxiv
accumulated precipitation data (PCP) for Tornadizos de Ávila (ZAV) and Soto del Real
(ZMA). Time units are represented as the X and Y-axis. Each time-unit is 8-days,
coincident with 8-day composed MODIS images during the study period (2002-2018) in
the CRPS. Lags are represented in days in the diagonal-wise recurrence profile The panels
[a,c,e,g] represent the CRPs of NDVI-TEMP and NDVI-PCP. The panels [b,d,f,h]
represent the diagonal-wise profile of the CRPS, respectively. ..................................... 67
Figure 23. Optimised Cross-Recurrence Plots (CRPs) and diagonal-wise recurrence
profiles between vegetation index data (MSAVI) temperature data (TEMP) and
accumulated precipitation data (PCP) for Tornadizos de Ávila (ZAV) and Soto del Real
(ZMA). Time units are represented as the X and Y-axis. Each time-unit is 8-days,
coincident with 8-day composed MODIS images during the study period (2002–2018) in
the CRPS. Lags are represented in days in the diagonal-wise recurrence profile The panels
(a,c,e,g) represent the CRPs of MSAVI-TEMP and MSAVI-PCP. The panels (b,d,f,h)
represent the diagonal-wise profile of the CRPS, respectively ...................................... 68
Figure 24. Soil slope (s) estimation for each plot by the minimum NIR method (Xu &
Guo, 2013). Inside each figure, the regression equation obtained from weighted least
squares method (Xu & Eckstein, 1995). Each point represents the RED and NIR
reflectance in an 8-days interval between 2002 and 2018. ZGU = Guadalix de la Sierra,
ZSO = Soto del Real. ...................................................................................................... 79
Figure 25. Soil factor adjustment for MSAVI (𝐿𝑀) average and standard error throughout
the year at: Guadalix de la Sierra (L_MSAVI_ZGU) and Soto del Real and Madrid
(L_MSAVI_ZSO). X-axis represent the month and the date of each measure. ............. 80
Figure 26. Yearly average of the VIs (NDVI and MSAVI), temperature (TEMP) and
annual accumulated precipitation (Precipitation) from 2002 till 2008. .......................... 81
Figure 27 Vegetation index statistics, from 2002 to 2018, along the year in each zone:
ZGU = Guadalix de la Sierra, ZSO = Soto del Real. Y-axis shows a box and whisker plot
of A) NDVI (Normalized Difference Vegetation Index) and B) MSAVI (Modified Soil-
Adjusted Vegetation Index) in each date. Error bars indicate maximum and minimum
values; horizontal line indicate median value; and box indicates the 25% and 75%
percentiles. Red and blue solid lines represent the median of the time series ................ 82
Figure 28. Accumulated 8-day precipitation (PCP) statistics along the year, including
from 2002 to 2018, in each zone: ZGU = Guadalix de la Sierra, ZSO = Soto Del Real. Y-
axis shows a box and whisker plot of A) PCP_ZGU, PCP_ZSO and B) TEMP_ZGU and

xxv
TEMP_ZSO in each date. Error bars indicate maximum and minimum values; horizontal
line indicate median value; and box indicates the 25% and 75% percentiles. ................ 84
Figure 29. Optimized recurrence plots using normalized VIs (NDVI and MSAVI) and
climate (TMIN and PCP) data and rescaled distance matrix for both zones (ZGU and
ZSO). Time units are represented as the X and Y-axis. Each time-unit is a period of 8-
days, coincident with 8-day composed MODIS images during the study period (2002-
2018). .............................................................................................................................. 86
Figure 30. Optimized Cross-Recurrence Plots (CRPs) between VIs (NDVI and MSAVI),
minimum temperature (TMIN) and precipitation (PCP) for both zones (ZGU and ZSO).
Time units are represented as the X and Y axis. Each time-unit is a period of 8-days,
coincident with 8-day composed MODIS images during the study period (2002-2018).
........................................................................................................................................ 89
Figure 31. Optimized recurrence plots using VIs anomalies (A_NDVI, A_MSAVI) for
both zones (ZGU and ZSO), minimum temperature anomalies (A_TMIN) and
precipitation anomalies (A_PCP). Time units are represented as the X and Y-axis. Each
time-unit is a period of 8-days, coincident with 8-day composed MODIS images during
the study period (2002-2018). ........................................................................................ 91
Figure 32. Optimized Cross-Recurrence plots (CRPs) between VIs anomalies (A_NDVI,
A_MSAVI), minimum temperature anomalies (A_MIN) and precipitation anomalies
(A_PCP) for both zones (ZGU and ZSO). Time units are represented as the X and Y-axis.
Each time-unit is a period of 8-days, coincident with 8-day composed MODIS images
during the study period (2002-2018). ............................................................................. 93
Figure 33. Trial raw data exploration. The x axis represents the wheat yield (kg ha-1) and
the y axis represents the wheat protein content (g kg-1). Each color represents a sector
from Figure 1. ............................................................................................................... 102
Figure 34. Histograms of vegetation indices. The y axis represents the percentage of
pixels in GS32 (11 March). The letters represent each of the vegetation indices used in
the study: (A) NDVI, (B) MSAVI, (C) NDRE, and (D) BRI1. ................................... 103
Figure 35. Coefficient of determination (R2) and root mean square error (RMSE)
obtained from linear regression between vegetation indices and protein content in each
segmentation at GS65: A (NDVI), B (MSAVI), C (NDRE), D (BRI1). Box labels are the
optimum VI threshold value. ........................................................................................ 104

xxvi
Figure 36. Application of TVO to the image. A) RGB image acquired at GS39. B)
Binarised NDVI image (Threshold value: 0.23). Black dots are soil. No nitrogen fertilizer
was applied to N0, 50 kg N/ha to N1, and 90 kg N/ha to N2 treatments ..................... 104
Figure 37. Coefficient of determination (R2) and root mean square error (RMSE)
obtained from linear regression between vegetation indices and protein content in each
segmentation at GS65: A (NDVI), B (MSAVI), C (NDRE), D (BRI1). Box labels are the
optimum VI threshold value. ........................................................................................ 105
Figure 38. Coefficient of determination (R2) and root mean square error (RMSE)
obtained from linear regression between vegetation indices and nitrogen output in each
segmentation at GS39: A (NDVI), B (MSAVI), C (NDRE), D (BRI1). Box labels are the
optimum VI threshold value. ........................................................................................ 106
Figure 39 NDVI with yield (A, B, C), MSAVI with protein content (D, E, F), and NDRE
with nitrogen output yield (G, H, I) linear regressions. Each column represents a different
GS (first column, GS32; second column, GS39; third column, GS65). Each color
represents the different sectors in the trial. ................................................................... 108
Figure 40. NDVI with yield (A, B, C), MSAVI with protein content (D, E, F), and NDRE
with nitrogen output yield (G, H, I) optimised linear regressions. Each column represents
a different GS (first column, GS32; second column, GS39; third column, GS65). Each
color represents the different sectors in the trial........................................................... 109

xxvii
List of abbreviations
ZSO: Soto del real’s Zone
ZGU: Guadalix de la Sierra’s Zone
IV: Vegetation index
VHI: Vegetation Health Index
SVHI: Standardized vegetation Health index
SPI: Standardized Precipitation Index
SPEI: Standardized Precipitation-Evapotranspiration Index
ZAV: Tornadizos de Ávila’s Zone
ZMA: Soto Del Real’s Zone
NDVI: Normalized Difference Vegetation Index
MSAVI: Modified Soil-Adjusted Vegetation Index
RP: Recurrence plot
CRP: Cross- Recurrence Plots
RQA: Recurrence Quantification Analysis
m: Embedding dimension
τ: Delay
𝜺: Threshold
ℓ: Time Lag
RR: recurrence rate
DET: Determinism
LT: Average length of diagonal structures
ENTR: Shannon’s Entropy
LAM: Laminarity
UAV: Unmanned aerial vehicle
NDRE: Normalized Difference Red-Edge Index
BRI1: Blue-Red Index
TVO: Threshold values optimization
Topt: Threshold Optimal value
GS: Growth stages

xxviii
1 INTRODUCTION

1
1.1 Research context
Ecological systems tend to present non-linear dynamics, combining chaotic and periodic
cycles, whose equations governing the systems are unknown (Proulx and Parrott, 2009;
Storch and Gaston, 2004). As a rule, ecosystems are an excellent example of complex
dynamical processes continuously subjected to and updated by non-linear feedforward
and feedback inputs (Martínez and Gilabert, 2009; Zhong et al., 2010). These complex
interactions of the climate-plant-soil system occur at various hierarchical levels and
involve several interacting variables (Zhao et al., 2015), many of which are unavailable
for experimental measurement.
1.2 Climate-vegetation-soil system
1.2.1 Climate factors and activity
Multiple research studies (Li et al., 2002; Wang et al., 2003) show that climate variables
greatly influence in vegetation growing. Among them, precipitation, and temperature
(Cui and Shi, 2010) are pointed out as the most direct driver factors for vegetation growth.
Based on this assumption, Gessner et al. (2013) show that vegetation growth is strongly
correlated to temporal and spatial patterns of precipitation in semiarid ecosystems, mainly
because of the extreme precipitation events and seasonal shifts that affect vegetation
dynamics. In this line, temperature is also pointed as a critical driving factor in vegetation
activity (Braswell et al., 1997). In fact, Piao et al. (2006) report that temperature showed
positive effects on vegetation growth; decreasing, as temperature rises in high-cold areas.
Strongly related to water availability and temperature fluctuations, drought is an
important driving force in vegetation dynamics and it is the main limiting factor in
Mediterranean areas (Vicente-Serrano et al., 2019). The amount and timing of
precipitation controls vegetation productivity, inducing fluctuations in soil water content
throughout the seasons (Scheuring and Riedi, 1994). Therefore, it is essential to
understand how drought affects vegetation, as these events are expected to increase with
climatic change (Gao et al., 2016). All in all, climate and vegetation are strongly related
and there is a neccesity to implement adequate methodologies and novel tools to reveal
the intricate and non-linear relatiosnhips between both of them.

1.2.2 Vegetation as a complex system

Several studies (Hobbs et al., 2007; Scheuring and Riedi, 1994) present vegetation as a
spatial temporal-varying complex ecological system involving multiple variables. In this
sense, Mediterranean grasslands and high input crops are one of the most vulnerable

2
agroecosystems in Europe, being drought a major risk due to their location in an arid-
semiarid region.
In one hand, natural grasslands are considered one of the most important ecological
systems in the Iberian Peninsula, providing multiple agro-environmental services
(Konings et al., 2017), including the increase of biodiversity, conservation of landscape,
preservation of traditional values and fixation of rural population. Grasslands activity is
tightly linked with hydrological fluxes and terrain dynamics in river basins at a wide range
of space-time scales (Scheuring and Riedi, 1994), showing a great diversification of
scenarios (Herrick and Whitford, 1995). Thus, historical time-series analysis (Serrano et
al., 2019) is considered an proper method to examine and characterise pasture grasslands
complexity.
On the other hand, rural landscapes present a high spatial complexity as a result of the
dynamic interaction between the variable human actions and the spatial distribution of
biophysical cues (Laterra et al., 2012). In this sense, multiple authors attempt to reduce
the enormous intravariability in a crop system to improve its management by monitoring
each growth phase. Particularly, high input crops requires adequate fertilization and
management to maximize yield and obtain the targeted grain quality (Xue et al., 2007).
In fact, the misapplication of N fertilizers is a frequent problem in high-input systems,
which contributes to water and atmospheric pollution (Gabriel et al., 2016; Galloway et
al., 2008). Optimal fertilization of crops requires systematic monitoring of the N status
throughout the growing stages (Feng et al., 2016). On a field scale, data collection with
manual labour is often arduous, involving destructive sampling (De Castro et al., 2021;
Ramirez-Garcia et al., 2012), and is too time-consuming and expensive for rapid and
precise biomass and N status estimations (Potgieter et al., 2017; Yang et al., 2020).
Consequently, it is essential to further research rapid and non-destructive measurements
of crops traits at large spatial and multiple time scales.
1.2.3 Soil characterization
Vegetation activity is tightly linked with hydrological fluxes at a wide range of space-
time scales (Scheuring and Riedi, 1994), wherein soil plays an essential role, showing a
great diversification of scenarios (Herrick and Whitford, 1995). Soil moisture play as a
linking factor between precipitation, vegetation growth and soil; being the controller of
rainfall inputs and outputs, that define the temporal variability of soil moisture and, in
turn, vegetation activity in soils (Laio et al., 2001; Rodriguez-Iturbe, 2000; Scanlon et al.,
2005). In fact, Holzapfel et al. (2006) report that there is a clear relationship existing
3
between soil and water availability. Furthermore, among all the variables involved in the
water balance, clay content is one of the keys in the soil water content and the subsequent
vegetation response. This is in line with Baret et al. (1993) results that positively
correlates the clay content with the soil moisture.
A large body of scientific literature discusses the effect of soil when monitoring areas
with partial vegetation coverage, as it is in semiarid areas (Dyson et al., 2019; Li et al.,
2021; Qiao et al., 2022). These fluctuations are a result of the brightness and colour of
bare soils (Huete and Tucker, 1991; Xue and Su, 2017), soil-vegetation spectral mixing
(Chuvieco, 2002; Thomasson et al., 2001), soil texture (Danson and Plummer, 1995;
Quemada and Daughtry, 2016), among others.
The reduction of this soil effect continues to be a challenge in monitoring vegetation
(Duan et al., 2017; Mulla, 2013), and thus numerous techniques have been implemented
to distinguish and mitigate the background effect. Thus, it is important to consider the
soil influence, by quantifying the soil effect or reducing its effect in the semiarid-arid
zones.

1.3 Climate-vegetation-soil monitoring.

Monitoring large surfaces is still a great challenge. Although, remote sensing techniques
are pointed out as appropriate and promising tools to supervise grassland ecosystems and
vegetation cover changes in Mediterranean areas (Ali et al., 2016; Schmidtlein and Sassin,
2004). Primarily, because remote sensing is statistically dependable, operationally robust
and economically cost-effective. Nowadays, satellites allow obtaining historical real-time
acquisition data about the vegetation status in vast areas (Blanco et al., 2016). Optical
remote sensing techniques allow tracking vegetation cover due to its specific spectral
behaviour in the visible (0.4–0.7 mm) and infrared (0.74–1.11, 1.3–2.5 mm) bands of the
electromagnetic spectrum Remote sensing approaches have already been used in the
characterization of grassland type and vegetation change, the derivation of biomass or
grass cover (Boschetti et al., 2007), as well as for mapping grazing intensity (Kurtz et al.,
2010). In all these studies, remote sensing data are used to improve our understanding of
grasslands, revealing intra-and inter-annual variabilities.

Furthermore, remote sensing is not limited to monitoring ecological systems. Multiple

authors reported its use in the real-time monitoring of the N status of high-input crops
(Orlando et al., 2015; Řezník et al., 2020), enabling the collection of important data for
decision-making processes. Nitrogen application based on sensors has been reported to

4
provide N fertilizer savings between 5-45% without significant grain yield losses in
cereals (Colaço and Bramley, 2018).
Among the most recent approaches, sensors on board unmanned aerial vehicles (UAVs)
or drones have shown high efficiency in tracking and assessing vegetation status. Their
main advantages include the capability to fly at low altitudes providing ultra-high spatial
resolution imagery, great flexibility to schedule flights in the critical stages of crop
growth, and the opportunity to operate various sensors to acquire different ranges of the
electromagnetic spectrum (visible, infrared, thermal) (Del Cerro et al., 2021).

1.3.1 Precipitation indices

Drought is perceived as a deficiency in rainfall with respect to normal levels, namely, a
meteorological drought (Mishra and Singh, 2010). The standardised precipitation index
(SPI) is the most widely used indicator of drought episodes. This index can quantify the
degree of wetness by comparing the same time frame in an entire time series. Usually,
the time frame is defined from 1 up to 24 months. As a result of the statistical techniques
applied in SPI estimation, a probabilistic approach uses a z-score transformation to
convert the precipitation probability density function as a distribution with mean zero and
variance of one (Loukas et al., 2003; Vicente-Serrano et al., 2010). SPI is successfully
applied in tracking drought effects on irrigated areas (Khan et al., 2008), drought
monitoring in semiarid regions (Hadri et al., 2021; Zarei et al., 2021) and studying climate
change trends through historical rainfall data (W. Li et al., 2008).
The Standardised Precipitation Evapotranspiration Index (SPEI) is presented as an
improvement of the SPI index (Vicente-Serrano et al., 2010) and is an standard index for
monitoring drought episodes. To obtain SPEI, potential evapotranspiration (ETP) is
calculated and subtracted from precipitation data. Then, the same statistical approach to
estimate the SPI is applied to the previously obtained water balance data. In general, SPEI
was shown to be a better drought indicator than SPI (Pei et al., 2020; Tan et al., 2015)
due to the addition of the temperature effect in its calculation.
1.3.2 Vegetation indices
Vegetation indices (VIs) based on remote sensing are used to investigate spatial and
temporal patterns of grassland production. They are defined as a combination of two or
more spectral bands related to vegetation status. Vegetation indices present a close
relationship with climate (Mao et al., 2012) across different biomes and bioregions (B.
Guo et al., 2014). Significantly, the Normalised Difference Vegetation Index (NDVI)

5
demonstrates to be an excellent indicator of the vegetation growth conditions (Huete et
al., 2002; Rouse J. et al., 1974) and the biophysical characteristics of ecosystems.
Multiple studies reveal (Jafari et al., 2017; Yang et al., 1998) that NDVI is an adequate
tool to monitor grasslands conditions, being the most widely used spectral VI (Yagci et
al., 2014) by ecologists and agriculturalists until nowadays. It is employed in many cases,
such as drought assessment or biomass estimation (Karnieli et al., 2010; Mkhabela et al.,
2011), herbaceous and woody grassland characterisation (Blanco et al., 2016; Chen and
Weber, 2014) and for damaged pasture and forage insurance over the last few years in
Spain (Martín-Sotoca et al., 2019). NDVI values tend to increase during the growing
season, showing the biomass increase due to the intense photosynthetic activity. On the
contrary, there is a gradual reduction of NDVI values when there is a lack of water, or the
temperature is excessively higher.
Vegetation indices are employed in multiple circumstances depending on the purpose of
the research. They can also be used in drought monitoring (Gidey et al., 2018a; Kogan,
1995), in that case, authors suggest that Vegetation Health Index should be employed
(VHI) instead of NDVI, due to its great sensitivity to detect drought periods.
There are also VIs specialised in the monitoring of high-input crops traits, that correlates
closely with crop cover and biomass growth (Benedetti and Rossini, 1993; Rouse J. et al.,
1974). Therefore, they can be used to support decision-making processes, such as
variable-rate N fertilization (Argento et al., 2021; Sanz et al., 2021b; Tremblay et al.,
2009) and grain yield estimation (R. Liu et al., 2022; Magney et al., 2016). Among them,
the Normalized Difference Red–Edge index (NDRE) (Barnes et al., 2000) seems to be a
promising index which includes the red-edge spectral region (690–730 nm) in its
formulation. Several studies (Jiang et al., 2021; Marszalek et al., 2022; Raya-Sereno et
al., 2021) support that the red-edge channel is more sensitive to chlorophyll content.
Furthermore, VI based only on the visual region such as the Blue–Red (BRI1) are really
interesting due to its use in conditions when the camera is the major restriction (Zarco‐
Tejada et al., 2005) . It is suggested that BRI is sensitive to changes or disturbances in the
photosynthetic apparatus (Li et al., 2014; Lichtenthaler et al., 1996), enabling the
identification of early stress in plants.
Overall, VIs encompass a wide variety of applications, but it is also necessary to account
for their limitations, such as VI saturation (Zarco‐Tejada et al., 2005), bare soil influence
(Huete et al., 1985; Ren et al., 2018), and atmospheric noise (Hui and Huete, 1995;
Pancorbo et al., 2021a).

6
1.3.3 Bare soil role in vegetation monitoring
Several authors (Eisfelder et al., 2012; Huete et al., 1985; Ren et al., 2018) point out that
distinguish bare soil and vegetation is still a challenge due to the considerable impact that
background reflectance soil has on VIs (Almeida-Ñauñay et al., 2022b), Regions with
sparse vegetation tend to generate high reflectance values that might saturate sensors or
produce biased biomass and vegetation cover estimations. Particularly, NDVI might
present several limitations when bare soil and organic material represent a substantial
portion of the analysed surface. In those cases, NDVI may not appropriately represent
vegetation activity (Eisfelder et al., 2012; Huete, 1988; Qi et al., 1994).
Different approaches have been proposed to distinguish and mitigate the background soil
effect. On one hand, several authors focus on the correction of the VIs obtained through
optical monitoring. Huete (1988) proposed a correction based on the utilization of a
constant parameter, adjusted to a predefined background reflectance. Qi et al. (1994)
added a variable parameter based on the soil. Thus, a modified soil-adjusted vegetation
index (MSAVI) was proposed as a solution to consider the soil background effect in
semiarid areas. MSAVI has been successfully applied in numerous studies (Jiang et al.,
2013; Li et al., 2019), especially in the estimation of above-ground biomass in semiarid
areas.
On the other hand, alternative approaches, such as radar techniques, have been pointed
out as a promising tool for vegetation monitoring, crop mapping and soil monitoring.
Synthetic aperture radar (SAR) based on techniques are the most widely used (Steele-
Dunne et al., 2017) in the agricultural field of study. Their principal advantages are the
capacity to obtain information independent of the weather conditions, minimising the
atmospheric effects, and the capacity to penetrate through the soil. However, they also
present disadvantages compared to optical remote sensing techniques, such as possible
speckle effects (Liu et al., 2021), that could reduce the quality of SAR imagery and the
disturbance of topography (Das et al., 2015) in hilly regions.

1.4 Climate-Vegetation-Soil complexity

Vegetation indices time series are essential to remote sensing vegetation phenology and
obtaining analytical measurements of vegetation dynamics. Determining variability and
quantification from remotely sensed data could be challenging (Verbesselt et al., 2010)
since time series consists of a combination of seasonal fluctuations, gradual and sudden
changes, and long-term and short-term variations that alter its overall structure.

7
1.4.1 Temporal complexity
In ecological systems, temporal variability is frequently measured as the standard
deviation of the records in a time series, though, these systems present nonlinear
characters as in any complex system. In particular, VIs time-series present time cycles,
allowing agro-environmental system dynamics description (Martínez and Gilabert, 2009;
Zhao et al., 2015; Zhong et al., 2010).
As it is reported by several authors, time scale is a critical factor to study the relationships
between VIs and climate. Multiple studies are reported through the literature regarding
the multi-time-scale approach. Among them, it is the concept of drought in semiarid
grasslands. The resilience and stability to drought is dependent on the grassland growth
phases (Zhang et al., 2017). Moreover, multi time scale drought assessment depends on
the vegetation cover, providing different responses vegetation on semiarid and semi
humid biomes.
Previous studies suggest that VIs response to climate differs at different timescales across
the year (Chamaille-Jammes et al., 2006; Ma et al., 2013; Zhang et al., 2018), showing
differences inter-monthly, inter-seasonal and inter-annual time scales (Li et al., 2018;
Zhang et al., 2018). Even more, it has been revealed optimal relationships with a lagged
response between VIs and climate (Mu et al., 2013).
On one hand, a high correlation between precipitation and VIs has been reported at a
yearly scale (Wang et al., 2003). On the other hand, several studies reveal a delayed
correlation between precipitation and NDVI on a monthly scale (Chen et al., 2020;
Gessner et al., 2013). Relationships between NDVI and temperature are also reported at
different time scales. Tibetan Plateau (China) grasslands present positive correlations
during the growing season (May–September) (Cong et al., 2017). Whereas, (Wang et al.,
2015) obtained correlations that present spatial–temporal variations, probably due to the
differences in environmental factors or plant functional traits over which the correlations
were calculated. In this way, researchers suggest that the optimal VI delayed response
depends on the climate variable, shifting from one to two months in temperature and
precipitation (Shen et al., 2014).
As it is shown in previous sections, VIs time series are complex, and quantitative methods
(Martínez and Gilabert, 2009) may ignore that are usually non-stationary and inherently
non-linear. Therefore, they should be analysed with a corresponding methodology. In this
line, Eckmann et al. (1987) introduced recurrence plots (RPs) and cross recurrence plots
(CRPs) as a simple way to visualize the periodic or chaotic behaviour of a dynamic

8
system. Basically, RP takes a single time series measurement, projecting it into
multidimensional space by embedding procedures, and identifies time correlations
(recurrences) with itself that is not apparent in the one-dimensional time series. Then, the
original description of RP is extended by computing an array of specific recurrence
variables that quantify the deterministic structure and complexity of the RP (Marwan et
al., 2007).
Furthermore, RPs can be extended to include multivariate relationships through Cross-
Recurrence Plots (CRPs). CRP is defined as a bivariate extension of RP (Marwan et al.,
2007). It is computed to analyse two variables by comparing their states and studying the
dependencies between two different systems; it may be regarded as a nonlinear cross-
correlation function. Unlike other conventional time series techniques, RP and CRP
analyses are not limited by data stationarity constraints (Marwan and Kurths, 2002) and
allow to visualize and quantify the topology hidden in the time series data (Thiel et al.,
2004).
Recurrence plots are used to detect dynamical patterns in time series (Marwan et al.,
2007), and Recurrence Quantification Analysis (RQA) quantifies and characterises the
small-scale structures in RP. Recurrence plots can be visually interpreted to distinguish
non-stationary dynamics with either smooth or abrupt transitions (Proulx et al., 2008) and
finding the presence of periodic and non-periodic processes. Recurrence Plots–
Recurrence Quantification Analysis (RP–RQA) are able to measure temporal
determinism and predictability (Proulx et al., 2015). In this line, several parameters are
obtained: determinism measures the time series stochasticity, and the average length of
diagonal structures is interpreted as the predictability degree of a state.
There were several works in the framework of recurrence analysis and Vis; S. C. Li et al.
(2008) applied RPs to measure the determinism and predictability of the NDVI series and
spatial patterns. Zurlini et al. (2014) detected land cover transformation by NDVI
measurements; Sparavigna and Marazzato (2015) studied seasonal oscillations of the
vegetation health index (VHI) in Italy. Zhao et al. (2015) applied a quantification analysis
to RPs, named Recurrence Quantification Analysis (RQA), to several NDVI series to map
their determinism and analyse the spatial pattern of different zones. Zurlini et al. (2018)
studied rangelands' periodic and chaotic components to differentiate recovery capacity
after uncontrolled fires. Recently, Semeraro et al. (2020) applied CRPs and RQA to VI
time series to evaluate the response of the forest ecosystem to disruptive events, such as
droughts, with success. Including bidimensional RPs of NDVI series (Dias et al., 2020)

9
to establish a pixel-wise classification applied in a eucalyptus forest. Overall, Marwan et
al. (2015) revealed the usefulness of RPs to describe non-linear behaviours in high-
dimensional systems, such as VI time series.
1.4.2 Spatial complexity
Spatial self-organization in a landscape is driven by feedbacks between vegetation and
the environment. The emergence of these patterns can be represented in continuous or
high-resolution spatially explicit models (Franklin et al., 2020), which reveal that the
underlying positive feedback loop is a common organizing principle for spatial patterns
across distinct locations and systems. Unluckily, rural spatial distribution is a
multidimensional property of landscapes (Roschewitz et al., 2005) depending on their
spatial configuration metrics (e.g. number, size, shape) and their composition (e.g.
proportion of non-arable land, covert types diversity) which are partially correlated
among them.
Spatially distributed soil pixels represent a problem in high resolution monitoring.
Numerous techniques have been implemented to classify soil and vegetation
independently. The object-based image analysis (OBIA) method reduces intraclass
spectral variability by grouping homogeneous and contiguous vegetation pixels (Torres-
Sánchez et al., 2015; Zarco-Tejada et al., 2019). The OBIA reduces the spectral variability
efficiently. However, the main inconvenience is to correctly set the segmentation
parameters for the varying sized, shaped, and spatially distributed image-objects
composing the image.
Machine-learning classification methods use supervised and unsupervised algorithms
such as neural networks and support vector machines (SVM) to identify vegetation pixels
(Dyson et al., 2019; X. Zhou et al., 2021). However, depending on the algorithm, several
challenges should be faced on, such as: time-consuming processes (computational time
and model training), the incomplete detection or insensitivity to similar targets and the
highly dependence on the training data set reliability (Zhang et al., 2020).
One of the most accepted methods for agricultural purposes is the threshold calculation
based on the Otsu methodology (Guijarro et al., 2011; Otsu, 1979; Torres-Sánchez et al.,
2015). It considers that the image contains two classes of pixels (bare soil and vegetation)
and calculates the optimum threshold based on minimizing the intraclass variance. In
general, thresholding values methodologies are appreciated due to its simplicity in the
implementation (Zhu et al., 2018).

10
Even though all the implemented methodologies, there is still a necessity to further
develop an accurate vegetation segmentation. Mainly, to find an optimal VI threshold
value that sets the breaking point between vegetation and bare soil reducing the variability
in high-resolution imagery.

1.5 Research objectives

Vegetation indices time series are essential to remote sensing vegetation phenology and
traits. However, quantitative methods (Martínez and Gilabert, 2009) might ignore that
VIs time series are usually non-stationary and inherently non-linear. Thus, the main
objective of this thesis is to implement methodologies and strategies to manage the spatial
and temporal variability in agricultural systems.
Therefore, the main question of the thesis arises:

Is it possible to implement non-linear and optimization tools to unravel the temporal

variability and reduce spatial variability of the vegetation-soil-climate system to improve
the monitoring of agricultural systems?

a) Is it feasible to ascertain the complexity of vegetation time series related to climate

fluctuations and drought episodes through VIs (VHI and SVHI) and precipitation
indices (SPI and SPEI)? Is it possible to determine which index performed better
and the most adequate time scale?
b) Are the VIs and recurrence techniques an adequate tool to picture the complexity
of the relationships between vegetation and climate in grasslands? If it is, which
is the most adequate VI to be used in this case?
c) Is it possible to identify variations in the temporal dynamics of two different agro-
ecological zones through recurrence techniques and VIs applied in the anomalies
time series?
d) Is it viable to implement a methodology to reduce the high spatial intra-variability
in a crop by removing the soil background influence with a defined threshold? If
it is, is it possible to determine the optimal threshold and evaluate the predictive
performance of the methodology?

The chapters that have been developed to address the stated objectives are detailed below:
- Chapter 3. Assessment of Drought Indexes on Different Time Scales: A Case in
Semiarid Mediterranean Grasslands
• The results obtained in the article: (Almeida-Ñauñay et al., 2022a).

11
- Chapter 4. The Vegetation–Climate System Complexity through Recurrence Analysis
• The results obtained in the article: (Andrés F. Almeida-Ñauñay et al., 2021)
- Chapter 5. Recurrence plots for quantifying the vegetation indices dynamics in a
semi-arid grassland
• The results obtained in the article: (Almeida-Ñauñay et al., 2022b)
- Chapter 6. Optimization of soil background removal to improve the prediction of
wheat traits with UAV imagery
• The preliminary results of the article in the peer revision process titled:
Optimization of soil background removal to improve the prediction of wheat
traits with UAV imagery

1.6 Related publications

The doctoral thesis was elaborated as a part of the project No. PGC2018-093854-B-I00
of the Spanish Ministerio de Ciencia Innovación y Universidades of Spain and the funding
from the Comunidad de Madrid (Spain), Structural Funds 2014–2020 512 (ERDF and
ESF), through project AGRISOST (S2018/BAA-4330) the founds obtained through the
Boosting Agricultural Insurance based on Earth Observation data—BEACON European
project under agreement No. 821964, funded under H2020_EU, DT-SPACE-01-EO-
2018-2020 and the support of the grant from Agencia Estatal de Investigación (PID2019-
106433GB-I00 / AEI / 10.13039/501100011033) of Spain,. The work was developed in
the CEIGRAM (Research Centre for the Management of agricultural and Environmental
Risks) between 2019 and 2022.

As a result of this work, three scientific publications were published, and one additional
paper, is currently in the reviewing process. Additionally, we have contributed to national
and international scientific congresses, which are listed below:
- Scientific articles

Andrés Felipe Almeida-Ñauñay, Rosa María Benito Zafrilla, Miguel Quemada Sáenz-
Badillos, Juan Carlos Losada, and Ana María Tarquis Alfonso, 2021. The Vegetation–
Climate System Complexity through Recurrence Analysis. Entropy.
DOI: https://doi.org/10.3390/e23050559

Andrés Felipe Almeida-Ñauñay Benito, R.M., Quemada, M., Losada, J.C., Tarquis,
A.M., 2022. Recurrence plots for quantifying the vegetation indices dynamics in a semi-
arid grassland. Geoderma.
12
DOI: https://doi.org/10.1016/j.geoderma.2021.115488

Andrés Felipe Almeida-Ñauñay, María Villeta, Miguel Quemadaa and Ana María
Tarquis Alfonso, 2022. Assessment of Drought Indexes on Different Time Scales: A Case
in Semiarid Mediterranean Grasslands. Remote Sensing.
DOI: https://doi.org/10.3390/rs14030565

Andrés F. Almeida-Ñauñay, Ana M. Tarquis, Juan López-Herrera, Enrique Pérez-

Martín, José L. Pancorbo, María D. Raya-Sereno, and Miguel Quemada. Optimization of
soil background removal to improve the prediction of wheat traits with UAV imagery.
Computers and electronics in agriculture.
DOI. (Under review)

- Other scientific contributions

Andrés Felipe Almeida-Ñauñay, Rosa María Benito Zafrilla, Miguel Quemada Sáenz-
Badillos, Juan Carlos Losada, and Ana María Tarquis Alfonso, Recurrence
Quantification Techniques of vegetation time-series indices in semiarid
grasslands. European Geosciences Union (EGU) 2020. (V.pico presentation)
https://meetingorganizer.copernicus.org/EGU2020/EGU2020-9937.html

Andrés Felipe Almeida-Ñauñay, Rosa María Benito Zafrilla2, Miguel Quemada Sáenz-
Badillos3, Juan Carlos Losada4, and Ana María Tarquis Alfonso, Complexity of
the Vegetation-Climate System Through Data Analysis, in: Benito, R.M.,
Cherifi, C., Cherifi, H., Moro, E., Rocha, L.M., Sales-Pardo, M. (Eds.), Complex
Networks & Their Applications IX. Volume 1, Proceedings of the Ninth
International Conference on Complex Networks and Their Applications.
COMPLEX NETWORKS 2020. Springer International Publishing, Cham, pp.
609–619.
https://doi.org/10.1007/978-3-030-65347-7_50

Andrés Felipe Almeida-Ñauñay, Rosa María Benito Zafrilla, Miguel Quemada Sáenz-
Badillos, Juan Carlos Losada, and Ana María Tarquis Alfonso, Revealing
climate and vegetation indices interactions through Cross Recurrence
Techniques. European Geosciences Union (EGU) 2021. (V.pico presentation)
https://meetingorganizer.copernicus.org/EGU21/EGU21-12723.html

13
Andrés Felipe Almeida-Ñauñay, Rosa María Benito Zafrilla, Miguel Quemada Sáenz-
Badillos, Juan Carlos Losada, and Ana María Tarquis Alfonso, 2021.
Recurrence techniques for the analysis of vegetation indices and climate
anomalies: a study case in semiarid grasslands, in: Proc.SPIE. p. 118560R.
https://doi.org/10.1117/12.2600222

14
2 MATERIALS AND
METHODOLOGY

15
Different methodologies were employed according to the objectives and the questions of
this thesis; however, all the methodologies share the disentanglement of complex
information in the image analysis. The section 2.1 consists in the description and the
analysis of the complexity in the vegetation-climate-soil system. The section 2.2
introduce the complexity of the climate-vegetation-soil system by a correlation and cross-
correlation analysis. The section 2.3 shows the estimation of the climate and VIs
anomalies time series. The section 0 introduces the recurrence plots methodology, applied
to analyse the complexity of the system in combination with anomalies time series.
Finally, in the section 2.5, it is introduced a novel methodology to reduce the spatial
variability of the soil pixels in high-resolution imagery, by a sequential thresholding
removal of soil pixels based on a percentile approach.

2.1 Complexity of vegetation and drought indices at different time Scales.

Standardization and drought episodes.
2.1.1 Area of study
Two study areas were considered in the first section (Figure 1): ZGU in Guadalix de la
Sierra and ZSO in Soto Del Real, both in Madrid (Spain). These sites are characterised
by a Mediterranean climate with scarce precipitation, warm summers and cold winters
(Rivas-Tabares et al., 2019). The altitude is similar in ZGU and ZSO, 820 and 958
m.a.s.l., respectively, whereas the slope is higher in ZGU than in ZSO (Table 1).

16
Figure 1. Site and climate station’s location, ZSO is on the left and ZGU is on the right. Images were
obtained from the Mosaic of current orthophotos corresponding to sheet 0509 of the MTN50 in Spain during
the years 2017,2018. Pixel plot size: 250x250 m2. Pixel orthophotographs size: 0.05 x 0.05 m2. RGB
composition. Blue crosses are the locations of the climate station locations and red pinpoints are the soil
trench locations.

Table 1 Soil horizons and zone characteristics of both study areas. ZGU = Guadalix de la Sierra, ZSO =
Soto del Real

ZGU ZSO
Thickness (cm) 5 Thickness (cm) 3
Colour 10YR3/4 Colour 10YR3/2
Silt (%) 29 Silt (%) 19
Coarse Sand (%) 12 Coarse Sand (%) 37
A Fine Sand (%) 25 Au Fine Sand (%) 38
Clay (%) 34 Clay (%) 6
Bulk Density (g/cm3) 1.2 Bulk Density (g/cm3) 1.5
Water Holding Water Holding
13,3 11,7
Capacity (%) Capacity (%)
Thickness (cm) 20 Thickness (cm) 12
Colour 10YR3/4 Colour 10YR4/4
Silt (%) 22 Silt (%) 18
Coarse Sand (%) 24 Coarse Sand (%) 30
AB Fine Sand (%) 13 Au2 Fine Sand (%) 47
Clay (%) 41 Clay (%) 5
Bulk Density (g/cm3) 1.2 Bulk Density (g/cm3) 1.7
Water Holding Water Holding
12,6 11.1
Capacity (%) Capacity (%)
Slope (%) 11.7 Slope (%) 4.7
Zone Height (m) 820 Zone Height (m) 958
attributes Precipitation (mm) 576 attributes Precipitation (mm) 550
Temperature (ºC) 12.6 Temperature (ºC) 13.6

17
Both study areas are located on the down hillsides of the Guadarrama Sierra (Central
Spain). The northern piedmont slope of Madrid's basin belongs to the Upper Aragonian,
and it is constituted of an arkose nature substrate. The Manzanares-Jarama tertiary
interfluvial presents the typical morphological features of a foothill in destruction due to
the enclosure of the Tagus river that is emptying the Madrid's Basin (Vaudour et al.,
1979). Old surface degradation has originated in various complex forms. Fragile surface
deposits related to the denudation process are combined with arkose lithofacies. These
result in severe problems of substrate differentiation (Carral et al., 1996). Particularly,
ZSO parent material is situated over pre-hercenic igneous rocks material, composed
mainly of glandular orthogneisses. Meanwhile, ZGU parent material belongs to the
tertiary period and is mainly composed of blocks and conglomerates of stone edges and
polimyctic blocks (Bellido Mulas et al., 2004; Vaudour et al., 1979).
There are differences in physical and chemical properties found in the soil trench dug in
each study area, as showed in Table 1 (Schmid et al., 2000). The ZGU soil is a Calcaric
Regosol (FAO, 2014) with thin topsoil (0-5 cm) followed by AB horizon (5-25 cm). Both
present a clay loam texture, 5% organic matter, a pH of 7.2 and a bulk density of 1.20 g
cm-3. The ZSO soil is a Dystric Cambisol composed of three different horizons. The Au
topsoil (3 cm) and Au2 horizon (20 cm) are similar, presenting the same sandy loam
texture, a pH ranging from 5.4 to 5.8 and a bulk density varying from 1.5 to 1.7 g cm -3.
These are followed by a Bw horizon (20 cm), with a sandy loam texture with higher clay
content and a similar pH and bulk density to the rest of the horizon. Water-holding
capacity (WHC) was estimated in the topsoil and underneath horizon from the
pedotransfer function proposed by (Saxton and Rawls, 2006), being higher in Guadalix
de la Sierra (13%) than that in Soto del Real (11.1%).
The areas selected are inland Mediterranean grasslands that grow during spring and
autumn and have vegetative winter dormancy and a summer senescence period. These
vegetative periods depend mainly on regional climatological conditions. The images
analysed in the current study represent the seasonal cycle. Both areas belong to public
natural dehesas, periodically grazed all year round by sheep and cattle, at low stocking
densities (less than 0.5 livestock units/ha) (San Miguel-Ayanz et al., 2009). At the end of
the summer, these grasslands expose a mix of bare soil and vegetation (live or dead)
captured during imagery acquisition.
ZGU grasslands tend to develop on heavy texture soils, providing moisture for most of
the year, except in midsummer, for climatic reasons. Common species in these natural

18
pastures are Brachypodium phoenicoides, Elytrigia hispida, Elytrigia campestris,
Dactylis hispanica, Medicago sativa, Poa angustifolia, and Arrhenatherum album (Peco
et al., 2006; San Miguel-Ayanz et al., 2009). ZSO area is an early pasture grassland rich
in relative thermophiles belonging to the Agrostion castellanae alliance. They are
acidophilic, Mediterranean pastures, dominated by annual lively herbaceous, mostly
gramineous of medium to tall size, tending to grow in the transition zones to Guadarrama
Sierra. The dominant species in these natural pastures are Agrostis castellana, Festuca
ampla, Dactylis lusitanica (San Miguel-Ayanz et al., 2009).
2.1.2 Satellite imagery and vegetation indices calculations
In both areas, grassland plots were selected based on four criteria: i) closer to the soil
trench dug, ii) maximum surface covered by grassland without wood, iii) continuous
pasture use during the analysed period, and iv) grassland cover in the surrounding area.
ZSO was composed of four pixels of 250×250 m2 between (3º45'00" W, 3º46'00" W) and
(40º43'00", 40º44'00" N). The same number of pixels were selected in the ZGU area
delimited by (3º38'00" W, 3º39'00" W) and (40º46'00" N, 40º47'00" N). Each one of these
sites represents an area of 1 km2 approximately.
MODIS (Terra and Aqua) and Landsat are the most employed imagery in long-term
vegetation research (Reinermann et al., 2020), being successfully applied in the study of
semiarid grassland cover and playing an important role in the pasture insurance policy in
Spain (Sanz et al., 2021b). The MOD09Q1 product offers imagery at 250 m resolution of
the RED (620-670 nm) and NIR (841-876 nm) surface reflectance. Observations from the
same orbit are composited by observational coverage and the highest coverage
observations are stored in an 8-days period. Each orbit's observation is rated based on
multiple criteria: clouds, cloud shadow, aerosol quantity, and view angle (Vermote,
2015). Atmospheric correction is computed using an LPDAAC algorithm specialised in
curing MODIS imagery data (Vermote et al., 2002; Vermote and Saleous, 2006). The
reflectance of the study plot was monitored from 2001 to 2020. Each year, 46 images
were acquired, giving a total of 920 images during the study period. Vegetation indices
were calculated from the reflectance bands obtained in the MOD09Q1 product for each
selected pixel (Table 2).

19
Table 2. Selected vegetation-drought indices

Index Equation Reference

Normalized difference NIR (860 nm) − 𝑅𝐸𝐷(660 nm) (Rouse J. et al.,
𝑁𝐷𝑉𝐼 =
vegetation index (NDVI)1 NIR (860 nm) + 𝑅𝐸𝐷(660 nm) 1974)
Vegetation condition index 100 ∗ (𝑁𝐷𝑉𝐼 − 𝑁𝐷𝑉𝐼𝑚𝑖𝑛 )
𝑉𝐶𝐼 = (Kogan, 1997)
(VCI) (𝑁𝐷𝑉𝐼𝑚𝑎𝑥 − 𝑁𝐷𝑉𝐼𝑚𝑖𝑛 )
Temperature condition index 100 ∗ (𝐿𝑆𝑇𝑚𝑎𝑥 − 𝐿𝑆𝑇)
𝑇𝐶𝐼 = (Kogan, 1997)
(TCI) (𝐿𝑆𝑇𝑚𝑎𝑥 − 𝐿𝑆𝑇𝑚𝑖𝑛 )
Vegetation health index
𝑉𝐻𝐼 = (𝛼 ∗ 𝑉𝐶𝐼) + (1 − 𝛼) ∗ 𝑇𝐶𝐼 (Kogan, 1997)
(VHI)2
1
NIR is the near-infrared reflectance (860 nm), and RED is the red reflectance (660 nm). 2α is a coefficient
that depends on the dry conditions of the area.

The normalized difference vegetation index (NDVI) is an adequate estimator of the

photosynthetic activity and vegetation growth. Dense and healthy vegetation shows high
values (>0.5), whereas sparse or stressed vegetation presents low values, due to adverse
conditions, mainly climate and soil. NDVI series was calculated in each pixel, and their
average value was employed in the analysis.
According to scientific literature, the vegetation health index (VHI) is shown to be an
improved version of NDVI, allowing to detect drought and non-drought periods (Kogan,
1997). Two components define it mathematically. The first represents the vegetation
component, named as vegetation condition index (VCI), and the second component
corresponds to the land surface temperature and is defined as temperature condition index
(TCI). TCI is considered a thermal stress indicator to quantify temperature-related
drought periods. This index is sensitive to vegetation drought stress based on soil moisture
variations during wet and drought periods. Land surface temperature (LST) was obtained
from MOD11A2 product, which provided the temperature average 8-day per-pixel with
a spatial resolution of 1 Km (Wan et al., 2015).
2.1.3 Climate data and precipitation indices
Daily meteorological data were obtained from two different AEMET (Agencia Estatal de
Meteorologia) stations. ZGU station is established at Venturada (40° 47' 53.79" N, 3° 37'
30.49" W) and ZSO meteorological station is situated between Soto del Real and
Colmenar Viejo (40° 41' 46.008" N, 3°3 ° 45' 54.019" W). Both are nearby to the selected
zones (Figure 1).
Daily series of average temperature and precipitation from 2001 to 2020 were extracted.
Annual average precipitation and temperature were obtained from a 20-year time series
(Table 1) at each location. The average precipitation of Guadalix de la Sierra (576 mm)

20
was higher than Soto del Real, whereas the temperature (12.6 °C) was lower than Soto
del Real (560 mm), respectively.
Precipitation and temperature series were resampled to an eight-day frequency to be
analysed on the same time scale as the VHI time series. Precipitation (PCP) is the
accumulated precipitation every eight days, and the temperature (TEMP) is the average
temperature in the same period. Both series were analysed using descriptive statistics to
characterise the climate conditions of the study. Then the Hargreaves-Samani expression
(Hargreaves and Samani, 1985) was used to calculate the reference evapotranspiration
(ET0; mm d-1) based on equation [1].

𝐸𝑇0 = 0.0135 ∗ 𝑘𝑅𝑆 ∗ 0.408 ∗ 𝐻0 ∗ (𝑇𝑚 + 17.8) ∗ (𝑇𝑥 − 𝑇𝑛 )0,5 [1]

where kRS is the Hargreaves empirical coefficient; H0 is extraterrestrial radiation (MJ m-2
d-1); and Tm, Tx and Tn are the daily averages, maximum and minimum air temperature
(ºC), respectively. Daily extraterrestrial radiation (H0) is calculated based on the day of
the year and the geographical position of each plot (Duffie and Beckman, 2013). The
value of kRS was set at 0.16 as recommended for arid and semiarid interior regions
(Hargreaves, 1994).
2.1.4 Standarised indices
As mentioned above, standardised indices are calculated on various time scales. In our
case, we have used monthly (1 month) and quarterly (3 months) time scales in this work,
as these are the most usual time scales found in the literature (An et al., 2020; Kang et al.,
2018; Pei et al., 2019) and the most suitable to be studied. Annual scale was not
considered in this work because quarterly trends and tendencies of VHI could be
vanished. On the quarterly scale, each trimester presented a constant time length and a
VHI trend, representing the semi-arid grasslands studied in Spain. The first one goes from
January to March, and the last one from October to December.
The SPI is computed by fitting a given precipitation series' gamma probability density
function (PDF). This adjustment is performed separately for each temporal basis on which
the precipitation is built. Then, the data is transformed to a normal distribution where the
mean SPI for the station and the needed period is zero.
Positive SPI values show excess rainfall, while negative SPI values correspond to a
precipitation deficit. Further details on the SPI method could be found in the (Asadi Zarch
et al., 2015; Malik and Kumar, 2021) works.

21
To obtain the SPEI, we calculated the daily water balance, which is the difference
between the daily precipitation and the daily potential evapotranspiration. We then
aggregated the values at the time scales of interest to acquire the starting data (Jiang et
al., 2020). Later, the same standardisation method of SPI is applied to the water balance
time series. The method used to calculate SPEI is fully described in (Vicente-Serrano et
al., 2010) work.
The approach to calculating the standardised vegetation health index (SVHI) is primarily
based on the standardisation method used in SPI calculation applied to VHI values.
Standardisation is applied to the VHI values of the month's last date when working on a
monthly scale. In the case of the quarterly scale, it is applied at VHI values of the
trimester´s last date.
The scientific review revealed that it is recommended to use a variable α in the VHI
calculation depending on the study area (Bento et al., 2018a; Bento et al., 2018b).
Therefore, it was maximised the correlation between SVHI, SPI, and SPEI to obtain the
optimal α in each zone and time scale. From now on, the subsequent analyses were
applied to the optimal SVHI. The optimal VHI was estimated using the same α values
obtained in the optimization of SVHI.
2.1.5 Correlation, classification and quantification analysis of drought Episodes
We used Pearson's correlation analysis to examine the standardised indices relations in
each zone. In the first step, we calculated the correlation values along with the whole time
series. Equation [2] is used to compute Pearson's correlation values.

𝑐𝑜𝑣(𝑆𝑉𝐼𝑖 , 𝑦𝑖,𝑙 )
𝜌= [2]
𝜎(𝑆𝑉𝐼𝑖)∗ 𝜎(𝑦𝑖,𝑙)

where SVIi is the SVHI or VHI values at i time scale, and the yi,l is the l indices (SPI or
SPEI) values at the i time scale. We calculated the correlation for each unit scale: month
and quarterly to determine the best correlated time frame in the series.
Based on a bibliography review in the field of meteorological and agricultural drought
indices (Ali et al., 2017; Gidey et al., 2018b; Ye et al., 2015), we classified SPI, SPEI,
VHI, and SVHI to define four different drought classes. Values were classified as follows
in Table 3:

22
Table 3. Classification of meteorological and agricultural drought indicators using the index value (IV).
The threshold values of IV are adapted from (Ali et al., 2017; Gidey et al., 2018b; Ye et al., 2015).

SPI, SPEI and SVHI VHI

Extremely drought IV < -2 IV <10
Severe drought -2 < IV < -1.5 10 < IV < 20
Moderate drought -1.5 < IV < -1 20 < IV < 35
No drought IV > -1 IV > 35

2.2 The Vegetation–Climate system complexity

2.2.1 Area of study
In the second section of this thesis two study areas were studied (Figure 2): one located
in Tornadizos de Ávila, Ávila (ZAV) and the other in Soto Del Real, Madrid (ZMA). Both
sites are characteristics of inland Mediterranean grasslands under a Mediterranean climate
with continental influence (temperate with a dry season and hot summer (Cfa), according
to Köppen classification). Both grasslands rise during spring and autumn and have a
summer senescence period and a vegetative winter dormancy of variable annual length
(Table 4).

Figure 2. Localisation of the study zones and the three selected plots in each one. Left: ZAV, Tornadizos
de Ávila, Ávila (ZAV_1, ZAV_2, and ZAV_3), Right: ZMA, Soto del Real, Madrid (ZMA_1, ZMA_2 and
ZMA_3). Mosaic of the most current orthophotographs corresponding to sheet 0509 of the MTN50 in
Spain, during the years 2017, 2018. Pixel size: 0.25 km. RGB composition.

23
Table 4. Annual climate and soil characteristics of both study zones. ZAV: Tornadizos de Ávila (Avila),
ZMA: Soto Del Real (Madrid).

Variable ZAV ZMA

Slope (%) 4.2 (±1.10) 4.7 (±1.60)
Height (m) 1290 (±70) 958 (±30)
Silt (%) 20 (±2.00) 18 (±1.00)
Sand (%) 60 (±2.00) 76 (±1.00)
Clay (%) 20 (±2.00) 6 (±1.00)
Bulk Density (g/cm3) 1.3 (±0.10) 1.6 (±0.10)
pH 6.5 (±0.30) 5.6 (±0.20)
Organic Matter (%) 3.8 (±0.20) 3.0 (±0.10)
Water Holding Capacity (%) 14.4 (±1.00) 11.1 (±1.00)
Precipitation (mm) 400 (±150.00) 560 (±80.00)
Temperature (°C) 11.6 (±0.60) 13.6 (±0.60)

Plots were chosen based on three criteria: (i) maximum surface covered by pasture
grassland without woodland, (ii) continuous pastureland practices during the analysed
period, (iii) pastureland cover in the surrounding area. Thus, the ZAV zone was composed
of three pixels of 500 × 500 m between (3°45’00” W, 3°46’00” W) and (40°43’00” N,
40°44’00” N). ZMA was composed of three pixels (500 × 500 m) enclosed between
(4°32’00” W, 4°33’00” W) and (40°37’00”, 40°39’00” N).

2.2.2 Satellite imagery and meteorological data

Like in the previous section, it was selected a product from the MODIS collection. In this
case, MOD09A1 was chosen for this study. This product is a level-3 composite of the
500-m resolution, and the best pixel observation was chosen within eight days. The
selection criteria for the best pixel observation were the aerosol content, view and solar
zenith angle, cloud presence and clouds shadows (Martín-Sotoca et al., 2019; Vermote,
2015).
Study plots reflectance was monitored from 2002 to 2018. Each year, 46 images were
acquired, giving 782 images in the study period. In our case, only two of all the spectral
bands were extracted from the imagery collection: band 1 (RED: 620–670 nm) and band
2 (NIR: 841–876 nm). An average of each band per zone was applied in the VIs
calculation to assure a correct spectral characterisation.
Two VIs were calculated using the same spectral bands (RED and NIR), but a different
approach in reducing the soil effect. The first one was the NDVI, calculated by the
following equation in Table 2.

24
The second index was MSAVI, proposed as an improved version of NDVI. It includes
bare soil and dead vegetation effect by adding a new variable named soil factor
adjustment (LM). LM is calculated by the following formula [3]:

2 ∗ 𝑠 ∗ (NIR − RED) ∗ (NIR − 𝑠 ∗ RED)

𝐿𝑀 = [3]
NIR + RED

where s is defined as the soil line given by a plot of RED vs. NIR brightness.
To obtain s, a set of points is extracted, characterised by the minimum NIR value within
the RED breaks (0.005) in the RED vs. NIR plot (Xu and Guo, 2013). A weighted least-
squares linear regression is fitted over these points, being s the slope of the regression
(Xu and Eckstein, 1995). This method is recommended to be implemented when non-
photosynthetic vegetation or bare soil is predominant in the scene, as it is our case in the
Mediterranean summer.
Once the s parameter was calculated for each plot, the average s was used in Equation [3]
to obtain the LM factor. Then, MSAVI was calculated by Equation [4]:

NIR − RED
MSAVI = ∗ (1 + 𝐿𝑀 ) [4]
NIR + RED + 𝐿𝑀

Two different AEMET (Agencia Estatal de Meteorología, 2019) stations were considered
to provide daily measurements of average air temperature (Tmean), minimum
temperature (Tmin), maximum temperature (Tmax) and precipitation. The
meteorological station of ZAV is settled at the centre of Ávila (40°39’33.024” N,
4°40’48.000” W) and located at 1130 m.a.s.l. In the ZMA zone, the meteorological station
is sited between Soto del Real and Colmenar Viejo (40°41’46.008” N, 3°45’54.019” W)
at 1004 m.a.s.l).

2.2.3 Inter-Annual and Intra-Annual Analysis and time-Series and VIs Phase Cross-
Correlations
On a first approach, we estimated the annual average of VIs, temperature, and
accumulated precipitation. Then, a linear fitting was conducted to analyse the VIs and
temperature annual trends. Precipitation was plotted to show the water availability along
the time. A statistical test was performed to detect significant data trends (Tarquis et al.,
2017). The slope of the linear regression was compared to 0 using a Student t-test. If the
t-estimated value was inferior to the critical t-value at the 95% level, then the slope was
not significantly different from zero.

25
In the next step, a min–max normalisation (0–1) method was applied to determine
whether annual climate variations were fluctuating with VIs behaviour. Each VI was
compared to each climate variable in both zones (ZAV and ZMA).
A more in-depth analysis was performed to characterise date-to-date VIs behaviour.
Average temperature and accumulated precipitation in an 8-day (PCP) period were
estimated based on daily meteorological data. Descriptive statistics were applied to
characterise each one of the 46 dates using box-plots charts.
Based on the VIs trend changes in the date-to-date series, we distinguished several annual
pasture phases. Individualised phase linear regressions were compared using the Chow
test to confirm each phase datasets differences (Chow, 1960). Then, linear regressions by
phase were plotted to analyse the relationship between VIs and climate variables.
Correlation Pearson’s coefficients (CR) were calculated to reveal relationships between
VIs and climatic variables series along the time. Additionally, partial correlation
coefficients (PCR) were calculated to detect the most determinant variable in the VIs
response. Then, time-series lagged cross-correlation analysis was performed to obtain an
optimal lag (ℓ) where the correlation between climate variable and VI is maximum. The
following formula [5] is applied:

𝑐𝑜𝑣(𝑥(𝑗, 𝑡), 𝑦(𝑗, 𝑡 − ℓ))

𝜌(ℓ) = [5]
𝜎𝑥(𝑗,𝑡) 𝜎𝑦(𝑗,𝑡−ℓ)

where x(j, t) are the VI values, NDVI or MSAVI, at year j and time t. The y(j, t- ℓ) are the
meteorological values, temperature, or precipitation, at year j and delayed ℓ times the lag
time, which is eight days.
Pearson’s coefficients analysis was replicated in each previously defined VI phase. Then
the expression becomes [6]:

𝑐𝑜𝑣(𝑥𝑖 (𝑗, 𝑡), 𝑦𝑖 (𝑗, 𝑡 − ℓ))

𝜌(𝑖 , ℓ) = [6]
𝜎𝑥𝑖(𝑗,𝑡) 𝜎𝑦𝑖(𝑗,𝑡−ℓ)

where xi(j,t) are the VI values at year j and time t that belong to phase i. The yi(j,t- ℓ) are
the meteorological values at phase i at year j and delayed ℓ times the lag time, which is
eight days.

26
2.3 Vegetation and climate anomalies time series.
2.3.1 Area of study, satellite data and vegetation index estimation
In the third section of this thesis, the two areas studied are the same as in section 2.1:
ZGU in Guadalix de la Sierra and ZSO in Soto del Real (Figure 1). Like in the previous
section, MOD09Q1 product from MODIS imagery was selected for this study. Study plot
reflectance was monitored from 2002 to 2018. Each year, 46 MODIS images free of cloud
were acquired, giving 782 images in the study period. The spectral vegetation indices
considered in this section were NDVI, calculated following the equation in Table 2, and
MSAVI calculated following equation [4]. Both indices were calculated for each plot and
the average of the plots was considered for the time series analysis. VIs anomalies series
(A_NDVI and A_MSAVI) were obtained by subtracting the average per date of the year
and obtaining a seasonally adjusted [7] . Annex 1 show VIs series and their respective
anomalies.

𝐴_𝑉𝐼𝑖 = 𝑉𝐼𝑖 − ̅̅̅̅

𝑉𝐼𝑖 [7]
Where i is the date of the time series, going from 1 to 46.
2.3.2 Climate data
The two different AEMET (Agencia Estatal de Meteorología, 2020) stations from section
2.1.3 were used to obtain daily meteorological data: Venturada (40° 47' 53.79" N, 3° 37'
30.49" W) for ZGU and Colmenar Viejo (40° 41' 46.008" N, 3° 45' 54.019" W) for ZSO.at
1004 m.a.s.l.. Both stations are near to the areas selected.
From these stations, a daily series of average temperature and precipitation from 2002 till
2018 were extracted. In general, Guadalix de la Sierra average precipitation and
temperature were lower than those of Soto del Real) (Table 1). To analyse the climate
effect, temperature and precipitation series were resampled from daily to an eight-day
frequency. The precipitation (PCP) represents the accumulated precipitation every eight
days, and the temperature (TEMP) represents the average temperature of each eight days.
These series were used for descriptive statistics. Following Zhao et al. (2011) work to use
temperature for RPs and CRPs, the temperature series were created with the lowest value
of daily average temperature each 8 days (TMIN).
Temperature anomalies series (A_TMIN) were obtained by subtracting the average per
date of the year in the case of TMIN [8] . In the PCP case, we preferred to subtract the
median per date of the year due to its tailed frequency distribution in some dates (A_PCP)
[9], see Annex 1.

27
 𝐴_𝑇𝑀𝐼𝑁𝑖 = 𝑇𝑀𝐼𝑁𝑖 − ̅̅̅̅̅̅̅̅
𝑇𝑀𝐼𝑁𝑖 [8]

̃𝑖
𝐴_𝑃𝐶𝑃𝑖 = 𝑃𝐶𝑃𝑖 − 𝑃𝐶𝑃 [9]

Where i is the date of the time series, going from 1 to 46.

2.4 Recurrence Plots and Recurrence Quantification Analysis

2.4.1 Recurrence analysis for vegetation complexity
Recurrence plots (RP) were introduced by Eckmann et al. (1987) and allowed to visualize
system states. In complex dynamical systems, recurrence is a concept related to the
temporal evolution of dynamical systems trajectories. To create an RP, the reconstruction
of the system states using a time-delay (τ) and an embedding dimension (m) are necessary
(Takens, 1981). Then, m represents the number of independent variables needed to
characterise the system. Delay, τ, is the minimum time lag to minimize the autocorrelation
of a time series (Proulx et al., 2015). In this step, from the original series (𝑥1 , 𝑥2 ⋯ 𝑥𝑁 ),
being N the series length, a new series is constructed whereby each 𝑥𝑖 is replaced by a
vector (𝑥⃗𝑖 ) as follows:

𝑥⃗𝑖 = (𝑥𝑖 , 𝑥𝑖+𝜏 , 𝑥𝑖+2𝜏 , ⋯ , 𝑥𝑖+(𝑚−1)𝜏 ) [10]

The result is a series of vectors:

𝑉 = {𝑥⃗1 , 𝑥⃗2 , 𝑥⃗3 ⋯ 𝑥⃗𝑁−𝜏(𝑚−1) } [11]

If N is the series length, one can only reconstruct N- τ (m-1) system states. Each possible
system state is represented by a point in the space of dimension m, named the phase space.
Plotting all the system states describes a dynamical system trajectory curve.
The RP is then a square matrix, with time on both axes, of pairwise Euclidean distances
between the reconstructed system states to which a distance threshold or radius (𝜀) is
applied (Marwan et al., 2007). Mathematically RP is defined as:

𝑹𝑖𝑗 = 𝛩(𝜀 − ‖𝑥⃗𝑖 − 𝑥⃗𝑗 ‖), 𝑥⃗𝑖 ∈ ℝ𝑚 , 𝑖, 𝑗 = 1. . . 𝑁, [12]

where N is the number of measured states 𝑥⃗𝑖 𝑜𝑟 𝑥⃗𝑗 , i and j are the indexes of the states, 𝛩
is the Heaviside step function (i.e. 𝛩(x) = 1, if ‖𝑥⃗𝑖 − 𝑥⃗𝑗 ‖ ≤ 𝜀, and 𝛩(x) =0 otherwise),
‖·‖ is a norm, and 𝜀 is a threshold previously defined based on the time-series properties.
In this study, the phase space trajectories are based on the Euclidean distance between 𝑥⃗𝑖

28
and 𝑥⃗𝑗 of the series. If 𝑹𝑖𝑗 = 1 at a time (i, j), is marked as a black dot in the position (i,
j). Otherwise, if 𝑹𝑖𝑗 = 0 recurrence states will be represented as white dots.
In this example, it is visualised a numerical case to construct a RP. An artificial series of
random numbers has been created = {2, 1, 8, 8, 9, 0, 2, 7, 7, 10, 1, 5, 5, 9, 9, 9, 4, 7, 6, 1,
8, 0, 1, 9, 4} and an absolute distance matrix (DM) of the series with itself is built as
presented in Figure 3. The elements of the line of identity (LOI), or principal diagonal of
the matrix, are zero as it is the differences of the two elements of the series with the same
value. Applying a threshold 𝜀 on the DM, an RP is obtained as showed in Figure 3B (𝜀 =
3) and Figure 3C (𝜀 = 1). Values inferior than 𝜀 are represented as coloured cells showing
the recurrence of the system. Values higher than 𝜀 are represented as white cells. The
number of coloured cells varies depending on the threshold value and the proportion of
coloured cells in the RP is defined as recurrence rate (RR).
In the RP of this example, 𝑚 was equal to 1, and τ was 0. In the case that 𝑚 =4 and τ =5,
we reconstruct the system state from the above series yielding the successive states of 𝑥⃗1
= [2, 0, 1, 9]; 𝑥⃗2 [1, 2, 5, 4]; 𝑥⃗3 [8, 7, 5, 7]; ⋯ 𝑥⃗10 [10, 9, 1, 4]. In a Vector Space of
dimension 4, the representation of these vectors will be the trajectories described by the
system in the phase space.

29
 A B C
Figure 3. Random series distance matrix (DM) and recurrence plot (RP) example. The length of the series is of 25 points. Absolute distance is calculated and represented as a
coloured matrix (A). Different thresholds were applied to obtain RPs: 𝜀 < 3 (B) and 𝜀 < 1 (C).

30
Once the embedding dimension and time delay are chosen, an RP is created, and the
obtained patterns characterised it. In the RP, homogeneous distributions of points are
related to stationary stochastic processes. Whereas long diagonal structures lead to
periodic behaviours. The RPs of a white-noise stochastic and periodic sine series are
shown in Figure 4A and 3B, respectively. Structures as single points are non-persistent
states in time, as in Figure 4A. On the other hand, continuous diagonal lines represent the
same phase space visited by a trajectory at different times (Marwan et al., 2007). As
shown in Figure 4B, sine series RP is characterised by long repeating black lines,
representing the periodicity and frequency of the sine series.

A) B)

Figure 4. Artificial series RPs adapted from (Zhao et al., 2015): A) White-noise series, embedding
dimension (m) = 1, delay (τ) = 0, recurrence rate (RR) = 5%, B) Sine series, m = 2, τ = 6, RR = 5%. Both
series have a length of 387 data points.

RPs were difficult to analyse visually due to their dependency on the number of points in
a screen. Webber and Zbilut (Webber and Zbilut, 2005) introduced a quantification
method to RP, namely Recurrence Quantification Analysis (RQA), based on quantifying
the small-scale structures in RPs.
Several measures of complexity have been proposed, in this work, RQA formulas are
obtained from (Marwan et al., 2007); for further details, it is recommended to revise the
work mentioned above. The first step to obtaining RQA measures is the calculation of RP
histogram (P) from the total number of diagonal lines, of length (l) [13], and the total
number of vertical lines of length (v) [14]. The following formulas define these
histograms:

31
 𝑁 𝑙−1
𝑃(𝑙) = ∑ (1 − 𝑅𝑖−1,𝑗−1 (𝜀))(1 − 𝑅𝑖+1,𝑗+1 (𝜀)) ∏ 𝑅𝑖+𝑘,𝑗+𝑘 (𝜀), [13]
𝑖,𝑗=1 𝑘=0

𝑁 𝑣−1
𝑃(𝑣) = ∑ (1 − 𝑅𝑖,𝑗 )(1 − 𝑅𝑖,𝑗+𝑣 ) ∏ 𝑅𝑖,𝑗+𝑘 [14]
𝑖,𝑗=1 𝑘=0

Where i and j are the row and columns of the matrix R, k is the diagonal ranging from 0
till l-1 length in Equation [11], and k is the vertical ranging from 0 to v-1 in Equation [12].
DET is the proportion of recurrent points assembling diagonal structures of all recurrence
points. Therefore, it can be interpreted as the predictability of the system's state but not
of the process. Temporal predictability can be low even if determinism is high (Proulx et
al., 2015). DET is calculated by [15]:

∑𝑁
𝑙=𝑙𝑚𝑖𝑛 𝑙𝑃(𝑙)
𝐷𝐸𝑇 = , [15]
∑𝑁
𝑖,𝑗 𝑅𝑖,𝑗

Where lmin is the minimal length to consider a diagonal line, in this case, as suggested by
(Webber and Zbilut, 1994), it was established as 2.

LT is the average time passed when two trajectories are close to each other, and it is
calculated by the following equation [16]:

∑𝑁
𝑙=𝑙𝑚𝑖𝑛 𝑙𝑃(𝑙)
𝐿𝑇 = , [16]
∑𝑁𝑙=𝑙𝑚𝑖𝑛 𝑃(𝑙)

Shannon’s entropy or ENT points out the probability to find the same trajectory of length
l in the whole RP. The consequent formula is applied [17]:

𝑁
𝐸𝑁𝑇𝑅 = − ∑ 𝑝(𝑙)𝑙𝑛𝑝(𝑙) [17]
𝑙=𝑙𝑚𝑖𝑛

Laminarity or LAM represents the number of vertical structures, independent of their

length (Marwan and Kurths, 2002). It represents an overall measure of the signal stability,
and the following equation [18] calculates it:

∑𝑁
𝑣=𝑣𝑚𝑖𝑛 𝑣𝑃(𝑣)
𝐿𝐴𝑀 = [18]
∑𝑁
𝑣=1 𝑣𝑃(𝑣)

32
An extension of RP is the Cross Recurrence Plot (CRP), which checks the co-occurrence
of similar states of two different time series. Mathematically, a CRP between two
different series, (𝑥1 , 𝑥2 , ⋯ 𝑥𝑖 , ⋯ 𝑥𝑁 ) and (𝑦1 , 𝑦2 , ⋯ 𝑦𝑗 , ⋯ 𝑦𝑁 ), is calculated by:

𝑪𝑖𝑗 = 𝛩(𝜀 − ‖𝑥⃗𝑖 − 𝑦⃗𝑗 ‖), 𝑥⃗𝑖 , 𝑦⃗𝑗 ∈ ℝ𝑚 , 𝑖, 𝑗 = 1. . . 𝑁, [19]

This equation [19] is similar to equation [12] except that we are comparing two different
series corresponding to two different variables. CRPs compare the occurrences of similar
states of two systems. They can be used to analyse the similarity of the dynamical
evolution between two different series or to study the time-relationship of two similar
series whose timescale differ (Marwan and Kurths, 2002). The RQA can be applied to
CRP with the same meaning that with RP.
Furthermore, RQA was extended by computing the diagonal-wise recurrence
quantification profile (Coco and Dale, 2014). The recurrence rate around the line of
coincidence (LOC) and the surroundings time lags was calculated to measure the two
time-series coupling as a lag function. The maximum number of lags to be analysed was
six, the same as the cross-correlation method.
CRQA R package (Coco and Dale, 2014; Marwan, 2007) was used to construct RP, obtain
RQA measures and compute diagonal-wise recurrence profile. First, the VIs series were
normalised using a z-score normalisation; then, the distance matrix was rescaled based
on the maximum value following the recommendations of Webber and Zbilut (Webber
and Zbilut, 2005). Optimizeparam function is then computed to find the three parameters’
optimal values (τ, m, and ε). The delay (τ) was found by obtaining the local minimum
where mutual information drops to both series (Fraser and Swinney, 1986). The
embedding dimension (m) was calculated by the false nearest neighbours’ algorithm
(Cao, 1997). The threshold ε was estimated by an iterative process based on the time-
series’ standard deviation (SD). In this cases, ε was limited to 5% of the recurrence rate
(RR) in all the cases. When multiples values of m, τ, were obtained by the optimisation,
the maximum of them was selected as the optimal value to apply in the construction of
CRPs.
The quantification of RP and CRPs structures was calculated with the Crqa function using
the values obtained from the optimisation function. Then, the drpfromts function was
computed to plot diagonal-wise recurrence profiles in the RPs and CRPs.

33
2.4.2 Recurrence analysis for anomalies time series.
As it is said in the above section, recurrence plots (RP) allow visualising system states in
the phase space. In complex dynamical systems, recurrence is related to the temporal
evolution of dynamical systems trajectories in the phase space. In this section, original
and seasonally adjusted series (anomalies series) were studied with RP analysis. In
addition, CRP analysis was applied using VIs and climatic original and anomalies series.
Again, the CRQA R package (Coco and Dale, 2014), based on the Cross Recurrence Plot
Toolbox developed by Marwan (Marwan, 2007), was used to construct RP and CRP,
obtaining their corresponding RQA measures. First, the Optimizeparam function was
computed to obtain the optimal values of the three parameters (m, τ, and 𝜀). The threshold
𝜀 is estimated by an iterative process based on the standard deviation of the time series
and a limit in the RR value (Thiel et al., 2002). Based on different studies of RR, a
maximum value of 2.50 % was set (Marwan, 2007) in this case. The series analysed were
normalised using z-score, and the distance matrix (DM) was rescaled based on the
maximum value following the recommendations by Patro & Sahu (Patro and Sahu, 2015)
and Webber & Zbilut (Webber and Zbilut, 2005). Then, the Crqa function was computed,
using the optimization values obtained earlier, to calculate the complexity measures.
Finally, the drpfromts function was computed to plot diagonal-wise recurrence profiles
in the RPs and CRPs.

2.5 Threshold values optimization (TVO) to reduce soil spatial variability

in a wheat (Triticum aestivum L.) crop with UAV imagery
2.5.1 Study area and experimental design
The study was conducted at La Chimenea farm station (40º 04 ′ N, 03º 32′ W, 550
m.a.s.l.), near Aranjuez (Madrid, Spain), throughout the wheat growing season of 2019.
The study site was uniform, with a slope lower than 1%. According to the World reference
base for soil resources (2014), the soil is classified as Haplic Calcisol, with a medium
organic matter content (topsoil organic C 1.01 g kg-1), a pH of approximately 8.1, and a
silty clay loam texture with low stone content. The mean annual temperature is 14.2 ° C
with an average rainfall of 373 mm. According to the Koppen classification, the area was
classified as semi-arid cold (Bsk).

34
Figure 5. Location of the trial and sectors (S1, S2, S3, S4) over a normalized difference vegetation index
(NDVI) map retrieved from the UAV platform during stem elongation in 2019. The 133 plots monitored in
each flight are shown. Coordinate system: ETRS89 / UTM zone 30N (EPSG: 25830)

Four different previous experiments involving 133 plots randomly assigned to different
N and water doses conducted in a field with circular pivot (220 m in radius) (Figure 5)
were used in the current optimization work. Each plot was georeferenced with the real-
time kinematic (RTK) technique based on the National Geodetic Network of Reference
Stations GNSS (ERGNSS), using the stable Aranjuez (Madrid) and Sonseca (Toledo)
stations, with a Topcon HiPer Pro receptor® (Topcon Singapore Holdings Pte. Ltd.,
Singapore).
All plots used in this trial were obtained from different experiments fully explained in the
works of Pancorbo et al. (2021b) and Raya-Sereno et al. (2022). Each experiment,
referred to in this study as “sector,” was designed with a varying dose of fertilization and
irrigation. Thus, the primary purpose behind the design of this work Table 5 was to
maximize the intravariability of the study area. All plots were sown on the same date
(15 November 2018) with the same winter wheat cultivar (Triticum aestivum L. cv.
Nogal) at a rate of 220 kg seeds ha-1. Nitrogen fertilization was hand-broadcasted with
ammonium sulphate nitrate (ASN, 26% N) in two applications: half at the beginning of
tillering (GS22) and half at stem elongation (GS35). Phosphate and potassium were
applied to all plots to ensure no wheat deficiency based on the soil analysis results.

35
Table 5. Summary of the sectors. Number and size of plots, fertilization and irrigation treatments.

Number and size Fertilization

Sector Fertilization (2019) Irrigation
of plots (2018)
-1
N0 (0 kg N ha ),
46 plots
S1 No N1 (150 kg N ha-1), Yes
(3.5 m × 8 m)
N2 (190 kg N ha-1)
N0 (0 kg N ha-1),
16 plots
S2 N1 (50 kg N ha-1), No No
(25 m × 25 m)
N2 (100 kg N ha-1)
N0 (0 kg N ha-1),
42 plots
S3 No N1 (50 kg N ha-1), No
(8.3 m × 8.3 m)
N2 (90 kg N ha-1)
N0 (0 kg N ha-1),
29 plots N1 (42 kg N ha-1), Yes (16 plots)
S4 No
(22 m × 22 m) N2 (93 kg N ha-1), No (13 plots)
N3 (142 kg N ha-1)

Sector S1 was composed of 16 plots (8 m × 10.5 m) that were fertilised with different
nitrification inhibitor treatments in 2018. The following year (2019), each plot was
subdivided into three subplots that received different fertilizer doses, giving a total of 48
plots (3.5 m × 8 m). The N doses established were 0, 150, and 190 kg N ha-1. Sector S2
was made up of 16 plots (25 m × 25 m). In the previous year (2018), they were also
cultivated with wheat that received 0, 50, 100, and 150 kg N ha-1. In 2019, the plots did
not receive any fertilization treatment; therefore, all were considered control plots. Sector
S3 was composed of 16 plots (25 m × 25 m) cultivated with wheat in 2018 and that
received the same N doses as sector S2. In 2019, each plot was subdivided into three
subplots that received 50 kg N ha-1, 90 kg N ha-1, or no N fertilizer. Therefore, the total
number of plots in this sector (8.3 m × 8.3 m) was 48. Finally, in sector S4 there were 32
plots (22 m × 22 m) that were distributed into four different N levels: 0, 42, 93, and 142
kg N ha-1. From the total of 144 plots, 133 plots were finally used in this optimization
work.
Additionally, to evaluate the effect of water availability, two water levels were considered
in the experiment. Sector S1 and half of the plots of sector S4 were irrigated during spring
using a center-pivot system; whereas sector S2 and S3 did not receive irrigation. More
details on the irrigation in sector S1 and S4 can be found in Pancorbo et al. (2021b) and
Raya-Sereno et al. (2022).

2.5.2 Wheat measurements, imagery acquisition and postprocessing

All plots were harvested on the same date (15 July 2019) and the wheat traits were
measured following the same procedure. An experimental combiner registered wheat
yield by harvesting a central 1.4-m-wide strip and leaving a 1-m buffer at the beginning

36
and end of each plot. The grain N concentration was determined using the Dumas
combustion method (LECO FP-428 analyser, St. Joseph, MI, USA) and then transformed
into grain protein content by multiplying N concentration 6.25 times (Kruger, 2009).
Finally, the N output (kg N ha−1) was calculated as the product of the grain yield (kg ha−1)
and the grain N concentration (% N). Further information about the laboratory and field
measurements can be found in Pancorbo et al. (2021b) and Raya-Sereno et al. (2022).
Multispectral imagery was acquired on three dates, corresponding to the different growth
stages of wheat: beginning of stem elongation (GS32, 11 March 2019), final stem
elongation (GS39, 12 April 2019), and flowering (GS65, 13 May 2019). All the images
were taken close to solar noon local time (GMT+1). Surface reflectance was collected
with an UAV M-600 PRO (Shenzhen, China), a six-rotor flying platform capable of
following a pre-designed track. The flight was carried out by optimizing the UAV
trajectory in parallel with an automatic pilot function according to the planned project,
using a ground sampling distance (GSD) of 4.65 cm/px. The drone hexacopter used the
dual-frequency GPS system A3 Pro and a Romin-MX stabilizer for the cameras with an
inertial system for positioning the photo-centres.
The UAV was equipped with a Micro-MCA multispectral sensor, Tetracam Inc.
(Shenzhen and Chatsworth, CA, USA), a configurable camera matrix from 5.2 to 15.6 MP
of six multispectral channels; the sensor systems installed are 1.3 MP CMOS sensors.
The camera had six independent image sensors that captured narrow wavelength bands
centered at blue (490 nm), green (550 nm), red (671 nm), red edge (700 nm), far red (760
nm), and near infrared (NIR; 800 nm) with a bandwidth of 10.0 ±2 nm. Reflectance was
calibrated by an incident light sensor (ILS) integrated with the camera, which corrected
the incident radiation for each shot in each band during the multispectral survey. The
result of each shot was a high-resolution multispectral image composed of six reflectance
channels: (490 nm), (550 nm), red (671 nm), red edge (700 nm), far red (760 nm), and
NIR (800 nm). These images were processed with the Tetracam PixelWrench2 software
(Chatsworth, CA, USA), which aligned the image and released a multi-frame and single
TIF image file.
The images were processed with PixelWrench2 for correction and calibration, and then
Agisoft PhotoScan (St. Petersburg, Russia) was used to make the orthomosaic image of
the entire plot. The ortho-image and R statistical software (version 4.0.3; R Core Team,
2021) were used to extract and calculate VIs from the different experimental plots and to
perform statistical analysis. The VIs calculated were NDVI and MSAVI using the same

37
equation as in the previous sections, and two new indices common in characterization of
green vegetation studies. The Normalized different red-edge index (NDRE) that
incorporates into the equation the red-edge band, particularly sensitive to chlorophyll
pigment concentration Equation [20]. The Blue-red index (BRI1) has been broadly used
when only RGB cameras were available Equation [21].

NIR (860 nm) − 𝑅𝐸𝐷𝐸𝐷(700 nm)

𝑁𝐷𝑅𝐸 = [20]
NIR (860 nm) + 𝑅𝐸𝐷𝐸𝐷(700 nm)

BLUE(490 nm)
𝐵𝑅𝐼1 = [21]
𝑅𝐸𝐷(660 nm)
Where REDED is the red-edge reflectance (770 nm) and BLUE is the blue reflectance
(490 nm).

2.5.3 Threshold value optimization (TVO)

As a preliminary analysis, we estimated the histograms of the VI distribution in GS32.
The main objective was to detect anomalies in the distribution of VIs and to identify the
potential VIs that can be optimised using the segmentation method.
The main idea behind TVO was to optimize the removal of the soil background effect
based on the VI distribution. We established percentile probabilities from 0% to 100%
with an interval of 5%, as smaller increments did not decrease the RMSE and increased
the computational time. We then estimated the percentile value (xi) of the VI distribution
in each plot of the experiment with the following formula Equation [22]:

𝑣𝑣 ∗ 𝑖
𝑥𝑖 = [22]
100

where 𝑣𝑣 are the VI values ordered from minimum to maximum, and i is the percentile
probability. Then, the minimum xi of all plots was selected as the threshold value Tv. We
computed an iterative process using all the percentile probabilities to obtain a list of
threshold values (sequentially 𝑇𝑣1 , 𝑇𝑣2 … 𝑇𝑣𝑛 ). The data are presented in Annex 2.
A single Tv of the list was used to segment the VI distribution of the plots. Once
segmented, a linear regression was performed between the median values of each plot
and the measurements of the wheat traits. The method was repeated with each T v of the
list to obtain the optimal threshold (Topt), where the coefficient of determination (R2) is
maximised and the RMSE Equation [23] is minimised.

38
 𝑛
1
𝑅𝑀𝑆𝐸 = √ 𝑥 ∑(𝑃𝑖 − 𝑂𝑖)2 [23]
𝑛
𝑖=1

Pi and Oi are the predicted value and the measured value, respectively, and n is the number
of samples. The main objective of the analysis was to compare the predictive performance
of each VI segmentation and to obtain the Topt. The same analysis was repeated with all
VIs, wheat traits, and flights to compare and evaluate the best correlated VI for each flight
and its corresponding wheat trait.

39
 3 ASSESMENT OF DROUGHT
INDICES ON DIFFERENT TIME
SCALES: A CASE IN SEMIARID
MEDITERRANEAN GRASSLANDS

40
3.1 Objectives, area of study and materials
The aims of this chapter were: i) to understand the complexity of vegetation time series
related to climate fluctuations and drought episodes through VIs (VHI and SVHI) and
precipitation indices (SPI and SPEI) and ii) to determine which index performed better
and the most suitable time scale.
Two study areas were considered in section 3, ZGU in Guadalix de la Sierra and ZSO in
Soto Del Real, both in Madrid (Spain). ZSO was composed of four pixels of 250×250 m2
between (3º45'00" W, 3º46'00" W) and (40º43'00", 40º44'00" N). The exact number of
pixels was selected in the ZGU area delimited by (3º38'00" W, 3º39'00" W) and
(40º46'00" N, 40º47'00" N). Each one of these sites represents an area of approximately
1 km2.
In this case, it was used the MOD09Q1 product which offers imagery at 250 m resolution
of the RED (620-670 nm) and NIR (841-876 nm) surface reflectance. Observations from
the same orbit are composited by observational coverage, and the highest coverage
observations are stored in 8 days. Two VIs were selected, NDVI, an adequate estimator
of the photosynthetic activity and vegetation growth, and the vegetation health index
(VHI), an improved version of NDVI, allowing for the detection of drought and non-
drought periods. Two components define VHI mathematically. The first represents the
vegetation component, named the vegetation condition index (VCI), and the second
corresponds to the land surface temperature. The latter is defined as the temperature
condition index (TCI). It is essential to note the importance of the α coefficient, which
defines the weight of each component in the VHI formulation.
Daily meteorological data were obtained from two different AEMET stations.
Precipitation and temperature series were resampled to an eight-day frequency to be
analysed on the same time scale as the VHI time series. Precipitation (PCP) is the
accumulated precipitation every eight days, and temperature (TEMP) is the average
temperature in the same period.
Four precipitation and VIs were employed in a correlation analysis: SPI, SPIE, VHI and
SVHI, which use the same standardisation process of SPI and SPIE, applied to a VHI
time series. Then, based on a bibliography review, SPI, SPEI, VHI, and SVHI are
classified to define four different drought classes.

41
3.2 Descriptive Statistics
We estimated the VHI time series to reveal the drought epochs in the vegetation. VHI
was composed of two elements. VCI, which characterised the vegetation behaviour
(Figure 6a) and TCI, which represented the land surface temperature (Figure 6b).

(a)

(b)

Figure 6. Vegetation Condition Index (VCI) (a) and Temperature Condition Index (TCI) (b) statistics from
2001 to 2020, along the year in each zone: ZGU = Guadalix de la Sierra, ZSO = Soto del Real. The Y-axis
shows a box and whisker plot of VCI and TCI on each date. The error bars indicate the maximum and
minimum values; horizontal lines indicate the median value, and the box indicates the 25% and 75%
percentiles. Q1, Q2, Q3 and Q4 are the quarters of the year. The blue and red line are the medians of each
zone.

It was important to notice that the higher the value of TCI and VCI was, the better the
vegetation condition was. In one hand, we observed that VCI is higher in ZGU during Q2
and Q3, during the most drought vulnerable seasons in the year (spring and the summer).
On the other hand, the TCI of ZGU is higher in the entire year, showing that vegetation
is in a better state than ZGU. Overall, we could expect a minor number of drought events
in ZGU compared to ZSO.
The dispersion and evolution of the VHI through the different quarters showed a similar
pattern for the two zones under study (Figure 7).

42
Figure 7. VHI vegetation index statistics, from 2001 to 2020, throughout the year in each zone: ZGU =
Guadalix de la Sierra, ZSO = Soto del Real. The Y-axis shows a box-and-whisker plot of VHI (Vegetation
Health Index) on each date. The error bars indicate the maximum and minimum values; horizontal lines
indicate the median values, and the box indicates the 25% and 75% percentiles. Q1, Q2, Q3 and Q4 are the
quarters of the year. The blue and red lines are the medians of each zone.

The summer presented the lowest values, which began to increase in Q3 and peaked in
Q4. The variation in VHI data throughout the year ranged from 10 to 82, with the highest
dispersion in Q2 and Q4. Generally, the VHI values at ZGU were higher than those of
ZSO in the Q2 and Q3 quarters. This difference is probably because of the difference in
temperature in both sites (Table 1). However, during the middle of Q4 and Q1, the VHI
values in ZSO were higher than in ZGU.
A similar quarterly pattern obtained in VHI time series is shown in the precipitation data
(Figure 8a). Both zones shared similar behaviour in the volume and dispersion of
precipitation. Although precipitation appeared to be erratic, it presented a significant
dispersion in Q4 and Q2. Meanwhile, the lowest data dispersion was observed in Q3. The
highest temperature dispersion was found in spring (Figure 8b). Temperature variation at
both sites was around 1ºC. Differences were more noticeable during Q3 (mainly summer)
and at the beginning of Q4 (October).

43
 (a)

(b)

Figure 8. Accumulated 8-day precipitation (PCP) (a) and temperature (TEMP) (b) statistics throughout the
year, including from 2001 to 2020, in each zone: ZGU = Guadalix de la Sierra, ZSO = Soto del Real. The
Y-axis shows a box and whisker plot of a) PCP_ZGU, PCP_ZSO, and b) TEMP_ZGU and TEMP_ZSO on
each date. The error bars indicate the maximum and minimum values, the horizontal lines indicate the
median values, and the box indicates the 25% and 75% percentiles. Q1, Q2, Q3 and Q4 are the quarters of
the year.

Summarising the annual cycle in both zones, precipitation and temperature patterns were
quite similar. The lack of precipitation was noticeable during summer; autumn rain events
were synchronised with the increase in VHI. A similar result is obtained at the beginning
of spring when precipitation and VHI increased simultaneously in both zones.

3.3 Standardised Indices Time Series

The SVHI pattern showed a better similarity with the meteorological drought indices in
the quarterly division than on the monthly time scale (Figure 9 and Figure 10). There is
no easily perceived remarkable difference between SPI and SPEI. We could also note that
the ZSO time series might be better correlated than the ZGU time series. Especially at the

44
 end of the time series, we could detect that SPI and SPEI are better overlapped with SVHI
in ZSO than ZGU.

SPI-SVHI SPEI-SVHI
Monthly
Quarterly

Figure 9. SVHI (Standardised Vegetation Index), SPI (Standardised Precipitation Index), and SPEI
(Standardised Precipitation Evapotranspiration Index) time series at different time scales in ZGU (Guadalix
de la Sierra). The red dotted lines mark ±1 values.

SPI-SVHI SPEI-SVHI
Monthly
Quarterly

Figure 10. SVHI (Standardised Vegetation Index), SPI (Standardised Precipitation Index), and SPEI
(Standardised Precipitation Evapotranspiration Index) time series at different time scales in ZSO (Soto del
Real). The red dotted lines mark ±1 values.

We could observe in Figure 11 the effects of standardisation and the use of a variable
alpha in the correlation between SVHI and drought indices. We recognised a parabolic
trend on the monthly scale and a logarithm shape on the quarterly time scale in both zones.
We could detect that both time scales presented a remarkable maximum, being (α = 0.2)
in the monthly’s monthly scale and (α = 0.6) in the quarterly case.

45
 ZGU ZSO

Monthly
Quarterly

Figure 11. Correlation between meteorological drought indices (SPI and SPEI) and standardised vegetation
index (SVHI) at different time scales (M: Monthly, Q: Quarterly) in ZGU (Guadalix de la Sierra) and ZSO
(Soto del Real). The alpha corresponds to the α value in Table 2.

3.4 Correlation Analysis

A considerable difference in the correlation values between the entire time series of the
VHI, SVHI, and drought indices (SPI, SPEI) at different time scales was observed (Table
6). The correlation increased when standardisation was applied to the VHI in both zones.
At the same time, we confirm the supposition exposed in the time series plots, where the
quarterly division seemed to be the most suitable time scale to analyse drought indices.

Table 6. Pearson series correlation values between VHI (Standardised Vegetation Index) and SVHI with
SPI (Standardised Precipitation Index) and SPEI (Standardised Precipitation-Evapotranspiration Index) at
different time scales. M is the monthly time scale; Q is the quarterly time scale.

Zone Scale VHI-SPI VHI-SPEI SVHI-SPI SVHI-SPEI

M 0.14 0.13 0.40 0.42
ZGU
Q 0.30 0.30 0.64 0.65
M 0.13 0.14 0.41 0.43
ZSO
Q 0.28 0.30 0.69 0.71

46
The correlation between SVHI and SPI revealed that the most correlated months were
March and May in both zones (Table 7). This fact is repeated in SPEI, in exception of
ZGU, which showed the higher correlation in May and September. The most correlated
seasons were spring (April-June) and winter (October-December) in both zones on the
quarterly scale, achieving values greater than 0.68 in all cases. Nevertheless, we could
still observe differences between zones. Generally, ZGU obtained values smaller than
ZSO. However, it is interesting to note that the trend reversed in (Q3) mainly summer
months when ZGU obtained higher values than ZSO in both indices.
Table 7. Correlation values between SVHI (Standardised Vegetation Health Index), SPI (Standardised
Precipitation Index), and SPEI (Standardised Precipitation Evapotranspiration at different time frames in
both zones: ZGU (Guadalix de la Sierra) and ZSO (Soto del Real). M is the monthly scale; Q is the quarterly
scale.

SVHI-SPI
Jan. Feb. Mar. Apr. May June July Aug. Sept. Oct. Nov. Dec.
M 0.19 0.33 0.63 0.14 0.63 0.34 0.37 0.50 0.59 0.07 0.28 0.66
ZGU
Q 0.51 0.69 0.66 0.72
M 0.18 0.35 0.72 0.30 0.69 0.47 0.05 0.54 0.64 0.32 0.36 0.27
ZSO
Q 0.72 0.80 0.56 0.69
SVHI-SPEI
Jan. Feb. Mar. Apr. May June July Aug. Sept. Oct. Nov. Dec.
M 0.21 0.21 0.64 0.17 0.69 0.43 0.47 0.45 0.67 0.12 0.29 0.66
ZGU
Q 0.48 0.70 0.69 0.73
M 0.18 0.26 0.73 0.32 0.78 0.57 0.24 0.50 0.66 0.30 0.40 0.28
ZSO
Q 0.68 0.83 0.61 0.74

Overall, the previous time series plots and correlation analysis support the idea that the
quarterly time scale is the most suitable for analysing the studied semiarid Mediterranean
grasslands.

3.5 Drought Classification

The drought classification suggested that ZGU suffered a minor number of drought events
compared to ZSO (Table 8). As we could observe, the SPI did not show a notable
difference between zones, neither on the monthly nor quarterly time scales. SPEI and
SVHI showed better results, especially on the quarterly time scale, where moderate
drought increased compared to SPI. However, the VHI was the best approach to
distinguish between both zones. The VHI showed a remarkable difference in the case
number of extreme drought cases between both zones monthly.

47
Meanwhile, on the quarterly scale, the critical point was the number of severe drought
cases. In general, we speculate that LST influenced the TCI and VHI values. Therefore,
VHI was able to better distinguish between both zones. However, it is important to note
that SVHI was better correlated with SPI and SPEI at meteorological drought periods.

Table 8. Drought classified data frequency of drought index (SPI, SPEI), non-standardised vegetation
health index (VHI), and standardised vegetation health index (SVHI) in both zones, Guadalix de la Sierra
(ZGU) and Soto del Real (ZSO). M is the monthly scale. Q is the quarterly scale.

Zone Scale Class drought SPI SPEI VHI SVHI

No drought 202 201 171 205
Moderate drought 17 30 48 19
M Severe drought 12 9 20 9
Extremely drought 9 0 1 7
TOTAL 240 240 240 240
ZGU
No drought 71 68 57 68
Moderate drought 6 9 21 7
Q Severe drought 2 2 2 2
Extremely drought 1 1 0 3
TOTAL 80 80 80 80
No drought 203 193 162 205
Moderate drought 17 36 36 22
M Severe drought 13 11 31 7
Extremely drought 7 0 11 6
TOTAL 240 240 240 240
ZSO
No drought 73 65 49 66
Moderate drought 4 12 15 10
Q Severe drought 2 2 16 1
Extremely drought 1 1 0 3
TOTAL 80 80 80 80
3.6 Discussion
Vegetation indices are an effective indicator of grassland cover. Their response is highly
interrelated with climate and agricultural conditions (Almeida-Ñauñay et al., 2021). In
fact, multiple authors (Liu et al., 2017; Martín-Sotoca et al., 2019; Wu et al., 2019) have
reported a clear relationship between grasslands vegetation and water availability. In this
line, the amount, variability, and timing of precipitation and temperature are determining
factors of grassland production.
Regarding VIs, when precipitation rise, generally, the VI value increases with a potential
delay between both of them (Fabricante et al., 2009). Our results reach the same
conclusion, VCI and VHI showed greater values during the autumn and spring transition
phases and achieved low stable values during winter and summer, when temperature and
precipitation are limiting. Similar trends for MODIS NDVI were reported by (Sanz et al.,

48
2021a; Wang et al., 2013), suggesting that NDVI is restricted, as well as grassland
vegetation, during the summer and winter seasons.
The precipitation amount and dispersion increased during autumn (Figure 8a); as a result,
VIs spread simultaneously in the same time frame (Chandrasekar et al., 2010; Iglesias,
2003). This fact could be explained because of the rainfall patterns that define the
temporal variability of soil moisture and, in turn, vegetation activity in soils (Harper et
al., 2005). At both zones, the precipitation pattern is similar. However, the ZSO VIs were
higher than in the ZGU, suggesting that the temperature difference could explain the
decrease in VIs due to cold stress induced in grassland vegetation during winter and
drought stress during summer (Bonan, 2016).
Analysing the effect of temperature on the VIs, during autumn, the VIs (VCI and VHI)
in ZSO indicate a higher increase with similar precipitation, most likely due to a higher
temperature in ZSO. At the start of spring, the VIs in ZGU show a remarkable increase
whereas, in ZSO the VIs remained constant. This fact is reflecting the growth of biomass
as the temperature increases, producing the Mediterranean spring herbage flush (Bugalho
and Abreu, 2008). In the second part of the spring, the VIs decreased more quickly in
ZSO than in ZGU. At the end of spring, both zones achieved similar VIs. However, the
value of ZSO was slightly lower during the summer. The WHC (water holding capacity)
suggest that the ZGU soil is able to retain more water compared to the ZSO soil. Thus,
the soil mixed interactions between moisture and lower temperatures in ZGU might
explain the difference with ZSO during spring and summer. In general, soil moisture is
defined as a connection element between vegetation growth and precipitation, regulating
rainfall storage and availability, therefore determining vegetation dynamics (Laio et al.,
2001; Quemada, 2005; Rodriguez-Iturbe, 2000; Scanlon et al., 2005).
Standardised vegetation index series pattern was more distinctive on the quarterly time
scale than in the monthly scale. These results are in line with those obtained by (Marini
et al., 2019). We speculate that the main reason is the time scale increment (from 1 month
to 3 months). This increase allows changing the number of drought events as the time
series becomes smoother with fewer abrupt changes. As the reference period rises, the
aggregate rainfall value changes from one month to three months; hence, the variations
are minor and less significant (Mishra et al., 2009).
It has been thoroughly discussed the optimal time scale to be used in the SPI and SPEI
studies. At monthly scale Pei et al. (2019) work suggest that vegetation appears to show
a more obvious response to hydrothermal conditions. Whereas quarterly time scale seems

49
to reveal seasonal trends and patterns (3-months) between drought and vegetation
response. In our case, we could observe that the highest values were obtained in the
quarterly time scale. This in line with (Zhao et al., 2018) research, revealing that grassland
cover showed the highest correlation with the quarterly SPEI. Although, other works
suggest that monthly and annual time scales might be more explanatory (Liu et al., 2017;
Pei et al., 2020). Thus, it is essential to consider the influence of the local conditions in
the vegetation response to drought.
Our results suggest that spring and autumn, and particularly May and September, are the
most correlated time frames throughout the year. Similar conclusions were drawn from
the (Hadri et al., 2021) work that obtained higher correlations in the crop growth period.
In our case, spring, and autumn are the seasons when the NDVI reaches its peak.
All standardised indices showed significant indicators of SVI status in the critical phases,
spring, and autumn. SPEI stands out over SPI; this probably results from the addition of
evapotranspiration to the analysis. This result aligns with research (Zarei et al., 2021),
where SPEI showed better results than SPI. Therefore, we should consider the growth
period of grassland vegetation in these zones, and this reaffirms the utility of standardised
indices in the critical growth phases.

50
4 THE VEGETATION–CLIMATE
SISTEM COMPLEXITY
THROUGH RECURRENCE
ANALYSIS

51
4.1 Objectives, area of study and materials
The aims of this chapter were: i) to determine if the VIs and recurrence techniques were
an adequate tool to picture the complexity of the relationships between vegetation and
climate in grasslands, and ii) to assess the most suitable VI to use in this case.
In section 4 of this thesis, two study areas in the opposite hillsides of Sierra de
Guadarrama were studied; one located in Tornadizos de Ávila, Ávila (ZAV) and the other
in Soto Del Real, Madrid (ZMA). ZAV zone was composed of three pixels of 500 × 500
m between (3°45’00” W, 3°46’00” W) and (40°43’00” N, 40°44’00” N), while ZMA was
composed of three pixels (500 × 500 m) enclosed between (4°32’00” W, 4°33’00” W) and
(40°37’00”, 40°39’00” N).
As in the previous section, a product from the MODIS collection was selected. In this
case, MOD09A1 was chosen for this study. This product is a level-3 composite of 500-m
resolution, and the best pixel observation was chosen within eight days. The study plots'
reflectance was monitored from 2002 to 2018. Each year, 46 images were acquired,
giving 782 images in the study period. In our case, only two spectral bands were extracted
from the imagery collection: band 1 (RED: 620–670 nm) and band 2 (NIR: 841–876 nm).
An average of each band per zone was applied in the VIs calculation to ensure a correct
spectral characterisation. In this case, we use NDVI and MSAVI index, which includes
bare soil and dead vegetation effect, by adding a new variable named soil factor
adjustment based on the soil line approach.
Descriptive statistics were applied to characterise each one of the 46 dates using box-plots
charts. We distinguished several annual pasture phases based on the VIs trend changes in
the date-to-date series. Correlation Pearson’s coefficients (CR) were calculated to reveal
relationships between VIs and climatic variables series over time. Additionally, partial
correlation coefficients (PCR) were calculated to detect the most determinant variable in
the VIs response. Then, time-series lagged cross-correlation analysis was performed to
obtain an optimal lag (ℓ) where the correlation between the climate variable and VI is
maximum. Then, the climate-vegetation-soil system's complexity was characterised and
quantified by phase-state analysis: recurrence and cross-recurrence plots, diagonal-wise
recurrence quantification profile and recurrence quantification analysis.

52
4.2 Soil Line Acquisition
The soil line was calculated for each studied plot in each zone, and the results are
displayed in (Figure 12). RED-NIR method’s linear regression displayed an R2 > 0.90 in
all the cases. ZAV s values were higher than ZMA in most cases. We could not detect
significant differences between the s values from ZAV and ZMA, even if the s average
in ZAV (1.40) tends to be higher than in ZMA (1.17).

(a) (b) (c)

(d) (e) (f)

Figure 12. Soil slope (s) estimation from the spectral bands NIR vs. RED plots adapting the minimum NIR
method (Xu and Guo, 2013). Inside each figure, the regression equation was obtained from the Weighted
Least Squares method. Each black point represents the NIR vs. RED reflectance in an 8-day interval
between 2002 and 2018. Red points are the minimum value of NIR at 0.005 RED breaks. Panels (a–c)
correspond respectively to the soil slope of pixels 1, 2 and 3 for the Tornadizos de Ávila (ZAV) zone.
Panels (d–f) correspond respectively to the soil slope of pixels for the Soto del Real (ZMA) zone.

4.3 Inter-Annual Analysis and intra-annual analysis

In the time series, NDVI, MSAVI, temperature, and precipitation values were higher in
ZMA than ZAV (Figure 13). VIs and precipitation displayed a descending trend over the
years in both zones. In contrast, temperature showed an ascending trend in both zones.
All the estimated slopes were non-statistically significant (Table 9).

53
 (a) (b) (c)

(d) (e) (f)

Figure 13. Inter-annual variations of vegetation indices (Normalised Difference Vegetation Index—NDVI,
modified soil-adjusted vegetation index—MSAVI) and meteorological parameters of water (Annual
accumulated precipitation) and energy (Annual average Temperature) from 2002 to 2018 in Tornadizos de
Ávila (ZAV) (a–c) and Soto del Real (ZMA) (d–f).

Table 9. Statistical trends significance of vegetation indices ((Normalised Difference Vegetation Index—
NDVI, modified soil-adjusted vegetation index—MSAVI)) and climatic variables (TEMP: Temperature;
PCP: Precipitation) in ZAV: Tornadizos de Ávila and ZMA: Soto del Real.

Standard
t-estimated t-value
Zone Variable Slope Error R2 Signif*

NDVI -0.0014 0.0013 0.0678 1.0446 2.1314 n.s.

MSAVI -0.001 0.0011 0.0573 0.9552 2.1314 n.s.
ZAV
TEMP 0.0384 0.0262 0.1254 1.4665 2.1314 n.s.
PCP -2.613 4.2582 0.0245 0.6137 2.1314 n.s.
NDVI 0.0001 0.0019 0.0002 0.0529 2.1314 n.s.
MSAVI -0.0003 0.0015 0.0032 0.2186 2.1314 n.s.
ZMA
TEMP 0.0563 0.0277 0.2158 2.0318 2.1314 n.s.
PCP -5.7312 7.7954 0.0348 0.7352 2.1314 n.s.
Note: *Significance at 95% level of confidence: s, significant; n.s., no significant

Generally, an inverse relationship was observed between VIs and temperature

fluctuations (Figure 14). For instance, in 2009 and 2017, the temperature was remarkably
higher in both zones; consequently, NDVI and MSAVI showed a severe fall.
In contrast, we detected a direct relationship between VIs and precipitation. Usually,
when precipitation was limited, VIs tended to be reduced. This tendency was observed
over 2009 and 2017 in ZAV, and over 2003–2004 in ZMA, respectively.

54
ZAV

(a) (b)
ZMA

(c) (d)
Figure 14. Relationship between the standardised NDVI and MSAVI with min-max standardised climate
parameters of energy (Annual average Temperature) and water (Annual accumulated precipitation) in
Tornadizos de Ávila (ZAV) (a-b) and Soto del Real (ZMA) (c-d) at a yearly scale from 2002 to 2018.

In general, we detected that both VIs showed similar inter-annual variations in both zones.
However, there were unusual VIs changes in certain years, i.e., in 2007, VIs showed a
remarkable rise, even when temperature and precipitation did not show a notable change
of trend.

The VIs phases were closely related to the Mediterranean climate seasons, Chow’s test
revealed that all of them were significantly different from each other (Table 10), being
(P1 and P2) the cold season, (P4) the hot season and (P3 and P5) the transitional periods
(Table 11).

55
Table 10. Chow’s test of annual pasture phases. ZAV: Tornadizos de Ávila, ZMA: Soto Del Real. REG_1
(linear regression of the first phase data), REG_2 (linear regression of the second phase data, REG_T (both
data regression.) B0 is the intercept, and B1 is the slope of the linear regression.

P1-P2 P2-P3 P3-P4 P4-P5 P5-P1

B0 0,51 0,42** 1,09** 0,37** -0,86**
REG_1
B1 -0,01 0,01** -0,03** -0,01* 0,03**
B0 0,41** 1,09** 0,37** -0,83** 0,51**
REG_2
B1 0,01** -0,03** -0,01* 0,03** -0,01*
ZAV
B0 0,50** 0,60** 0,91** -0,16** 0,05 *
REG_T
B1 0,01** -0,01** -0,02** 0,02** 0,01**
F-chow 25,18* 197,02* 51,75* 56,88* 38,26*
T-Chow
F-test 3,02 3,02 3,02 3,03 3,03
B0 0,58** 0,52** 1,30** 0,30** -1,39**
REG_1
B1 0,01* 0,01** -0,04** -0,01* 0,05**
B0 0,52** 1,30** 0,30** -1,39** 0,58**
REG_2
B1 0,01** -0,04** -0,01* 0,05** 0,01*
ZMA
B0 0,58** 0,76** 0,95** -0,45** -0,09*
REG_T
B1 0,01** -0,01** -0,02** 0,02** 0,01**
F-chow 6,19* 174,12* 83,10* 60,80* 40,84**
T-Chow
F-test 3,02 3,02 3,02 3,03 3,03

Table 11. Annual pasture phases based on vegetation index (VIs) trend at the studied zones (Tornadizos de
Ávila and Soto del Real).

Code Months Seasons

Initial Date Final Date VIs Trend
Phase Implied Implied
November-December-
P1 25 November 25 January Autumn–Winter Constant
January
P2 2 February 23 April February-March-April Winter–Spring Increasing
P3 1 May 20 June May-June Spring–Summer Decreasing
July-August-
P4 28 July 22 September Summer–Autumn Constant
September
September-October-
P5 30 September 17 November Autumn Increasing
November

We found that VIs, temperature, and precipitation were higher in ZMA than ZAV on an
intra-annual scale (Figure 15). We also observed that VIs and precipitation were notably
different between zones in P2 and at the beginning of P3 and P5. It should be noted that
ZMA VIs declined faster in P3 and increased quicker in P5 than ZAV VIs. NDVI
dispersion was generally greater than MSAVI. At the same time, VIs and precipitation
showed a higher data dispersion in ZMA than ZAV. Both variables reached their
dispersion peak in the same period of the year (P3 and P5).

56
 (a) (b) (c)

(d) (e) (f)

Figure 15. Boxplots of vegetation indices and meteorological parameters of energy (Average temperature)
and water (Accumulated precipitation) in 8 days periods from 2002 to 2018 in Tornadizos de Ávila (ZAV)
(a–c) and Soto del Real (ZMA) (d–f).

Based on the box plots results, it was observed that P1 and P4 phases did not show any
variation along the time, being stable and less dispersed than the other phases. For this
purpose, linear regression analysis between VIs and climate variables was conducted only
in the phases in which a trend was observed in VIs (Figure 16 and Figure 17). In this way,
the most critical vegetation–climate driving factors were detected. Generally, the
temperature was identified as the potential driving factor in the vegetation–climate system
as it showed R2 > 0.9 in all the studied phases in both zones. We observed that the
temperature trend varied throughout the year, being positive in P2 and negative in P3.
Instead, precipitation showed lower R2 values than temperature, being the highest (>0.7)
in P3, and maintained the same positive trend in all the phases.

57
 (a)

(b)

Figure 16 Growing phases response of NDVI to water (Accumulated precipitation) and energy (Average
temperature) parameters at 8 days in Tornadizos de Ávila (ZAV) (a) and Soto del Real (ZMA) (b). In each
figure, the regression equation was obtained from the Least Squares method.

(a)

(b)

Figure 17. Growing phases response of MSAVI and water (Accumulated precipitation) and energy
(Average temperature) parameters, in 8 days periodsat Tornadizos de Ávila (ZAV) (a) and Soto del Real
(ZMA) (b). In each figure, the regression equation was obtained from the Least Squares method.

58
4.4 Time-Series Correlation and correlation by phases
Pearson’s coefficients analysis revealed that all correlation coefficients between VIs and
meteorological time-series were statistically significant along the time (Table 12 and
Table 13). The corresponding Pearson’s coefficients for temperature showed a negative
relationship, whereas precipitation displayed a positive relationship. The temperature was
the most correlated variable, and partial coefficients indicated that temperature is the main
driving factor in the relationships between climate variables and VIs.

Table 12 Time-series Pearson correlation coefficients (CR) and partial correlation coefficients (PCR)
between NDVI time-series and meteorological time-series for each study zone (ZAV: Tornadizos de Ávila;
ZMA: Soto del Real) during the period of 2002-2018. TEMP is 8-day average air temperature (°C), and
PCP is the accumulated precipitation in 8-day (mm).

Zone CR PCR
TEMP PCP TEMP PCP
ZAV -0.587** 0.192** -0.487** 0.168**
ZMA -0.743** 0.270** -0.523** 0.104**
Note: ** represents P < 0.01 significance.

Table 13. Time-series Pearson correlation coefficients (CR) and partial correlation coefficients (PCR)
between MSAVI time series and meteorological time-series for each study zone (ZAV: Tornadizos de
Ávila; ZMA: Soto del Real) during the period of 2002–2018. TEMP is 8-day average air temperature
(°C), and PCP is the accumulated precipitation in 8-day (mm).

Zone CR PCR
TEMP PCP TEMP PCP
ZAV −0.435 ** 0.165 ** −0.571 ** 0.199 **
ZMA −0.572 ** 0.199 ** −0.504 ** 0.074 **
Note: ** represents P < 0.01 significance.

Cross-correlation showed the sinusoidal behaviour of the climate variables and the
seasonality of VIs (Figure 18 and Figure 19). The Lag τ varied between variables and
zones, i.e., MSAVI-PCP showed a Lag of −2 (16 days) in ZAV, while Lag was of −3 (24
days) in ZMA. The Lag also fluctuated depending on the VI used, NDVI could not
differentiate precipitation lags between zones; meanwhile, MSAVI distinguished them.

59
 NDVI-TEMP NDVI-PCP

ZAV
ZMA

Figure 18 Time series cross-correlation between NDVI and climatic variables time series, Temperature
(TEMP) and accumulated precipitation (PCP) in Tornadizos de Ávila (ZAV) and Soto del Real (ZMA).
Each lag is of an 8-days period. The blue dot line is the confidence interval at 95%.

MSAVI-TEMP MSAVI-PCP
ZAV
ZMA

Figure 19. Time series cross-correlation between MSAVI and climatic variables time series, Temperature
(TEMP) and Accumulated precipitation (PCP) in Tornadizos de Ávila (ZAV) and Soto del Real (ZMA).
Each lag is of an 8-days period; the blue dot line is the confidence interval at 95%.

Pearson correlation coefficients by phases (Table 14 and Table 15) indicated that VIs
dynamics varied over a year. Correlation coefficients varied between zones, being the
highest difference in P3 and P5. The temperature was the most correlated climate variable,
achieving values higher than 0.7 in P3 and P5 in both zones and VIs. As expected, there
were some inaccuracies in the precipitation; thus, it was not possible to obtain high

60
correlations values (<0.5). However, in P3, MSAVI showed significant results (0.270 in
ZAV and 0.248 in ZMA), most likely due to the decrease of MSAVI data dispersion
during the dry season.
Table 14 Pearson correlation coefficients (CR) and partial correlation coefficients (PCR) between NDVI
and meteorological parameters in the five phases. TEMP is 8-day average air temperature (°C), and PCP is
the accumulated precipitation in 8-day (mm). ZAV is Tornadizos de Ávila, and ZMA is Soto del Real.

CR PCR
PHASE NDVI xTEMP NDVI x PCP NDVI x TEMP NDVI x PCP
P1 0.203* 0.029 0.238** -0.014
P2 0.524** 0.050 0.082 0.157*
ZAV P3 -0.740** 0.297** -0.066 -0.010
P4 0.081 0.109 -0.039 0.150
P5 -0.578** 0.265** 0.092 -0.010
P1 0.040 0.130 0.133 0.105
P2 0.353** 0.099 0.134 0.180*
ZMA P3 -0.786** 0.402** 0.044 -0.009
P4 -0.019 0.194* 0.019 0.194*
P5 -0.654** 0.224* 0.077 -0.126
Note: *represents P < 0.05 significance, **represents P < 0.01 significance.

Table 15. Pearson correlation coefficients (CR) and partial correlation coefficients (PCR) of MSAVI and
meteorological parameters in five distinct phases. TEMP is 8-day average air temperature (°C), and PCP is
the accumulated precipitation in 8-day (mm). ZAV is Tornadizos de Ávila, and ZMA is Soto del Real.

CR PCR
Phase MSAVI x TEMP MSAVI x PCP MSAVI x TEMP MSAVI x PCP
P1 0.130 0.046 0.173* 0.008
P2 0.489 ** 0.066 −0.027 0.180 *
ZAV P3 −0.714 ** 0.283 ** −0.095 −0.003
P4 0.270 ** −0.037 −0.130 0.106
P5 −0.550** 0.216* 0.072 −0.052
P1 0.101 0.023 0.144 0.015
P2 0.428 ** 0.042 0.026 0.125
ZMA P3 −0.763 ** 0.359 ** −0.007 −0.053
P4 0.248 ** 0.029 0.007 0.135
P5 −0.594 ** 0.160 0.040 −0.157
Note: * represents P < 0.05 significance, ** represents P < 0.01 significance.

A fluctuating lag was observed in cross-correlation coefficients by phases, pointing out

the different VIs dynamics during a year (Table 16 and Table 17). The most correlated
variable was the temperature. Precipitation did not achieve higher correlation values,
though a lag was needed in all the cases. As a result, the correlation by phase method
improved the correlations in all the precipitation cases. Overall, cross-correlation by
phases was significant in P3 and P5, allowing us to achieve better correlation coefficients
than time-series analysis, regardless of VI and zone.

61
Table 16. Cross-correlation coefficients between NDVI and meteorological parameters with different
lags (ℓ) in different phases. ZAV: Tornadizos de Ávila and ZMA: Soto del Real. TEMP is 8-day average
air temperature (°C), and PCP is the accumulated precipitation in 8-day (mm). Each time lag is of 8 days.
The bold letter represents the maximum correlation in each row.

Time Lag (ℓ)

Phase Param. 0 1 2 3 4 5 6
TEMP 0.526 0.483 0.456 0.314 0.239 0.161 0.026
P2
PCP 0.049 0.051 0.116 0.108 0.053 0.052 0.078
TEMP -0.740 -0.762 -0.773 -0.756 -0.730 -0.700 -0.659
ZAV P3
PCP 0.297 0.347 0.401 0.274 0.269 0.186 0.067
TEMP -0.578 -0.592 -0.607 -0.616 -0.632 -0.635 -0.567
P5
PCP 0.265 0.397 0.451 0.415 0.249 0.238 0.201
TEMP 0.345 0.217 0.192 0.174 0.133 0.064 -0.019
P2
PCP 0.096 0.171 0.204 0.198 0.165 0.182 0.148
TEMP -0.786 -0.784 -0.753 -0.729 -0.715 -0.709 -0.647
ZMA P3
PCP 0.402 0.423 0.433 0.369 0.375 0.329 0.225
TEMP -0.654 -0.650 -0.650 -0.620 -0.645 -0.618 -0.523
P5
PCP 0.224 0.336 0.437 0.375 0.304 0.270 0.249

Table 17. Cross-correlation coefficients between MSAVI and meteorological parameters with different
lags (ℓ) in different phases. ZAV: Tornadizos de Ávila and ZMA: Soto del Real. TEMP is 8-day average
air temperature (°C), and PCP is the accumulated precipitation in 8-day (mm). The bold number represents
the maximum correlation in each row. Each time lag is of 8 days.

Time Lag (ℓ)

Phase Param. 0 1 2 3 4 5 6
TEMP 0.492 0.451 0.418 0.333 0.271 0.198 0.041
P2
PCP 0.067 0.039 0.074 0.133 −0.031 −0.003 0.079
TEMP −0.714 −0.739 −0.741 −0.726 −0.714 −0.681 −0.629
ZAV P3
PCP 0.283 0.312 0.378 0.243 0.256 0.186 0.064
TEMP −0.550 −0.553 −0.575 −0.584 −0.605 −0.613 −0.561
P5
PCP 0.216 0.359 0.396 0.424 0.235 0.214 0.241
TEMP 0.422 0.308 0.259 0.233 0.169 0.109 0.029
P2
PCP 0.032 0.088 0.186 0.221 0.156 0.102 0.093
TEMP −0.763 −0.757 −0.717 −0.699 −0.695 −0.697 −0.637
ZMA P3
PCP 0.359 0.396 0.381 0.321 0.357 0.309 0.230
TEMP −0.594 −0.600 −0.625 −0.602 −0.603 −0.581 −0.512
P5
PCP 0.159 0.297 0.393 0.397 0.293 0.275 0.304

In NDVI and temperature, ℓ varied in both zones from zero in P2 to 5 (40 days) in P5. In
the MSAVI case, temperature ℓ showed a similar pattern to the NDVI case, varying only
in the P5 phase on ZMA. Concerning precipitation, in the NDVI case, an ℓ of 2 (16 days)
was found in all the phases studied for both zones. Though, MSAVI showed different
precipitation ℓ depending on the zone and the phases. Precipitation ℓ was 3 (24 days) in

62
the phases when the VI increases (P2 and P5) in both zones. In P3, with a decreasing
trend in the VI, the precipitation ℓ was 2 in Avila and 1 in Madrid. This fact pointed out
a higher precipitation MSAVI sensitivity and recognised ℓ as a variable able to distinguish
different zones in the same phase.
MSAVI performance was better than NDVI in winter and at the beginning of the spring
(P2). It was remarkable because precipitation dispersion reached its peak during this
phase. This fact might reveal that MSAVI might improve NDVI results in the springtime
at a correct time scale. It is recommendable to analyse both series from a dynamic point
of view, emphasising the behaviours of MSAVI in both zones. From this point, we
detected that both indices showed similar behaviour, though, generally, MSAVI showed
better results, suggesting that its dynamics would be more distinct than NDVI.

4.5 Recurrence plots and recurrence diagonal profile characterisation

The Optimizeparam function was computed to estimate the parameters of RPs and
showed that the embedding dimension (m) increased in all VI series for both zones (Table
18 and Annex 3). In the case of NDVI, m was 2 for both zones, then for the MSAVI case,
m was 6 for ZAV and 8 for ZMA. The τ ranged from 8 to 11 varying between zones. The
same dimensionality increase was detected in the climate variables where precipitation
showed a higher dimension than the temperature in ZMA, pointing out a higher
precipitation complexity than temperature.

Table 18. Recurrence plot (RP) parameters in ZAV: Tornadizos de Ávila and ZMA: Soto del Real. MSAVI:
Modified Soil-Adjusted Vegetation Index, m: Embedding dimension, τ: Delay, 𝜀: threshold.

Zone RPs m τ 𝜺 RR (%)

MSAVI 6 9 20.91 4.99
ZAV TEMP 2 11 8.76 4.99
PCP 2 3 1.35 4.97
MSAVI 8 11 26.32 4.99
ZMA TEMP 2 11 9.00 4.99
PCP 10 9 13.25 5.00

NDVI RPs showed a noisy behaviour, characterised by many isolated points. Meanwhile,
MSAVI RPs showed white stripes on a large scale Figure 20. Furthermore, we could
observe that MSAVI RPs present small-scale structures and periodic patterns (diagonal
line like-shapes). This kind of structure is visible in the temperature RP, where we could
observe the temperature seasonality through diagonal-like structures. At the visual
inspection, we did not detect significant temperature changes between the two zones.

63
 NDVI MSAVI TEMP PCP

ZAV

(a) (b) (c) (d)

ZMA

(e) (f) (g) (h)

Figure 20. Optimised recurrence plots (RP) using normalised vegetation indices (MSAVI), Average temperature (TEMP) and Accumulated precipitation (PCP) data and rescaled
distance matrix for Tornadizos de Ávila (ZAV) and Soto del Real (ZMA). Time units are represented as the X and Y-axis. Each time-unit is 8-days, coincident with 8-day
composed MODIS images during the study period (2002–2018). Panels (a–d) correspond respectively to RP of NDVI, MSAVI, Temperature and Accumulated Precipitation
for the ZAV zone. Panels (e–h) correspond respectively to NDVI, MSAVI, Temperature and Accumulated precipitation for the ZMA zone.

64
 In contrast, precipitation RP showed a distinct pattern in each zone. In ZAV, we could
observe a block-like structure, whereas, in ZMA, we could distinguish a more line-like
pattern. This difference was likely due to the different precipitation regimes in each zone.
We found that the profile tendency varied between VIs and zones (Figure 21). NDVI
showed a more distinctive RR drop on the first days (0–8 days) in both zones. In contrast,
MSAVI maintained higher values of RR until 16 days. As expected, temperature showed
similar behaviour in the two zones. Meanwhile, PCP showed a different maximum lag.
Both PCP profiles showed lower values of RR from eight days till the end. We speculate
that different maximum lags are a consequence of the different PCP distribution in both
zones.

NDVI MSAVI TEMP PCP

ZAV

(a) (b) (c) (d)

ZMA

(e) (f) (g) (h)

Figure 21. Diagonal-wise recurrence profile of the RPs obtained from the MSAVI, Average temperature
(TEMP) and accumulated precipitation (PCP) in Tornadizos de Ávila (ZAV) (a–c) and Soto del Real
(ZMA) (d–f).

65
4.6 Cross recurrence plots and recurrence diagonal profile characterisation
Once RPs were constructed, the maximum m and τ for the two time series were selected
as the parameters to the construction of CRPS (Table 19 and Annex 3), then a RR of 5%
was selected for all of them.
Table 19. Cross-Recurrence plot (RP) parameters in ZAV: Tornadizos de Ávila and ZMA: Soto del Real.
MSAVI-TEMP: Cross-Recurrence plot between MSAVI and Average temperature. MSAVI-PCP: Cross-
Recurrence plot between MSAVI and Accumulated precipitation.

Zone CRPs m τ 𝜺 RR (%)

MSAVI-TEMP 6 11 23.04 5.00
ZAV
MSAVI-PCP 6 11 17.27 5.00
MSAVI-TEMP 8 11 25.60 5.00
ZMA
MSAVI-PCP 10 11 23.74 5.00

We detected the seasonal temperature effect on the VIs dynamics in the CRPs (Figure 22
and Figure 23) In the case of NDVI-TEMP, we could not distinguish between zones.
However, MSAVI-TEMP showed a different pattern in each zone, probably due to the
distinct MSAVI dynamics that distinguish between different CRP patterns. Cross-
recurrence profile allowed us to distinguish the interactions between VIs and climate
variables in the LOC and the surroundings lags. The temperature did not show a
difference between zones. As we observed in TEMP-VIs CRPs, the LOC was non-
existent (Figure 22 and Figure 23), and the surrounding regions were similar. Then it was
expected that RR was near zero on the first lags. We believe that this fact was due to the
temperature seasonality not being detected on the first lags.

66
 NDVI-TEMP NDVI-TEMP NDVI-PCP NDVI-PCP
ZAV

(a) (b) (c) (d)

ZMA

(e) (f) (g) (h)

Figure 22. Optimized Cross-Recurrence Plots (CRPs) and diagonal-wise recurrence profiles between vegetation indices data (NDVI) temperature data (TEMP) and accumulated
precipitation data (PCP) for Tornadizos de Ávila (ZAV) and Soto del Real (ZMA). Time units are represented as the X and Y-axis. Each time-unit is 8-days, coincident with 8-
day composed MODIS images during the study period (2002-2018) in the CRPS. Lags are represented in days in the diagonal-wise recurrence profile The panels [a,c,e,g]
represent the CRPs of NDVI-TEMP and NDVI-PCP. The panels [b,d,f,h] represent the diagonal-wise profile of the CRPS, respectively.

67
 MSAVI-TEMP MSAVI-TEMP MSAVI-PCP MSAVI-PCP

ZAV

(a) (b) (c) (d)

ZMA

(e) (f) (g) (h)

Figure 23. Optimised Cross-Recurrence Plots (CRPs) and diagonal-wise recurrence profiles between vegetation index data (MSAVI) temperature data (TEMP) and accumulated
precipitation data (PCP) for Tornadizos de Ávila (ZAV) and Soto del Real (ZMA). Time units are represented as the X and Y-axis. Each time-unit is 8-days, coincident with 8-
day composed MODIS images during the study period (2002–2018) in the CRPS. Lags are represented in days in the diagonal-wise recurrence profile The panels (a,c,e,g)
represent the CRPs of MSAVI-TEMP and MSAVI-PCP. The panels (b,d,f,h) represent the diagonal-wise profile of the CRPS, respectively

68
In contrast to the temperature, CRPs of precipitation and VIs were able to characterise a
different dynamic in each zone. In this case, we observed that NDVI-PCP CRPs showed
vertical lines in ZAV and diagonal-like structures in ZMA. In MSAVI-PCP CRPs, the
most distinct region zone occurred during 400–600 time-units, where ZMA presented an
isolated point structure, while ZAV did not show any recurrence in that timeframe. This
fact might be explained due to the increase of dimensionality produced by the
precipitation in the CRPs. VIs-PCP recurrence profile showed an evident maximum lag
of 16 days in the ZMA. In contrast, the ZAV presented a lower RR in precipitation, and
the maximum lag was not evident, presenting a more stable recurrence profile. This
difference can be explained because the CRQA analysis detected a higher number of
precipitation events coupled to the MSAVI index, showing a more evident maximum lag.
in the ZMA zone.

4.7 Recurrence Quantification Analysis of RPs and CRPs

We will present a more quantitative analysis by carrying out the RQA of the RPs and
CRPs considered for each study zone (Table 20 and Annex 3).

Table 20. Recurrence plot (RP) and Cross Recurrence plots (CRP) parameters and Recurrence
Quantification Analysis (RQA) using general z-score vegetation indices series in ZAV: Tornadizos de
Ávila and ZMA: Soto del Real. MSAVI: Modified Soil-Adjusted Vegetation Index, m: Embedding
dimension, τ: Delay, 𝑟: threshold, RR: Recurrence rate, DET: Determinism, LT: Average length of diagonal
structures, ENTR: Shannon Entropy, LAM: Laminarity, TT: Trapping time.

Zone RPs and CRPs RR (%) DET (%) LT ENTR LAM (%) TT
MSAVI 4.99 73.20 3.40 1.57 82.66 3.55
TEMP 4.99 40.26 2.31 0.72 52.89 2.49
ZAV PCP 4.97 14.65 2.10 0.32 33.72 2.35
MSAVI-TEMP 5.00 63.75 2.88 1.29 75.30 3.62
MSAVI-PCP 5.00 19.25 2.17 0.47 23.08 2.27
MSAVI 4.99 78.42 4.17 1.80 86.13 3.88
TEMP 4.99 42.02 2.35 0.76 54.35 2.60
ZMA PCP 5.00 6.57 2.03 0.14 20.72 2.11
MSAVI-TEMP 5.00 69.06 3.21 1.50 76.59 3.75
MSAVI-PCP 5.00 25.70 2.26 0.64 29.23 2.29

Now, we present the values of DET, LT, ENTR, LAM and TT. The DET is related to the
system’s random or periodic behaviour based on the density of recurrence points, being
higher when the system shows more periodical behaviour. The MSAVI presented a higher
DET in both zones, being ZMA the highest. DET obtained in precipitation RP was higher
in ZAV than ZMA.
Concerning CRPs, MSAVI-TEMP showed a higher DET than NDVI-TEMP. Both of
them showed a higher DET than PCP CRPs. MSAVI-PCP showed a DET increase in

69
comparison to NDVI-PCP in ZMA. We believe that MSAVI could characterise better
precipitation data dispersion, allowing us to improve the NDVI results in ZMA.
The LT is interpreted as the system’s predictability time, increasing when the
predictability time is longer. The LT values in both VIs obtained were low. Temperature
showed a higher LT than precipitation, suggesting that temperature predictability time
was higher than precipitation. In CRPs, we observed similar results, being the LT of VIs-
TEMP higher than VIs-PCP.
The ENTR value refers to the disorder of the system. MSAVI RPs showed a higher value
than NDVI, higher in ZMA than ZAV. Concerning climate variables, temperature showed
a higher ENTR than precipitation in their separate RPs and their CRPs with VIs.
The LAM value refers to the chaos–chaos transitions and is directly related to the
detection of laminar states. The MSAVI LAM was higher than NDVI. Generally,
temperature showed higher LAM values than precipitation. We also noticed that when
precipitation was involved, LAM values tended to decrease in the CRPs, being
temperature the highest in both cases.
The TT represents the average length of vertical structures and indicates how long the
state will be trapped at the same time. The MSAVI showed a higher value of TT in both
zones. In the climate variables RPs, Temperature TT was higher than precipitation TT.
The same phenomenon happened in the CRPs, where TT was higher in VIs-TEMP than
VIs-PCP.

4.8 Discussion
Concerning s values, we obtained different values of s in each zone. According to Baret
et al. (1993)’s findings, s increased when soil moisture grew. As shown in Table 4, ZAV
topsoil was more clayey and less sandy than ZMA soil. Thus, a higher water holding
capacity (WHC) was expected in ZAV than ZMA, explaining the differences between
zones. We could not obtain significant differences between the s values from ZAV and
ZMA, most likely because of the litter and non-photosynthetic material influence
(Quemada et al., 2018).
Temperature and precipitation yearly tendencies found in this work are consistent with
what was reported in previous studies, where the temperature is increasing, and
precipitation is decreasing in semiarid zones due to the climate change effect (Chuai et
al., 2013; Hao et al., 2012). We detected specific years when VIs dramatically dropped
(2005, 2009, and 2017). These years coincide with drought periods (Estrela and Vargas,

70
2012; Vicente-Serrano et al., 2017) that happened in Spain (2004–2008 and 2016–2017).
These phenomena most likely negatively affected the vegetation growth; thus, VIs values
decreased during this time. We expected a non-significant result in the trend slope
because environmental works suggested a significant data quota is needed to obtain
trustworthy climate trends at a yearly scale (Pang et al., 2017).
We obtained different inter-annual trends depending on the climate variable. An inverse
relationship between temperature and VIs was found, in agreement with previous research
(Xu et al., 2016), pointing out that NDVI and other optical indices are generally inversely
related to temperature.
Both VIs used the same NIR-RED spectral bands, leading to a similar performance (Liu et
al., 2004). However, the soil factor’s addition in the MSAVI case was expected to increase
the sensitivity in semiarid areas (Huete, 1989; Qi et al., 1994; Ren et al., 2018). In our
results, MSAVI showed lower dispersion than NDVI, pointing out a better potential for
characterisation of semiarid pasture grasslands. In contrast, the precipitation was directly
related to VIs. This fact ties nicely with previous studies (Liu et al., 2015) wherein
precipitation events were related to vegetation growth, leading to increased VIs values.
Several studies (Chuai et al., 2013; Hao et al., 2012; Sun and Qin, 2016) emphasise that
precipitation and temperature combine in a dynamic and complex system; thus, their
networks must be considered. As was reported Suzuki et al. (2007), NDVI could be
affected by other complementary variables, such as evapotranspiration, that depends on
the combination of local wetness and warmth. The lack of these variables might explain
the unusual behaviours in the VIs time series that are not directly related to temperature
or precipitation.
We believe that ZMA VIs were higher than ZAV because of the higher amount of
precipitation in ZMA during P2 and at the beginning of the P5 phase. Precipitation events
increased soil moisture leading to an increment in vegetation growth that VIs detected.
We speculate that the interaction between soil moisture and soil texture with the
vegetation in P2 and P5 might explain the differences between zones (Viola et al., 2008).
At the end of the rainy season (P2), both soils’ water storage is likely to be highest after
the winter water recharge. However, water holding capacity was lower for the sandy than
for the clayey soil (Table 4), and so less water was available in the ZMA than in the ZVA
soil. During P3, the temperature raised, then the pasture in the ZMA depleted the soil water
more quickly than in the ZAV. Therefore, ZMA vegetation decreased faster than ZAV
(Figure 15). A similar effect took place at the beginning of the P5. Water storage in both

71
soils was expected to be the lowest due to the previous dry season (P4). During P5,
precipitation increased in both zones. ZMA sandy soil permeability was higher than ZAV
clayey soil allowing a faster increment in ZMA vegetation than ZAV.
We detected an increase in the VIs dispersion during P3 and P5. This variability increase
was expected due to the precipitation variability increase that occurred during spring and
autumn. As reported by Grant et al. (2014), precipitation variability leads to increased
soil moisture variability. Therefore, grassland productivity is altered due to water
availability fluctuations, leading to higher variability in VIs results.
Fu and Burgher (2015) pointed out the temperature as the most potential driving factor in
NDVI dynamics. They also revealed that temperature harms NDVI as we detected the
same result in P3. This effect is explained by the limitation in vegetation growth produced
by higher temperatures and fewer precipitations during the dry season. These climate
conditions enhance the intensity of transpiration and reduce the available soil water (Mao
et al., 2012); thus, vegetation growth is expected to be limited in these unfavourable
conditions.
Precipitation was the only variable that maintained the same positive trend in all the
phases, pointing out that precipitation is regularly favourable in semiarid grassland
growth. The same conclusion was achieved by Sala et al. (1988), which presented a
positive relationship between precipitation and pasture grasslands growth because of the
soil moisture’s positive role in biomass production.
Multiple works have demonstrated transparent relationships between climate and VIs
response (Gessner et al., 2013; Olmos-Trujillo et al., 2020). Our results agree with
(Braswell et al., 1997), who showed a negative relationship between temperature and
NDVI. In line with Liu et al. (2015), we also found a positive correlation between
precipitation and VIs in the semiarid area pasture grasslands.
Cross-correlation results allowed to expose the seasonal behaviour of the VIs over time.
Simultaneously, we also observed that there was a lagged response between VIs and
climate variables. In line with this idea, most studies indicate an (up to) 3 months lagged
relationship between VIs response and climate variable effect (Xu et al., 2016). The range
of the period appears to be related to the studied area’s specific characteristics, such as
climate, topography, and soil type, affecting VI lagged response (Cao et al., 2011; L. Guo
et al., 2014).
The VIs seasonal behaviour also plays a crucial role in the VIs dynamics (Gong et al.,
2015); therefore, we obtained different strengths in the relationship between VIs and

72
climate over the year. As Helman et al. (2015) reported, NDVI showed a better response
to grassland vegetation during wet seasons due to the herbaceous vegetation’s growth in
the Mediterranean climate. This effect might explain the higher correlation found in P3
and P5 phases, coincident with the Mediterranean weather’s wet seasons. The same idea
might be suggested to the differences between ZMA and ZAV, being ZMA wetter than
ZAV.
We achieved better precipitation correlations when the year was divided into phases,
although they were not as high as the temperature. This fact is consistent with Ramos
(2001) work suggesting that Mediterranean precipitation is characterised by a complex
seasonal variability pattern, with large and unpredictable rainfall fluctuations from one
year to the other, hindering the relationship between precipitation and VIs.
Once the year is divided into phases, our results highlight a variable lag’s usefulness
depending on the year’s season to characterise the vegetation–climate system. This result
agrees with Zhang et al. (2018) study that supports the idea of a variable ℓ along the time.
In their case, ℓ varied from 0 to 90 days depending on the season and the climate variable.
In line with this idea, Chuai et al. (2013) suggested that the season of the year and the
type of vegetation cover are critical in the ℓ estimation. Even more, other authors indicate
that local conditions are crucial in the estimation of ℓ. Suzuki et al. (2007) revealed that
the NDVI lagged response changed inside the same study area. In the northern, NDVI
varied because of the warmth variations. On the other hand, in the southern, NDVI varied
due to the inter-annual wetness fluctuations instead of warmth changes.
From another point of view, several authors reported that ℓ depends on the observed time
scale. Cui and Shi (2010) found a 30-day NDVI lagged response to precipitation.
Meanwhile, Wang et al. (2003) suggested that the bi-weekly lag was the most correlated.
As we stated before, it is essential to note that these relationships were found in these local
conditions and the proposed phases. If the analysis is applied to a broader area, the
relationships might not be statistically significant as the conditions could differ in space and
time (Meroni et al., 2014).
We already emphasised the incredible complexity of the vegetation–climate system in the
previous analysis. Ecological systems present nonlinear dynamics, combining chaotic and
periodic cycles, whose equations controlling the systems are unknown (Proulx and
Parrott, 2009; Storch and Gaston, 2004). Thus, the nonlinear analysis provides
complementary information about the system. In our work, we detected a dimensionality
increase in MSAVI RPs. More detailed soil information was introduced through L in the

73
MSAVI series; thus, a higher embedding dimension was expected. This result agrees with
previous literature findings (Belaire-Franch et al., 2002; Donner et al., 2010), which relate
dimensionality increases to complexity increases. In this line, Marwan et al. (2015)
demonstrated the usefulness of RPs to describe nonlinear behaviours in high-dimensional
systems, such as VI time-series.
In MSAVI RPs, we found white stripes in the RPs pattern. These structures are related to
atypical values and an interruption in the vegetation pattern (Proulx et al., 2015). We
believe that this behaviour is due to an extreme climatic event that increased soil moisture;
consequently, VI series values atypically increased, being detected in the MSAVI RPs.
Diagonal-wise profiles of the CRPs revealed the maximum lag in each zone, and it is
expected to vary between zones, seasons, and vegetation cover. Several lags have been
reported in the literature. Cao et al. (2011) reported a twenty-day lag in precipitation and
temperature in Xinjiang’s arid area (China). In contrast, Richard and Poccard (1998)
reported a maximum lag of three months in South Africa.
Other authors refer to seasons as the most crucial factor in the variation of maximum lag.
As reported by Zhao et al. (2020), precipitation showed a 1–2 month lag in spring,
whereas maximum lag is reduced to 1 month in the Autumn season. They also revealed
that temperature showed the same lag as precipitation in the spring; however, maximum
lag might be increased up to 3 months in autumn. This result is in concordance with the
recurrence profile results, where the temperature did not show an evident maximum lag
in the first 50 days.
The CRQA analysis allowed us to quantify the different dynamics of the VIs and climate
variables. The DET value has been utilised to indicate climate stability (S. C. Li et al.,
2008) or detecting bioclimatic transitions (Zhao et al., 2015). The MSAVI time-series
showed a higher value of DET in comparison to NDVI. Our results suggest that MSAVI
allowed us to characterise the semiarid grasslands better. We speculate that soil moisture
is being detected by the MSAVI index, allowing us to improve the NDVI results in ZMA.
The same phenomena happened in the VI-PCP CRPs, where MSAVI achieved a higher
DET value than NDVI in ZMA. Our results agree with Marwan et al. (2015) that found a
higher DET in a humid grassland area than a dry grassland.
Regarding LT, both VIs obtained low values compared to the periodic series (Zhao et al.,
2015). This fact might indicate that vegetation may be predicted in the short term due to
the incredible complexity of ecological systems, as reported by Beckage et al. (2011).

74
Now, let us discuss the values of ENTR, which refer to the disorder of the system.
Standard values obtained by Zhao et al. (2015) noted that stochastic systems tend to obtain
lower ENTR values (0.2) in comparison with those of periodic systems (2.20). We
speculate that the high value of ENTR in the MSAVI case (see Table 20) is the
consequence of the high number of precipitations in ZMA. Marwan et al. (2015) sustained
this fact, suggesting that wet grassland areas tend to obtain higher ENTR values than dry
grassland areas.
The LAM and TT are related to the vertical structures created in the RPs and CRPs. LAM
refers to the chaos transitions and represents the number of laminar states (Marwan et al.,
2002). MSAVI presents a higher value than NDVI, indicating that values are trapped
during specific periods, decreasing time-series dispersion, and supporting the idea of
higher predictability and determinism of the MSAVI index. TT represents the average
length of vertical structures and indicates how long the state will be trapped, while
MSAVI showed a higher TT value than NDVI. We believe that this fact is the
consequence of the behaviour of each time series. MSAVI is less dispersed, and then it is
expected to be trapped in similar states much longer than NDVI. The same principle might
be applied to temperature and precipitation. Temperature is seasonal and did not
dramatically change between two consecutive measures. In contrast, precipitation is
erratic, and it is not equally distributed over time (Ramos, 2001), especially in the
Mediterranean climate.
Overall, our results highlight the incredible complexity of the grassland system. We
observed that the time scale is a critical component in the analysis of the VIs series. At
the same time, we prove that RPs, CRPs and CRQA are a promising analysis that could
provide complementary information about the system dynamics that the linear methods
could not describe.

75
5 RECURRENCE PLOTS FOR
QUANTIFYING THE
VEGETATION INDICES
DYNAMICS IN A SEMIARID
GRASSLAND

76
5.1 Objective, area of study and materials
This chapter aimed to identify variations in the temporal dynamics of two agroecological
zones through recurrence techniques and VIs applied in the anomalies time series.
The two areas studied were the same as in section 3: ZGU in Guadalix de la Sierra and
ZSO in Soto del Real. MOD09Q1 product from MODIS imagery was selected for this
study. Study plot reflectance was monitored from 2002 to 2018. Each year, 46 MODIS
images free of cloud were acquired, giving 782 images in the study period. The spectral
vegetation indices considered in this section were NDVI and MSAVI. Both VIs were
calculated for each plot, and the plots' average was considered for the time series analysis.
VIs anomalies series (A_NDVI and A_MSAVI) were obtained by subtracting the average
per date of the year and obtaining a seasonally adjusted.
The two AEMET (Agencia Estatal de Meteorología, 2020) stations, described in section
3, were used to obtain daily meteorological data. The precipitation (PCP) represents the
accumulated precipitation every eight days, and the temperature (TEMP) represents the
average temperature every eight days. Temperature anomalies series (A_TMIN) were
obtained by subtracting the average per date of the year in the case of TMIN. In the PCP
case, we preferred to subtract the median per date of the year due to its tailed frequency
distribution in some dates (A_PCP).
This section studied original and seasonally adjusted series (anomalies series) with RP
analysis. In addition, CRP analysis was applied using VIs and climatic original and
anomalies series. First, the Optimizeparam function was computed to obtain the optimal
values of the three parameters (m, τ, and 𝜀). The threshold 𝜀 is estimated by an iterative
process based on the standard deviation of the time series and a limit in the RR value
(Thiel et al., 2002). Based on different studies of RR, a maximum value of 2.50 % was
set (Marwan, 2007) in this case. The series analysed were normalised using a z-score, and
the distance matrix (DM) was rescaled based on the maximum value. Then, the Crqa
function was computed, using the optimization values obtained earlier, to calculate the
complexity measures. Finally, the drpfromts function was computed to plot diagonal-wise
recurrence profiles in the RPs and CRPs.

5.2 Soil slope and soil adjustment factor

The values obtained in the soil slope (s) calculation presented R2 values higher than 0.90
(Figure 24). In general, the ZGU area presented higher s values than those in ZSO, the
average value of 0.98 and 0.91, respectively. The average soil factor used for MSAVI

77
(𝐿𝑀 ) showed a similar yearly trend for both areas, increasing during summer and
decreasing in wet seasons (Figure 25). Although, 𝐿𝑀 values showed visible differences
from the beginning of autumn till the middle of spring. During this period, ZGU showed
a higher 𝐿𝑀 than ZSO.

78
 Pixel 1 Pixel 2 Pixel 3 Pixel 4
ZGU

Y = 0.95*x + 0.09 Y = 0.97*x + 0.08 Y = 1*x + 0.07 Y = 0.98*x + 0.08

s = 0.95 s = 0.97 s = 1.00 s = 0.98

ZSO

Y = 0.93*x + 0.09 Y = 0.87*x + 0.08 Y = 0.91*x + 0.07 Y = 0.91*x + 0.12

s = 0.93 s = 0.87 s = 0.91 s = 0.91

Figure 24. Soil slope (s) estimation for each plot by the minimum NIR method (Xu & Guo, 2013). Inside each figure, the regression equation obtained from weighted least
squares method (Xu & Eckstein, 1995). Each point represents the RED and NIR reflectance in an 8-days interval between 2002 and 2018. ZGU = Guadalix de la Sierra, ZSO =
Soto del Real.

79
Figure 25. Soil factor adjustment for MSAVI (𝐿𝑀 ) average and standard error throughout the year at:
Guadalix de la Sierra (L_MSAVI_ZGU) and Soto del Real and Madrid (L_MSAVI_ZSO). X-axis represent
the month and the date of each measure.

5.3 Descriptive statistics

The yearly average of the two VIs from 2002 till 2018 was calculated to study the trend
in those years (Figure 26). In ZGU, NDVI presented a positive trend higher than in ZSO.
However, these tendencies were not statistically significant (Table 21). In the case of
MSAVI, the tendencies were almost irrelevant with the same statistical result. Different
trends could be identified concerning the temperature and precipitation, but the slopes
were not statistically significant either.

80
 0.55 15.0

0.50 14.5

14.0
0.45
13.5
0.40
13.0
0.35
NDVI

Temp (ºC)
ZGU ZSO 12.5
0.30 ZSO ZGU
12.0
0.25
11.5

0.20
11.0

0.15 10.5

0.10 10.0

0.55
ZGU ZSO

0.50 800

ZGU ZSO
0.45
700

Precipitation (mm)
0.40
MSAVI

0.35 600

0.30

500
0.25

0.20
400
0.15

0.10 300

Figure 26. Yearly average of the VIs (NDVI and MSAVI), temperature (TEMP) and annual accumulated
precipitation (Precipitation) from 2002 till 2008.
Table 21. Statistical trend significance between the annual variables studied and time in each area (ZGU
and ZSO): NDVI, MSAVI, Temperature (TEMP) and Precipitation (PCP).

Site Lin. Reg. NDVI MSAVI TEMP PCP

slope 0.0004 0.004 0.0563 -5.7312
s.e. 0.0013 0.0010 0.0277 7.7954
ZGU R2 0.0051 0.0112 0.2158 0.0348
t - est 0.2762 0.4120 2.0320 0.7352
significance ns ns ns ns
slope 0.0023 0.0009 -0.0104 -1.4648
s.e. 0.0012 0.0004 0.0252 7.9764
ZSO R2 0.3208 0.2168 0.0113 0.0022
t - est 2.6614 2.0376 0.4134 0.1836
significance ns ns ns ns
s.e.: standard error; s : significant ; ns: not significant; t value : 2.1314; t-est : estimated t value

Generally, NDVI values were higher and more dispersed than MSAVI values (Figure 27).
The evolution of VIs through the different seasons displayed a similar pattern for both
zones. The summer presented the lowest values, which began to increase in autumn and
reached a peak in spring. ZGU VIs values were lower than those of ZSO, except in the
summer, where ZGU values were equal or slightly higher than ZSO.

81
 A)

B)

Figure 27 Vegetation index statistics, from 2002 to 2018, along the year in each zone: ZGU = Guadalix de
la Sierra, ZSO = Soto del Real. Y-axis shows a box and whisker plot of A) NDVI (Normalized Difference
Vegetation Index) and B) MSAVI (Modified Soil-Adjusted Vegetation Index) in each date. Error bars
indicate maximum and minimum values; horizontal line indicate median value; and box indicates the 25%
and 75% percentiles. Red and blue solid lines represent the median of the time series

A similar seasonal pattern as the one shown by VIs appeared in the precipitation (Figure
28A). Precipitation distribution throughout the year was similar in both zones, showing a
lack of precipitation during the summer and an increase during autumn and winter.

82
Interestingly, middle autumn precipitation dispersion was the highest, similar to the
highest VIs data dispersion time frame.
The temperature difference in both sites, around 2ºC, was noticeable during the summer
and at the beginning of autumn, showing the highest data dispersion in spring (Figure
28B). Observing the influence of temperature in the VIs pattern (Annex 4), during
autumn, VIs in ZGU showed a lower increase than ZSO, with similar precipitation,
probably due to a higher temperature in ZSO. During winter, VIs showed a steady trend
in both zones. At the beginning of spring, VIs showed an increase in ZGU. Meanwhile,
ZSO kept a steady trend. This fact suggests that lower temperatures might restrict the
biomass growth in ZGU during the cold season. In the second part of the spring, VIs
decreased more slowly in ZGU than ZSO and achieved the same value at the end. Finally,
during summer, VIs values in ZGU were slightly higher than in ZSO.

83
 A)

B)
))

Figure 28. Accumulated 8-day precipitation (PCP) statistics along the year, including from 2002 to 2018,
in each zone: ZGU = Guadalix de la Sierra, ZSO = Soto Del Real. Y-axis shows a box and whisker plot of
A) PCP_ZGU, PCP_ZSO and B) TEMP_ZGU and TEMP_ZSO in each date. Error bars indicate maximum
and minimum values; horizontal line indicate median value; and box indicates the 25% and 75% percentiles.

5.4 Recurrence and Cross-recurrence analysis of original series

An exploratory analysis of the RP, with an embedding dimension m= 1 and delay = 0,
for each VI series was carried out (Annex 5). The line of identity (LOI), characteristic of
RPs, was well perceived in all of them and different isolated group of black points was

84
observed. These structures were similar to the white noise showed in Figure 4A, which
may indicate noisy behaviour. This exploratory analysis indicated that it was necessary
to optimize RP to search for a more evident pattern.
A visual comparison between the RPs presented in Figure 29 and Annex 5 confirms the
optimization. Optimized NDVI RPs presented a few diagonal structures, and isolated
point structures were reduced compared to the non-optimized RPs. The NDVI RPs were
noisy and appeared not to show the difference between the two zones. In contrast, MSAVI
RPs showed a more evident diagonal structure and displayed structure differences
between both zones.

85
 NDVI MSAVI TMIN PCP

ZGU

ZSO

Figure 29. Optimized recurrence plots using normalized VIs (NDVI and MSAVI) and climate (TMIN and PCP) data and rescaled distance matrix for both zones (ZGU and
ZSO). Time units are represented as the X and Y-axis. Each time-unit is a period of 8-days, coincident with 8-day composed MODIS images during the study period (2002-
2018).

86
The embedding dimension (m) in the VIs series for both zones presented different
behaviour (Table 22). In ZGU, NDVI presented an m =2, and MSAVI had an m =3. When
the analysis was done in ZSO, NDVI maintained the same embedding dimension, and
MSAVI increased to 10. We speculate that the information introduced through 𝐿𝑀 in the
MSAVI series increased the complexity pattern, and a higher embedding dimension was
needed in both zones. Concerning climate series, a substantial difference was found in 𝑚.
TMIN series presented an 𝑚 = 2. On the other hand, PCP showed an 𝑚 = 10 implying
a higher complexity in the dynamic as expected.
A different range of complexity measures (Table 22), depending on the VI and zone, were
observed. Both VIs showed a DET increase from ZGU to ZSO zone, being more
noticeable in the MSAVI. MSAVI also showed an evident increment in LT, ENTR and
LAM from ZGU to ZSO.
Looking at the RQA of the climate variables, the DET and LAM parameters pointed out
the difference in TMIN that was not visually observed in the RPs. PCP presented time-
delay differences between zones; ZGU also showed a higher DET and LAM than ZSO.
It is important to note that descriptive statistics in the chapter’s section 5.3 showed that
the PCP of both zones was very close and presented similar patterns. However, the
dynamical analysis performed with RPs showed visual and quantitative differences.

Table 22. Recurrence plot (RP) parameters and Recurrence Quantification Analysis (RQA) of each study
area using normalized VI (NDVI, MSAVI), minimum temperature (TMIN) and accumulated precipitation
(PCP). ZGU = Guadalix de la Sierra, ZSO = Soto Del Real. NDVI = Normalized Difference Vegetation
Index, MSAVI = Modified Soil-Adjusted Vegetation Index, m= Embedding dimension, τ = Delay, 𝜀 =
threshold, RR = Recurrence rate. DET = Determinism, LT = Average length of diagonal structures, ENTR
= Shannon Entropy, LAM = Laminarity

RR DET LT LAM
m τ 𝜺 ENTR
(%) (%) (LTmax) (%)
NDVI 2 6 4.81 1.87 57.05 2.96 (14) 1.31 61.81
MSAVI 3 9 7.38 1.86 67.27 2.98 (24) 1.31 73.37
ZGU
TMIN 2 6 4.01 1.52 12.76 2.09 (4) 0.31 18.03
PCP 10 14 7.80 1.17 3.97 2.00 (2) 0.00 18.81
NDVI 2 11 4.50 1.87 60.70 2.78 (20) 1.21 64.87
MSAVI 10 9 19.89 2.03 92.69 5.16 (638) 1.85 93.14
ZSO
TMIN 2 6 3.01 0.94 6.65 2.04 (3) 0.17 11.24
PCP 10 9 9.52 1.12 1.02 2.00 (4) 0.00 5.37

87
CRPs between VIs and TMIN are shown in Figure 30. The LOI does not appear as we
are now comparing two different series. MSAVI with TMIN highlighted the differences
between both zones. In ZGU, CRP presented fewer straight diagonals and did not show a
clear white stripe, as we could see in ZSO. CRPs of VIs with PCP showed a similar plot
between them (Figure 30), though each site showed a different pattern that varied in the
quantity and the localization of high-density recurrence stripes.

88
 NDVI x TMIN MSAVI x TMIN NDVI x PCP MSAVI x PCP

ZGU

ZSO

Figure 30. Optimized Cross-Recurrence Plots (CRPs) between VIs (NDVI and MSAVI), minimum temperature (TMIN) and precipitation (PCP) for both zones (ZGU and
ZSO). Time units are represented as the X and Y axis. Each time-unit is a period of 8-days, coincident with 8-day composed MODIS images during the study period (2002-
2018).

89
Overall, the quantification of CRPs (Table 23)followed the same patterns achieved in the
RPs (Table 22). VIs with TMIN showed a higher DET, LT, ENTR, and LAM than VIs
with PCP, emphasizing the higher complexity of precipitation. At the same time, VIs with
PCP showed higher DET, LT, ENTR, and LAM in ZGU than ZSO, suggesting that PCP
is more periodic in ZGU than in ZSO.

Table 23. Cross-Recurrence plot (CRP) parameters and Recurrence Quantification Analysis (RQA) of VIs
(NDVI and MSAVI), minimum temperature (TMIN) and precipitation (PCP) in each study area. ZGU =
Guadalix de la Sierra, ZSO = Soto Del Real. NDVI = Normalized Difference Vegetation Index, MSAVI =
Modified Soil-Adjusted Vegetation Index, m = Embedding dimension, τ = Delay, 𝜀 = threshold, RR=
Recurrence rate, DET = Determinism, LT = Average length of diagonal structures, ENTR = Shannon
Entropy, LAM = Laminarity.

RR DET LT ENT LAM

m τ 𝜺
(%) (%) (LTmax) R (%)
NDVI XT_MIN 2 8 5 1.43 22.26 2.13 (7) 0.40 25.63
MSAVI X T_MIN 3 9 8 1.33 30.85 2.24 (6) 0.61 37.56
ZGU

NDVI X PCP 10 14 16 1.30 13.52 2.13 (5) 0.37 14.27

MSAVI X PCP 10 14 15 1.63 16.46 2.12 (5) 0.35 18.32
NDVI X T_MIN 2 11 6.5 1.65 23.70 2.17 (6) 0.47 26.32
MSAVI X T_MIN 10 9 22 1.57 47.65 2.47 (9) 0.92 48.35
ZSO

NDVI X PCP 10 17 21.5 1.33 11.59 2.16 (5) 0.47 11.84

MSAVI X PCP 10 20 20 1.43 15.90 2.10 (6) 0.33 17.83

5.5 Recurrence and Cross-recurrence analysis of the anomaly series

This chapter’s section will focus on seasonally adjusted series, climatic (A_TMIN,
A_PCP) and VIs (A_NDVI, A_MSAVI), designated anomaly series.
The optimised RPs obtained for both VIS showed a vanished diagonals lines structure
and a similar pattern in both zones (Figure 31). For climatic anomalies, a noisy pattern is
observed in A_TMIN, almost identical in both sites. In A_PCP, both sites presented
different patterns and resembled the obtained in the original PCP series.

90
 A_NDVI A_MSAVI A_TMIN A_PCP

ZGU

ZSO

Figure 31. Optimized recurrence plots using VIs anomalies (A_NDVI, A_MSAVI) for both zones (ZGU and ZSO), minimum temperature anomalies (A_TMIN) and
precipitation anomalies (A_PCP). Time units are represented as the X and Y-axis. Each time-unit is a period of 8-days, coincident with 8-day composed MODIS images during
the study period (2002-2018).

91
The embedding dimension between both VIs was quite close (Table 24). However, in
ZSO, the complexity was still higher. The delay in all the cases was 19-time units, close
to 5 months. Respect to climate variables, the embedding dimension and time delay were
the same in A_TMIN. A_PCP showed the same m and different time delay, 18 (close to
5 months) in ZGU and 6 in ZSO.

Table 24. Recurrence plot (RP) parameters and Recurrence Quantification Analysis (RQA) using VIs
anomalies (A_NDVI, A_MSAVI), minimum temperature (A_TMIN) and precipitation anomalies (A_PCP)
in each study area. ZGU series = Guadalix de la Sierra, ZSO = Soto Del Real. NDVI = Normalized
Difference Vegetation Index, SAVI = Soil-Adjusted Vegetation Index, MSAVI = Modified Soil-Adjusted
Vegetation Index, m= Embedding dimension, τ = Delay, 𝜀 = threshold, RR = Recurrence rate., DET =
Determinism, LT = Average length of diagonal structures, ENTR = Shannon Entropy, LAM = Laminarity.

RR DET LT LAM
m τ 𝜺 ENTR
(%) (%) (LTmax) (%)
A_NDVI 5 19 14.19 0.93 61.25 3.26 (18) 1.34 63.73
A_MSAVI 6 19 18.82 1.69 67.18 3.29 (137) 1.13 75.78
ZGU
A_TMIN 3 6 8.52 1.74 3.15 2.02 (3) 0.12 7.40
A_PCP 9 18 8.47 1.70 4.02 2.01 (3) 0.08 13.83
A_NDVI 6 19 22.19 1.71 77.10 3.61 (255) 1.31 83.93
A_MSAVI 7 19 21.88 1.68 80.86 4.31 (614) 1.43 86.42
ZSO
A_TMIN 3 6 7.51 1.18 2.19 2.02 (3) 0.12 3.49
A_PCP 9 6 11.28 1.47 0.66 2.00 (2) 0.00 5.07

The RQA of the VIs anomalies series revealed similar tendencies compared to the original
VIs time series, showing that DET, LT, ENTR and LAM increased from ZGU to ZSO.
Concerning the RQA of the climate anomalies, it was revealed that DET and LT showed
noticeable low values compared to the original series. Furthermore, ENTR values were
almost 2, which implies that ENTR was the lowest, implying the chaotic behaviour of
climate anomalies.
On another note, CRPs between VIs anomalies and A_TMIN presented a different pattern
in both sites (Figure 32). VIs and A_TMIN were more clamped in ZGU, whereas ZSO
presented a more linear structure. In the case of VIs anomalies and A_PCP, the line-like
patterns were similar in both zones. However, high co-recurrence periods were different
in each zone.

92
 A_NDVI X x A_TMIN A_MSAVI_x_A_TMIN A_NDVI x A_PCP A_MSAVI x A_PCP

ZGU

ZSO

Figure 32. Optimized Cross-Recurrence plots (CRPs) between VIs anomalies (A_NDVI, A_MSAVI), minimum temperature anomalies (A_MIN) and precipitation anomalies
(A_PCP) for both zones (ZGU and ZSO). Time units are represented as the X and Y-axis. Each time-unit is a period of 8-days, coincident with 8-day composed MODIS images
during the study period (2002-2018).

93
The RQA of the CRPs between VIs and climate anomalies did not reveal remarkable
differences in DET, LT, ENTR and LAM between zones (Table 25). Although, CRPs
with A_TMIN showed lower DET than the CRPs with A_PCP, suggesting that VIs
anomalies are less synchronised with A_TMIN than with A_PCP.

Table 25. Cross-Recurrence plot (CRP) parameters and Recurrence Quantification Analysis (RQA) of VIs
anomalies (A_NDVI, A_MSAVI). minimum temperature anomalies (A_TMIN) and precipitation
anomalies (A_PCP) in each study area. ZGU = Guadalix de la Sierra, ZSO = Soto Del Real. NDVI =
Normalized Difference Vegetation Index, MSAVI = Modified Soil-Adjusted Vegetation Index, m=
Embedding dimension, τ = Delay, 𝜀 = threshold, RR = Recurrence rate, DET = Determinism, LT = Average
length of diagonal structures, ENTR = Shannon Entropy, LAM = Laminarity.

RR DET LT LAM
m τ 𝜺 ENTR
(%) (%) (LTmax) (%)
A_NDVI X
5 8 17 2.11 8.49 2.05 (4) 0.21 8.80
A_T_MIN
A_MSAVI X
6 4 17 1.65 8.67 2.03 (3) 0.14 8.95
A_T_MIN
ZGU
A_NDVI X
9 18 14 1.23 11.30 2.15 (4) 0.43 15.01
A_PCP
A_MSAVI X
9 18 14 1.86 13.76 2.09 (4) 0.30 16.48
A_PCP
A_NDVI X
6 17 21 1.66 7.22 2.07 (4) 0.25 7.37
A_T_MIN
A_MSAVI X
7 17 25 1.84 8.60 2.07 (4) 0.26 9.26
A_T_MIN
ZSO
A_NDVI X
9 9 17.5 1.02 8.15 2.06 (3) 0.24 11.20
A_PCP
A_MSAVI X
9 9 18 1.97 11.77 2.08 (4) 0.29 15.19
A_PCP

5.6 Discussion
A significant difference with the ANOVA test for the s values was found, suggesting a
significant difference in both sites' soil surface and non-photosynthetic reflectance
(Quemada et al., 2018). We speculate that the main reason behind these differences is the
higher soil moisture content in ZGU than in ZSO, as the former showed a higher
percentage of clay. This reason is in line with Baret et al. (1993) results that positively
correlated the clay content with the soil moisture. On another note, 𝐿𝑀 trends obtained in
the work are in agreement with the results obtained by Qi et al. (1994), wherein low
vegetation density conditions (summer season) a large 𝐿𝑀 would best describe the soil-
vegetation interactions.
As expected, yearly climate trends were not statistically significant, even though a light
trend was found. These results can be partly explained by Zhao et al. (2011) work. Future

94
climate trends present non-linear behaviour resulting in highly diverse and even
counterintuitive observations that might lead to inaccurate statistical results.
The VIs are highly related to agro-climatic conditions and have proven to be an effective
indicator of vegetation cover, biomass production, plant growth and phenology (Pan et
al., 2015; Pérez-Hoyos et al., 2010). VIs trends tend to be linked to the variability, amount
and timing of water availability and temperature (Grime et al., 2008; Swemmer et al.,
2007). A similar MODIS NDVI trend was reported by (Wang et al., 2013). However, we
found much lower values for VIs in the dry season (beginning of June until the end of
September), as our studied zones were sited in an arid-semiarid climate.
As reported by (Catorci et al., 2021; Chen and Weber, 2014) it is important to note that
temperature can limit biomass growth in the cold season. This limitation is in line with
our results that suggest that winter temperature restricted the biomass growth in ZGU,
producing a remarkable increase in ZGU VIs when the winter finished.
Precipitation distribution and dispersion were reported (Chandrasekar et al., 2010; Chen
and Weber, 2014; Wang et al., 2003) to be positively correlated to VIs. This fact could
be observed in the VIs differences between the wet and dry seasons. This was particularly
during autumn, which showed the highest values and data dispersion in both variables:
VIs and precipitation.
It is crucial to consider the interactions between temperature, precipitation, and associated
variables. As it is the case of soil moisture, that has been identified as a linking factor
between precipitation, soil and vegetation growth, being the controller of rainfall inputs
and outputs, thus determining vegetation dynamics (Laio et al., 2001; Rodriguez-Iturbe,
2000; Scanlon et al., 2005). We speculate that soil moisture is involved in two exceptional
situations. In spring when VIs in ZGU decreased slower than ZSO. Precipitation
distribution was similar in both zones, though, according to WHC results (Table 1), ZGU
soils could retain more water than ZSO soils. Therefore, higher soil moisture combined
with lower temperatures in ZGU may explain the difference with ZSO. This phenomenon
is repeated in summer when a combination of higher soil moisture and lower temperatures
in ZGU resulted in higher VIs values.
This visual inspection of optimised RPs is in line with (Marwan et al., 2015) work that
revealed similar patterns in a northeast grassland zone in Spain. As we could see in the
RPs, some white stripes structures are related to atypical values and an interruption in the
vegetation pattern (Proulx et al., 2015). We believe this behaviour is due to an extreme
climatic event that atypically varied the VI series values.

95
As we have pointed out before, isolated-points structures in the RPs are related to
stochastic behaviours. In contrast, diagonal structures reveal periodic processes (Marwan
et al., 2007). An overall comparison suggests that MSAVI allow more precise observation
of the periodic vegetation patterns, distinguishing between both zones.
Optimised RP of the TMIN in both sites are similar and show a very persistent and cycled
pattern. This pattern agrees with Zhao et al. (2011), suggesting that temperature shows
temporal seasonality. Concerning PCP, visually, both sites present different structures in
the RPs. ZGU showed an irregular line-like pattern, whereas ZSO presented a more
irregular and block-like pattern. Both zones do not show a clear diagonal structure,
suggesting that precipitation presents a stochastic and erratic behaviour, typical of
Mediterranean precipitation (Ramos, 2001). It is important to note that RPs of
precipitation is allowed to distinguish between zones. In contrast, the annual cycle
suggests that precipitation presents a similar behaviour in both zones.
The increase of dimensionality between NDVI and MSAVI is in concordance with Zurlini
et al. (2018), showing that the Enhanced Vegetation Index (EVI) series presented an
embedding dimension of 3, higher than the NDVI series. Our findings are supported by
the literature (Belaire-Franch et al., 2002; Donner et al., 2010), which relate
dimensionality increment to system complexity increase and strengthen the idea that
MSAVI presents a higher complexity than NDVI.
Complexity measures allow to discern and characterise the non-linear behaviour of VIs
time series (Andrés F. Almeida-Ñauñay et al., 2021; Semeraro et al., 2020). Generally,
DET is a measure to quantify the stochasticity of a system (Thiel et al., 2004). It has been
successfully applied as an indicator of climate stability (S. C. Li et al., 2008) or the
detection of bioclimatic transitions (Zhao et al., 2015). Our result suggests that DET
increase in MSAVI (Table 3) is interrelated with a better characterisation of grassland
vegetation patterns through the addition of the 𝐿𝑀 soil factor. This fact might indicate a
better characterization of semiarid grasslands with the use of MSAVI instead of NDVI.
In almost all the cases in our studied zones, the LT values were close, between 2 and 3.
Increases in LT are interpreted as a long time of predictability, as previously reported by
(Frilot II et al., 2015; Syta and Grzegorz, 2015) in several contexts. In our case, MSAVI
offered a longer predictability time than that of the NDVI, being the highest in ZSO
(5.16). Therefore, our LT values suggest that VIs may be predicted in the short term due
to the incredible complexity of ecological systems not achieving the values obtained with
periodic systems (Beckage et al., 2011).

96
ENTR standard values obtained by Zhao et al. (2015) noted that stochastic systems tend
to obtain lower values (0.20) in comparison with those of periodic systems (2.20). One of
the zones (ZSO) presented higher ENTR values for MSAVI, implying that Soto Del Real
grassland was more periodic than Guadalix de la Sierra grassland.
LAM refers to the chaos-chaos transitions and is directly related to the detection of
laminar states (Marwan et al., 2002). MSAVI, with higher LAM values, clearly presents
more random chaotic transitions than NDVI. The increase in laminar states in MSAVI
indicates that VI values are trapped during certain time frames, decreasing time series
variability, and reinforcing the idea of higher determinism and predictability of MSAVI,
particularly in ZSO to the ZGU area.
Climate RPs also showed remarkable results in the RQA analysis. TMIN revealed a more
periodic behaviour than PCP as expected. This observation aligns with Wen et al. (2017)
work that suggests that non-linearity and uncertainty of precipitation are more complex
and diverse than temperature. PCP also revealed a time delay difference. ZSO with sandy
soil revealed a time delay of 4; meanwhile in ZGU, with more clay in the soil, showed a
time delay increased to 14. This difference in delay agrees with Piedallu et al. (2019),
suggesting that VIs dynamics are not only driven by climatic variability. Different
environmental factors act complementarily; such as soil parameters related to water
stress.
NDVI and TMIN CRPs do not show a characteristic pattern, though their structures
resembled the synchronization of oscillatory signals (Webber and Zbilut, 2005). In
contrast, MSAVI and TMIN CRPs distinguish between zones, ZSO showed a vertical
disruptive moment pointing out a desynchronization between both time series (Zolotova
and Ponyavin, 2007). The CRQA DET values revealed that TMIN is more synchronised
with VIs than PCP. Although Zhao et al. (2011) work was using a different recurrence
technique (joint recurrence plots), our results achieved the same conclusions, suggesting
that temperature is more synchronised with VIs than precipitation. A higher DET and
LAM reflect this. TMIN and PCP were also more synchronised with VIs in ZGU than
ZSO, most likely due to temperature and soil moisture differences. As reported by.
Holzapfel et al. (2006), there is a clear relationship existing between vegetation and water
availability. Clay content is one of the keys in the soil water content and the subsequent
VIs response. We speculate that the clay content is one of the reasons behind the
synchronization of VIs and climate. Soil with a higher content in clay, in moderate slopes,
such as ZGU, tend to be more time stable in humidity than sandy soils (Jacobs et al.,

97
2004). Therefore, we expected a higher synchronization in ZGU clayey soils than ZSO
sandy soils.
Nanzad et al. (2019) suggested that NDVI anomalies are closely related to temperature
and precipitation anomalies. These results agree with studies that indicated that VIs
behaviour is controlled by climatic fluctuations and revealed a delayed response (Shen et
al., 2014). These lags vary depending on the climatic variable local conditions, season,
and vegetation phenological phase (Almeida-Ñauñay et al., 2021; Zhang et al., 2018).
Although we could not observe differences between VIs anomalies visually, RQA
revealed dissimilarities between them. Once more, MSAVI, DET and LT were higher
than NDVI, reinforcing the idea that the soil factor adds valuable information to the VI
time series. At the same time, a lower DET and LT in the climate anomalies series suggest
that the original series seasonality was successfully removed in the anomalies time series.
However, their values are still higher than those presented by white noise. Therefore, we
expected a complexity increase of the anomalies time series, showing a decrease in their
periodic behaviour (DET) and the predictability degree (LT).
We have confirmed our hypothesis as we could observe that anomalies CRPs showed the
lowest DET and LT of all the cases. At the same time, CRPs did not show remarkable
differences in DET, ENTR and LAM between zones. We suspect that a clustering index
must quantify the visually recognised differences (Boers et al., 2021). Remarkably, CRPs
with A_PCP showed a higher DET than with A_TMIN, suggesting a higher
synchronization with VIs anomalies, especially with MSAVI. This point out the relevance
of include a soil factor in the semiarid-arid zones. Gessner et al. (2013) suggested that
precipitation anomalies are mainly related to VIs, showing a lagged response later
(Fabricante et al., 2009) and being sensitive in semiarid conditions. We should also
consider soil moisture's "memory effect" to explain the DET differences and the
synchronization between anomalies time series. Soil moisture memory is significantly
explained by two variables (Orth and Seneviratne, 2012): the accumulated precipitation
and initial soil moisture. As we have seen before, ZGU presents a higher potential water
content (Table 1) and is likely to have a higher memory than ZSO. Therefore, ZGU clayey
soil response is expected to be slower, longer, and more time stable, explaining the higher
DET achieved in ZGU than in ZSO. In summary, ZGU clayey soil displayed a higher
synchronization than ZSO in our case.

98
Overall, climate, soil and time VIs lagged response create the non-linearity and
complexity of the grassland vegetation system, and recurrence techniques have been
demonstrated to be a valuable tool to characterise them

99
 6 OPTIMIZATION OF SOIL
BACKGROUND REMOVAL TO
IMPROVE THE PREDICTION OF
WHEAT TRAITS WITH UAV
IMAGERY

100
6.1 Objectives, study area and materials
The aims of this chapter were: i) to assess if it is viable to implement a methodology to
reduce the high spatial intra-variability in a crop by removing the soil background
influence with a defined threshold, and ii) to determine the optimal threshold and evaluate
the predictive performance of the methodology.
The study was conducted at La Chimenea farm station, near Aranjuez (Madrid, Spain),
throughout the wheat growing season of 2019. Four previous experiments involving 133
plots randomly assigned to different N and water doses conducted in a field with a circular
pivot (220 m in radius were used in the current optimization work. Each experiment
referred to in this study as “sector” was designed with a varying dose of fertilization and
irrigation. All plots were harvested on the same date (15 July 2019), and the yield, protein
content and nitrogen output were measured following the same procedure. Multispectral
imagery obtained from UAV was acquired on three dates, corresponding to the different
growth stages of wheat: the beginning of stem elongation (GS32, 11 March 2019), final
stem elongation (GS39, 12 April 2019), and flowering (GS65, 13 May 2019). The result
of each shot was a high-resolution multispectral image composed of six reflectance
channels: (490 nm), (550 nm), red (671 nm), red edge (700 nm), far-red (760 nm), and
NIR (800 nm). Four vegetation indices were selected: NDVI, MSAVI, normalized
different red-edge index (NDRE) that incorporates into the equation the red-edge band,
susceptible to chlorophyll pigment concentration and the blue-red index (BRI) that is used
when only RGB cameras were available.
As a preliminary analysis, we estimated the histograms of the VI distribution in GS32 to
detect anomalies in the distribution of VIs and to identify the potential VIs that can be
optimised using the segmentation method. Then, TVO was applied to optimize the
removal of the soil pixels based on the VI distribution. We established percentile
probabilities from 0% to 100% with an interval of 5%, as smaller increments did not
decrease the RMSE and increased the computational time. We then estimated the
percentile value (xi) of the VI distribution in each plot of the experiment. Then, the
minimum xi of all plots was selected as the threshold value Tv. We computed an iterative
process using all the percentile probabilities to obtain a list of threshold values
(sequentially 𝑇𝑣1 , 𝑇𝑣2 … 𝑇𝑣𝑛 ). Each 𝑇𝑣 is used to threshold the plots, and then a linear
regression was performed between the median values and the wheat traits measurements.
This is calculated for each Tv to obtain the optimal threshold (Topt), where the coefficient
of determination (R2) is maximised and the RMSE is minimised.
101
6.2 Raw data and histogram analysis
We observed great intra-variability in the results of the different experiments (Figure 33).
There was not a well-defined relationship between yield and protein content. On the one
hand, we found that S2, composed of control plots without fertilization, achieved the
lowest yield values (2000–3500 kg ha-1) and protein content (100–120 g kg-1). On the
other hand, S1, S2, and S4 achieved a high yield variability (1100–5000 kg ha-1)
depending on the N treatments. The same was observed in the protein content that ranged
from 100 to 180 g kg-1.

Figure 33. Trial raw data exploration. The x axis represents the wheat yield (kg ha -1) and the y axis
represents the wheat protein content (g kg-1). Each color represents a sector from Figure 1.

The VIs distribution showed tails at the lowest values, probably due to the presence of
soil in the image (Figure 34). NDVI and NDRE were the VIs with the most extended tails.
On the contrary, MSAVI showed a shorter tail, probably because the MSAVI considered
a correction factor to mitigate the soil background reflectance effect.

102
Figure 34. Histograms of vegetation indices. The y axis represents the percentage of pixels in GS32 (11
March). The letters represent each of the vegetation indices used in the study: (A) NDVI, (B) MSAVI, (C)
NDRE, and (D) BRI1.

6.3 TVO performance

When increasing the number of segmentations, the linear regression between grain yield
and the VIs showed a clear trend of increasing R2 and decreasing RMSE (Figure 35).
These figures suggest that NDVI was an appropriate yield estimator, achieving an RMSE
as low as 578 kg ha-1 after segmentation (Figure 35A). MSAVI, which includes a soil
adjustment factor, achieved higher determination coefficients and lower RMSE than
NDVI and was less sensitive to the segmentation (less difference between R2 and RMSE
before and after the segmentation) (Figure 35B). NDRE was the best yield indicator as it
obtained the highest R2 and the lowest RMSE, probably due to the addition of the red-
edge channel. The segmentation improved the NDRE prediction, reducing the RMSE
from 459 to 427 kg ha-1 (Figure 35C). When comparing all VIs, the R2 for BRI1 was the
lowest and RMSE was the highest; both statistics were improved by the segmentation
(Figure 35D).

103
Figure 35. Coefficient of determination (R2) and root mean square error (RMSE) obtained from linear
regression between vegetation indices and protein content in each segmentation at GS65: A (NDVI), B
(MSAVI), C (NDRE), D (BRI1). Box labels are the optimum VI threshold value.

With respect to Topt, the values for yield estimation were between 0.15 and 0.34,
depending on the VI studied (Figure 35). Application of TVO to the acquired images
allowed identification of pixels with relevant soil exposure.
Overall, plots that received higher N-fertilizer rates had more crop coverage and less soil
was exposed (Figure 36).

A B

Figure 36. Application of TVO to the image. A) RGB image acquired at GS39. B) Binarised NDVI image
(Threshold value: 0.23). Black dots are soil. No nitrogen fertilizer was applied to N0, 50 kg N/ha to N1,
and 90 kg N/ha to N2 treatments

104
By contrast, there was no evidence that the TVO method improved the estimation of
protein content (Figure 37). In this case, the plots represent the performance of VIs at
flowering (GS65). The NDVI, MSAVI, and NDRE charts showed a flat trend in both R2
and RMSE. The R2 (≈0.60) and RMSE (≈16 g kg-1) values attained for NDVI, MSAVI,
and NDRE were similar (Figure 37A, B, and C). The MSE attained was 11.8%, with the
average protein content of all plots being 135 g kg-1 (13.5% protein content). The R2 and
RMSE achieved by using BRI1 as an indicator of protein content was inferior to the rest
of the VIs, and even worsen with the segmentation. The Topt for the protein content varied
between 0.1 and 0.21, but it was less relevant because the TVO performance was poor.

Figure 37. Coefficient of determination (R2) and root mean square error (RMSE) obtained from linear
regression between vegetation indices and protein content in each segmentation at GS65: A (NDVI), B
(MSAVI), C (NDRE), D (BRI1). Box labels are the optimum VI threshold value.

The N output is the joint response of yield and protein content; for this reason, the values
tend to be a combination of these elements. The TVO method improved the R2 and RMSE
measurements in the cases of NDVI, MSAVI, and NDRE (Figure 38A, B, C). In the case
of BRI1 (Figure 38D), TVO did not achieve a considerable performance. With respect to
Topt, it ranged from 0.15 to 0.31 depending on the VI employed.

105
Figure 38. Coefficient of determination (R2) and root mean square error (RMSE) obtained from linear regression
between vegetation indices and nitrogen output in each segmentation at GS39: A (NDVI), B (MSAVI), C (NDRE), D
(BRI1). Box labels are the optimum VI threshold value.

6.4 Pre- and post-segmentation analysis

There was a clear relationship between VIs and yield at GS32 and GS39 (Table 26). This
relationship tended to be stronger when the crop had reached its maximum concentration
of photosynthetic pigments and chlorophyll content (GS39) in the case of NDVI and
BRI1. However, MSAVI and NDRE, achieved their best performance in GS32.
Regarding protein content, the increase in R2 was gradual, and the highest value was
achieved at GS65. These results may be due to the date of the flight and crop growth.
With respect to N output, we observed the highest R2 values in GS39, because the
estimation based on the product of the variables involved in the calculation, whether yield
or protein content, surpassed the estimation based on the individual variables.

106
Table 26. Coefficient of determination (R2) between vegetation indices (NDVI, MSAVI, NDRE, and
BRI1), protein content, and nitrogen output before (pre-segmentation) and after (post-segmentation) the
segmentation methodology. GS32, GS39, and GS65 represent the winter wheat growth stages.

Vegetation
Yield Protein content Nitrogen output
index
GS32 GS39 GS65 GS32 GS39 GS65 GS32 GS39 GS65
Pre-thresholding
NDVI 0.50 0.58 0.22 0.03 0.22 0.59 0.42 0.75 0.61
MSAVI 0.56 0.52 0.17 0.05 0.24 0.62 0.50 0.71 0.56
NDRE 0.68 0.51 0.21 0.05 0.28 0.61 0.60 0.75 0.62
BRI1 0.23 0.38 0.11 0.00 0.18 0.45 0.17 0.51 0.41
Post-thresholding
NDVI 0.54 0.58 0.22 0.03 0.23 0.60 0.46 0.76 0.63
MSAVI 0.58 0.52 0.17 0.05 0.29 0.63 0.51 0.74 0.57
NDRE 0.72 0.51 0.21 0.06 0.29 0.61 0.64 0.76 0.63
BRI1 0.26 0.43 0.11 0.04 0.20 0.45 0.17 0.58 0.41

The post-segmentation results (Table 26) showed modest improvements between the pre-
and post-TVO methodology. In most yield cases, R2 was improved, especially in GS32
when the highest value was achieved in NDRE (0.72). We did not find an improvement
in the case of protein content, except for GS39. We observed a moderate increase in R2
value for N output (0.74-0.76), particularly in GS39. The VIs showed a similar
optimization. Among them, BRI1 was the most improved index, suggesting that ratio
indices could be enhanced by the TVO method. However, their R2 values were lower than
those of the rest of the VIs.
We observed an improvement in RMSE through TVO, as shown in Table 27. This
improvement was remarkable in the yield and N output assessment, especially in GS32.
On the contrary, the RMSE of the protein content was only slightly improved by the TVO
method.
Table 27. Root mean square error (RMSE) between vegetation indices (NDVI, MSAVI, NDRE, and BRI1),
protein content, and nitrogen output before (pre-thresholding) and after (post-thresholding) the thresholding
value optimization methodology. GS32, GS39, and GS65 represent the winter wheat growth stages.

Vegetation
Yield (kg ha-1) Protein content (g/kg) Nitrogen output (kg ha-1)
index
GS32 GS39 GS65 GS32 GS39 GS65 GS32 GS39 GS65
Pre-thresholding
NDVI 578.00 527.19 721.40 25.8 23.2 16.7 17.21 11.30 14.00
MSAVI 538.52 564.31 740.16 25.6 22.8 16.2 15.98 12.06 14.89
NDRE 459.39 570.20 724.32 25.5 22.2 16.4 14.23 11.24 13.90
BRI1 714.81 641.29 767.93 26.1 23.8 19.4 20.56 15.70 17.30
Post-thresholding
NDVI 551.65 526.44 720.86 25.7 23 16.6 16.53 10.95 13.74
MSAVI 528.69 564.11 740.16 25.5 22.1 16 15.72 11.51 14.80
NDRE 427.99 570.06 722.88 25.4 22 16.3 13.49 11.12 13.71
BRI1 700.02 613.67 767.93 25.7 23.4 19.4 20.56 14.52 17.30

107
6.5 Sector analysis
We observed an improvement in the linear regressions obtained by the TVO method
(Figure 39 and Figure 40), their equations, and the relative importance of each sector to
the total of the point cloud. We detected an effect of the optimization process, in that it
helped to reduce the dispersion of the point cloud (Figure 39A and Figure 40A; Figure 39
and Figure 40E).

Figure 39 NDVI with yield (A, B, C), MSAVI with protein content (D, E, F), and NDRE with nitrogen output yield
(G, H, I) linear regressions. Each column represents a different GS (first column, GS32; second column, GS39; third
column, GS65). Each color represents the different sectors in the trial.

108
Figure 40. NDVI with yield (A, B, C), MSAVI with protein content (D, E, F), and NDRE with nitrogen output yield
(G, H, I) optimised linear regressions. Each column represents a different GS (first column, GS32; second column,
GS39; third column, GS65). Each color represents the different sectors in the trial.

The highest values for yield were achieved in sector S4. At the same time, there was
saturation of the NDVI in the yield estimation, especially at GS39 (Figure 39B). The
results revealed that the prediction of protein content improved over time. At GS32
(Figure 39D) and GS39 (Figure 39E), the MSAVI values were highly dispersed and
corresponded to multiple protein content, especially in sector S3. At GS65, there was a
well-defined relationship between protein content and MSAVI (Figure 39F). Regarding
the N output, the maximum R2 value was achieved at the peak of the vegetation cycle
(GS39) (Figure 39H). Overall, the TVO methodology might be considered a pre-
treatment data curation tool for estimating yield and N output in earlier growth stages
(GS32).

6.6 Discussion
Previous research conducted in the same experimental field showed that VIs were a
feasible tool to monitoring N content. Raya-Sereno et al. (2022) observed that NDVI,
NDRE and BRI achieved R2 value ranging from 0.76 to 0.8 when assessing grain yield

109
using VIs and nitrogen nutrition index in the experiment from S1. In comparison, the
values reported in this work are lower. However, it is important to note that the work
presented in the current study includes the experiments from the other sectors, involving
water application as an additional factor and the water × N interaction. Furthermore, the
sensors used in Raya-Sereno et al. (2022) were highly sophisticated, involving hyper-
spectral airborne image acquisition with a very high-spatial resolution. High spectral and
spatial resolution equipment is expensive and rarely available beyond experimental fields,
therefore techniques that enhance reliability of accessible UAV imagery is needed.
A large body of scientific literature discusses the effect of soil reflectance background
when monitoring areas with partial vegetation coverage (Dyson et al., 2019; Li et al.,
2021; Qiao et al., 2022). In this study we showed that TVO may contribute toward
reducing the soil effect on remote-sensing indices and improve crop monitoring,
particularly in early growth stages. In our preliminary analysis, we observed that specific
pixels are grouped as a tail in the lowest values of the VI distribution. The results agree
with Qiao et al. (2020), even if our VI values did not display a distinct bi-modal
distribution (soil and vegetation) , the pixels with relevant soil exposure had VI values on
the lowest side of the distribution.
From a mathematical and probabilistic point of view, Topt has been defined by optimizing
functions and statistical models such as those by Duan et al. (2017) and Zhang et al.
(2019) to remove the effect of the soil background. In agreement with the results of Fu et
al. (2020), by thresholding the VI distribution, the R2 and the RMSE improved in the
present study. Our results also support the work of Cheng et al. (2008), which suggests
that an NDVI value lower than 0.22 is probably soil or non-photosynthetic material in
high-resolution imagery. In line with this, Campos et al. (2019) proposed a value of 0.3
at the beginning of biomass accumulation and 0.4 at the end of biomass accumulation or
grain maturity. To our knowledge, there is no literature on threshold optimization in
MSAVI, NDRE, or BRI1. It is important to note that MSAVI includes a soil adjustment
factor in the index, which is constant, variable, and even negative (Huete, 1988; Qi et al.,
1994; Ren et al., 2018; Xu and Guo, 2013). All these methodologies are mainly focused
on the NIR–RED relationship and the soil line concept that defines the limits between the
soil and the vegetation in the image.
The TVO method improved yield prediction during the early stages of crop growth
(GS32) in our study. Similar conclusions were drawn by Prey et al. (2018), who showed
that the best time frame to minimize the influence of the soil background is when the

110
wheat cover is still not fully developed. In our case, MSAVI and NDRE yielded better
results than NDVI. The results agree with the findings reported by Liaqat et al. (2017)
and Xie et al. (2016), showing a higher performance of MSAVI compared to NDVI in
semi-arid and low-leaf-surface areas. This fact might be due to the inclusion of a
correction factor in the MSAVI formula. In our case, NDRE is a better yield estimator
than NDVI. The inherent concept of NDRE explains this fact, since it is an index
specialised in the estimation of green biomass and chlorophyll content (Argento et al.,
2021; Magney et al., 2017; Yang et al., 2020). As multiple authors have reported, the
inclusion of the red-edge channel improves the VI sensitivity to green pigments and their
photosynthetic activity (Clevers et al., 2002; Dong et al., 2020; Li et al., 2021). We
believe that the outperformance of MSAVI in GS32 is related to a greater soil proportion
in comparison to the rest of growth stages (GS39 and GS65). NDRE, is an indicator of
green pigments content and it is more related to protein content than biomass. Therefore,
it was expected that at lower biomass quantities, green pigments were more concentrated
than in other stages; as a result, the performance of NDRE increased in GS32. Finally,
although the optical index BRI1 could be a valuable alternative when the sensor is the
main limitation (Li et al., 2014), it did not perform well in our study, most likely due to
the lack of NIR information in the VI calculation.
Regarding protein content, the TVO method improved the R2 and RMSE values at GS32
and GS39. However, R2 values were low enough to be considered non-relevant results in
the analysis. At GS65, TVO did not considerably improve estimation, with MSAVI and
NDRE being the most reliable indices. On the other hand, it was expected that protein
was better predicted in the last growth stages (grain filling) due to the movement of
nutrients from the plant to the grain content (Liu et al., 2006; Zhao et al., 2019). Therefore,
R2 values were better, ranging from 0.59 to 0.61 in GS65, but there is still room for
improvement to reach a reliable assessment.
Finally, we observed that the predictive performance of VIs (NDVI and BRI1) improves
during the peak of vegetation (GS39). These results are in line with those obtained by Liu
et al. (2022) and Salazar et al. (2007), who reported that the optimum period for
estimating wheat yield is during the booting–heading growth stages. On the other hand,
the predictability of MSAVI and NDRE decreased in GS39. We speculate that the R2 of
MSAVI decreased probably due to the absence of soil in the scene. Therefore, it was not
necessary to add a soil factor. In the case of NDRE, we believe that its close relationship

111
with the content of photosynthetic pigments is better employed in the prediction of protein
content instead of yield.
Additionally, this study showed a saturation of NDVI at GS39. The results reinforce the
idea that NDVI tends to saturate when the wheat cover reaches its green peak and the leaf
area index is the highest of the wheat cycle (Duan et al., 2017; Vannoppen and Gobin,
2021).
Although TVO seems to be a promising procedure, this work presents some limitations.
One of them is the spatial resolution to be employed. As multiple authors mentioned
(Hengl, 2006; López-Granados, 2011; Torres-Sánchez et al., 2015) it is required high-
resolution imagery (centimetre or millimetre scale) to successfully apply thresholding and
soil removing techniques in early stages of vegetation. Imagery with an inadequate size
of pixel may contain a mix of soil and vegetation, being difficult to distinguish between
both. Another issue is the presence of objects that are neither vegetation nor soil in the
UAV imagery (H. Zhou et al., 2021), such as rocks or crop residues (Hively et al., 2018).
In our case, it is an experimental field, assuring that there is no presence of foreign objects
inside the plots. However, this should be considered by researchers using segmentation
and thresholding methods in areas presenting a variety of land covers.
In summary, in this work, we demonstrated the potential of VIs from UAVs to predict
winter wheat traits. Furthermore, we successfully performed an iterative segmentation
method to select an optimal threshold value to minimize the soil background effect.
Further research should focus on evaluating and replicating TVO methodology in larger
datasets of high-input crops, such as maize and rice. Moreover, other potential
chlorophyll indices such as modified chlorophyll absorption ratio index (MCARI),
transformed chlorophyll absorption ratio index/optimized soil-adjusted vegetation index
(TCARI/OSAVI), Double peak canopy nitrogen index (DCNI), and Canopy chlorophyll
content index (CCCI) should be studied in future research.

112
7 CONCLUSIONS

113
7.1 Specific and general conclusions
This thesis shows that remote sensing techniques and imagery have an exciting potential
to assess the time and spatial variability of the climate-vegetation-soil system. It is
demonstrated that VIs are a feasible tool for monitoring vegetation and climate.
Furthermore, VIs time series can be used as a data input of advanced analysis data
techniques such as recurrence plots tools, revealing temporal dynamics of complex
processes such as the soil moisture role in the vegetation growing.
The spatial variability of high-resolution imagery remains to be a significant challenge in
the analysis of vegetation indices. Moreover, high-inputs crops require proper and precise
monitoring to adjust irrigation and fertilization to achieve commercial grain quality. In
this line, this thesis suggests a novel method to further improve wheat trait assessment by
reducing the spatial variability of soil pixels.
Finally, it is recommended to return to the central questions of the thesis to check whether
the specific objectives and the main objective were achieved:

a) Is it feasible to ascertain the complexity of vegetation time series related to climate

fluctuations and drought episodes through VIs (VHI and SVHI) and precipitation
indices (SPI and SPEI)? Is it possible to determine which index performed better and
the adequate time scale?

Choosing a suitable drought index to monitor grasslands ecosystems in arid and semiarid
conditions is difficult. In chapter 3, two standardised indices (SPI and SPEI) were selected
to examine their performance in semiarid grasslands. The statistical analysis shows that
a previous vegetation indices standardisation improves the correlation with climate
variables. Therefore, it is recommended to include a pre-treatment data to obtain the most
accurate relationships between drought indices (SPI and SPEI) and vegetation indices
(VHI and SVHI). It is also notable that optimisation of the α coefficient further improves
the correlation values.
We have obtained satisfactory results, showing that SPI and SPEI are suitable vegetation
and water status indicators. Furthermore, these results emphasise the importance of the
time scale as a determinant factor in monitoring semiarid grasslands. In this line, the
quarterly scale is the most appropriate for estimating these standardised indices in
semiarid grasslands. All in all, our approach can potentially improve the monitoring and
management of semiarid grasslands.

114
b) Are the VIs and recurrence techniques adequate tools to picture the complexity of the
relationships between vegetation and climate in grasslands? If it is, which is the most
suitable VI to be used in this case?

The relationships between VIs and the climate variables studied depend on the years’ time
and the climate variable. In chapter 4, the vegetation–climate complexity correlation
analysis suggests that each study area presents its features and local climate condition,
reinforcing that correlation analysis between VIs and climate variables should not be
restricted to year seasons.
The Cross-correlation method suggested that temperature was the most decisive driver
factor in this case. However, it was essential to note that precipitation showed a stable
positive trend along the phases suggesting that precipitation events were beneficial in
arid-semiarid grassland growth, regardless of the years’ time. Even though it was
challenging to study due to its irregular temporal distribution, precipitation showed a
significant positive correlation in phase 3 at the end of the spring, increasing when
MSAVI is used instead of NDVI.
The RP, CRPs and RQA were applied to VIs time series to measure the complexity of the
dynamics in the grassland–climate system. Both vegetation indices allowed us to
distinguish a different dynamic in each zone. However, we detected a characteristic
dynamic that points out the short-term predictability and high-dimensionality of the
MSAVI time-series. Moreover, this analysis allowed us to differentiate the precipitation
regime that affected MSAVI dynamics. This fact is shown in the CRPs, pointing out a
better grassland characterisation in the wet zone by MSAVI, likely due to the influence
of soil water availability.
Overall, our results suggest that bare soil influence and non-photosynthetic vegetation
residuals are included through the soil factor in the MSAVI formula, increasing MSAVI
sensitivity in semiarid pastures.

c) Is it possible to identify variations in the temporal dynamics of two different agro-

ecological zones through recurrence techniques and VIs applied in the anomalies time
series?

It is not simple to decide on a suitable vegetation index to monitor semiarid grasslands in

semi-arid conditions because of the significant influence of soil and non-photosynthetic
vegetation residuals. Because of the inherent non-linearity of these vegetation indices

115
series, RPs and CRPs offer new insight into the natural pasture system dynamics. These
advanced data analysis techniques allow us to visualize and quantify the system dynamics
subjected to several drivers, such as soil moisture and local climate events.
In this case, the differences in VIs behaviour between the two zones were explained by
the interactions of temperature and precipitation regimens with soil characteristics.
Furthermore, an important influence of the soil adjustment factor (LM) is detected on the
VIs data dispersion observed through boxplot analysis.
With respect to RP, the original and anomaly MSAVI series showed distinct temporal
patterns in each location. RPs showed different PCP dynamical patterns that distinguished
between zones, complementing classical statistical methods that revealed a similar PCP
distribution.
The CRPs of VIs with TMIN revealed a higher synchronization response than with PCP
and differentiated both zones when MSAVI was used instead of NDVI. The CRPs of VIs
with PCP anomalies showed a higher synchronization than TMIN anomalies, allowing
them to distinguish between both zones.
These results generally reinforce the idea of the high-dimensional dynamics of grassland
systems and the potential of RPs, CRPs, and RQA to characterise and quantify them.

d) Is it viable to implement a methodology to reduce the high spatial intra-variability in

a crop by removing the soil background influence with a defined threshold? If it is, is
it possible to determine the optimal threshold and evaluate the predictive performance
of the methodology?

Soil effect is not limited to time scale playing a key role in the spatial variability at high-
resolution vegetation monitoring. Till now, it is one of the most difficult challenges faced
by remote sensing techniques applied to agriculture. In this line, UAVs' potential was
demonstrated further to improve the monitoring and prediction of wheat traits. It was
detected that MSAVI and NDRE were the most sensitive indices, achieving a yield
prediction (R2 = 0.68) at the earliest stage (GS32) and a satisfactory prediction of protein
content (R2 = 0.6) at the later stage (GS65). This innovative method allowed the
establishment of an optimal threshold in each studied VI. The threshold value fluctuated
between 0.2 and 0.3 in predicting the yield at GS32 and ranged from 0.1 to 0.2 in protein
content predictions at GS65. Overall, TVO improved the prediction performance of VIs,
mainly in estimating yield and N output during GS32, by reducing the soil pixels’ effect
in high-resolution imagery.

116
 Is it possible to implement non-linear and optimization tools to unravel the temporal
variability and reduce spatial variability of the vegetation-soil-climate system to
improve the monitoring of agricultural systems?

This thesis emphasizes the need to implement and develop new methodologies to improve
our understanding of agro-ecological systems. In this line, vegetation indices are a
feasible tool to further expand our knowledge about grassland ecosystems and their
intricate temporal variability. Additionally, recurrence plot techniques demonstrate a
potential tool to handle complex and multidimensional systems involving variables that
are difficult to model.
During the thesis, it was possible to implement a novel methodology to reduce the spatial
variability in high-resolution imagery. Threshold value optimization reduced the intra-
variability of vegetation indices by removing soil pixels with an optimal threshold value.
As a result, the vegetation indices' predictive performance was moderately improved.
In summary, in conjunction with non-linear methodologies and optimization methods,
remote sensing tools allowed us to detangle the extraordinary complexity of agro-
ecological systems and further improve vegetation indices' performance.

7.2 Limitations and further research

7.2.1 Limitations
Although remote sensing techniques offer great advantages, this thesis has its limits.
Initially, this work is restricted by the local conditions of each zone, including different
optimal time scales that depend on the vegetation response to local meteorological
phenomena. In line with this, the precipitation study represents a significant challenge to
all researchers. The main problem is the erratic precipitation behaviour, which is not
equally distributed over time, especially in the Mediterranean climate. Furthermore,
precipitation events are closely related to other variables, such as soil moisture or
evapotranspiration, showing lagged responses and creating the non-linearity and
complexity of the climate-vegetation-soil system. For example, the "memory effect" of
the soil is explained by the accumulated precipitation and the initial soil moisture. In this
line, one of the main restrictions of the thesis was the availability of updated soil data.
Soil sampling and analysis are often arduous and time-consuming. Thus, finding a reliable
and consistent soil database at smaller temporal and spatial scales is challenging.
The temporal resolution of the imagery is a generalised issue in remote sensing. The
longer the time series is, the more possible it is to find complications such as clouds'

117
presence and atmospherical noise. Quality check procedures and atmospherical
corrections help to tackle the issue. In the thesis, obtaining the same temporal scale in the
meteorological (diary) and imagery data (one shot per 8 days) was impossible. However,
cloud noise was reduced because the MODIS product selected the best shot in the eight
days.
Another issue is the spatial resolution of the imagery. In the TVO methodology, high-
resolution imagery (centimetre or millimetre scale) is required to successfully apply
thresholding and soil-removing techniques. Furthermore, the presence of objects that are
neither vegetation nor soil in the UAV imagery may introduce noise in the imagery
analysis.

7.2.2 Further research

Future research should consider additional variables to characterise the climate-
vegetation-soil system better. In this line, one of the most promising is the water content
in the soil. Although, as is said before, a reliable soil database's availability seems
challenging.
Additionally, future studies should consider using vegetation indices under different
climates and vegetation covers. One of the most exciting lines of research is the use of
combined vegetation indices, showing the vegetation response through two indicators at
the same time.
Regarding recurrence plots, more tools are not implemented in this work. For instance, a
Joint recurrence plot seems to be a powerful and feasible tool for analysing simultaneous
processes.
In the same way, the TVO methodology applied in wheat cover should be tested in larger
datasets of high-input crops, such as maize and rice. Moreover, potential chlorophyll
indices such as modified chlorophyll absorption ratio index (MCARI), transformed
chlorophyll absorption ratio index/optimised soil-adjusted vegetation index
(TCARI/OSAVI), Double peak canopy nitrogen index (DCNI), and Canopy chlorophyll
content index (CCCI) should be studied to better monitoring the N translocation.

118
8 REFERENCES

119
Agencia Estatal de Meteorología, 2020. AEMET OpenData [WWW Document]. URL
https://opendata.aemet.es/centrodedescargas/productosAEMET (accessed
12.23.20).

Agencia Estatal de Meteorología, 2019. AEMET OpenData [WWW Document]. URL

https://opendata.aemet.es/centrodedescargas/inicio (accessed 5.24.19).

Ali, I., Cawkwell, F., Dwyer, E., Barrett, B., Green, S., 2016. Satellite remote sensing of
grasslands: from observation to management. J. Plant Ecol. 9, 649–671.
https://doi.org/10.1093/jpe/rtw005

Ali, Z., Hussain, I., Faisal, M., Nazir, H.M., Moemen, M.A. el, Hussain, T., Shamsuddin,
S., 2017. A Novel Multi-Scalar Drought Index for Monitoring Drought: the
Standardized Precipitation Temperature Index. Water Resour. Manag. 31, 4957–
4969. https://doi.org/10.1007/s11269-017-1788-1

Almeida-Ñauñay, Andres F., Villeta, M., Quemada, M., Tarquis, A.M., 2022. Assessment
of Drought Indexes on Different Time Scales: A Case in Semiarid Mediterranean
Grasslands. Remote Sens. 14, 565. https://doi.org/10.3390/rs14030565

Almeida-Ñauñay, Andrés F., Benito, R.M., Quemada, M., Losada, J.C., Tarquis, A.M.,
2022b. Recurrence plots for quantifying the vegetation indices dynamics in a semi-
arid grassland. Geoderma 406, 115488.
https://doi.org/10.1016/j.geoderma.2021.115488

Almeida-Ñauñay, A.F., Benito, R.M., Quemada, M., Losada, J.C., Tarquis, A.M., 2021.
The Vegetation–Climate System Complexity through Recurrence Analysis. Entropy
23, 559. https://doi.org/10.3390/e23050559

An, Q., He, H., Nie, Q., Cui, Y., Gao, J., Wei, C., Xie, X., You, J., 2020. Spatial and
Temporal Variations of Drought in Inner Mongolia, China. Water 12, 1715.
https://doi.org/10.3390/w12061715

Argento, F., Anken, T., Abt, F., Vogelsanger, E., Walter, A., Liebisch, F., 2021. Site-
specific nitrogen management in winter wheat supported by low-altitude remote
sensing and soil data. Precis. Agric. 22, 364–386. https://doi.org/10.1007/s11119-
020-09733-3

Asadi Zarch, M.A., Sivakumar, B., Sharma, A., 2015. Droughts in a warming climate: A
global assessment of Standardized precipitation index (SPI) and Reconnaissance
drought index (RDI). J. Hydrol. 526, 183–195.
https://doi.org/10.1016/j.jhydrol.2014.09.071

Baret, F., Jacquemoud, S., Hanocq, J.F., 1993. About the soil line concept in remote
sensing. Adv. Sp. Res. 13, 281–284. https://doi.org/10.1016/0273-1177(93)90560-
X

Barnes, E.M., Clarke, T.R., Richards, S.E., Colaizzi, P.D., Haberland, J., Kostrzewski,
M., Waller, P., Choi C., R.E., Thompson, T., Lascano, R.J., Li, H., Moran, M.S.,

120
 2000. Coincident detection of crop water stress, nitrogen status and canopy density
using ground based multispectral data. Proc. 5th Int. Conf. Precis Agric.

Beckage, B., Gross, L.J., Kauffman, S., 2011. The limits to prediction in ecological
systems. Ecosphere 2, 1–12. https://doi.org/10.1890/ES11-00211.1

Belaire-Franch, J., Contreras, D., Tordera-Lledó, L., 2002. Assessing nonlinear structures
in real exchange rates using recurrence plot strategies. Phys. D Nonlinear Phenom.
171, 249–264. https://doi.org/10.1016/S0167-2789(02)00625-5

Bellido Mulas, F., Casquet, C., González Lodeiro, F., Martin Parra, L.M., Martinez-
Salanova, J., Navidad, M., Del Olmo Sanz, A., 2004. Mapa geológico de la Hoja no
509 (Torrelaguna). Mapa Geológico de España E. 1:50.000. Segunda Serie
(MAGNA).

Benedetti, R., Rossini, P., 1993. On the use of NDVI profiles as a tool for agricultural
statistics: The case study of wheat yield estimate and forecast in Emilia Romagna.
Remote Sens. Environ. 45, 311–326. https://doi.org/10.1016/0034-4257(93)90113-
C

Bento, V., Trigo, I., Gouveia, C., DaCamara, C., 2018a. Contribution of Land Surface
Temperature (TCI) to Vegetation Health Index: A Comparative Study Using Clear
Sky and All-Weather Climate Data Records. Remote Sens. 10, 1324.
https://doi.org/10.3390/rs10091324

Bento, V.A., Gouveia, C.M., DaCamara, C.C., Trigo, I.F., 2018b. A climatological
assessment of drought impact on vegetation health index. Agric. For. Meteorol. 259,
286–295. https://doi.org/10.1016/j.agrformet.2018.05.014

Blanco, L.J., Paruelo, J.M., Oesterheld, M., Biurrun, F.N., 2016. Spatial and temporal
patterns of herbaceous primary production in semi-arid shrublands: a remote sensing
approach. J. Veg. Sci. 27, 716–727. https://doi.org/10.1111/jvs.12398

Boers, N., Kurths, J., Marwan, N., 2021. Complex systems approaches for Earth system
data analysis. J. Phys. Complex. 11–14. https://doi.org/10.1088/2632-072X/abd8db

Bonan, G., 2016. Ecological Climatology. Cambridge University Press, Cambridge.

https://doi.org/10.1017/CBO9781107339200
Boschetti, M., Bocchi, S., Brivio, P.A., 2007. Assessment of pasture production in the
Italian Alps using spectrometric and remote sensing information. Agric. Ecosyst.
Environ. 118, 267–272. https://doi.org/10.1016/j.agee.2006.05.024

Braswell, B.H., Schimel, D.S., Linder, E., Moore, B., 1997. The Response of Global
Terrestrial Ecosystems to Interannual Temperature Variability. Science (80-. ). 278,
870 LP – 873. https://doi.org/10.1126/science.278.5339.870

Bugalho, M.N., Abreu, J.M.F., 2008. The multifunctional role of grasslands, in: C., P.,
M.M., T. de S. (Eds.), Sustainable Mediterranean Grasslands and Their Multi-

121
 Functions, Options Méditerranéennes : Série A. Séminaires Méditerranéens.
Zaragoza : CIHEAM / FAO / ENMP / SPPF, pp. 25–30.

Campos, I., González-Gómez, L., Villodre, J., Calera, M., Campoy, J., Jiménez, N., Plaza,
C., Sánchez-Prieto, S., Calera, A., 2019. Mapping within-field variability in wheat
yield and biomass using remote sensing vegetation indices. Precis. Agric. 20, 214–
236. https://doi.org/10.1007/s11119-018-9596-z

Cao, L., 1997. Practical method for determining the minimum embedding dimension of
a scalar time series. Phys. D Nonlinear Phenom. 110, 43–50.
https://doi.org/10.1016/S0167-2789(97)00118-8

Cao, X.M., Chen, X., Bao, A.M., Wang, Q., 2011. Response of vegetation to temperature
and precipitation in Xinjiang during the period of 1998-2009. J. Arid Land 3, 94–
103. https://doi.org/10.3724/SP.J.1227.2011.00094

Carolina Sparavigna, A., Marazzato, R., 2015. Recurrence Plots of Geolocated Time
Series from Satellite Maps of NOAA STAR Vegetation Health Index. Int. J. Sci. 1,
47–54. https://doi.org/10.18483/ijSci.877

Carral, M.P., Martin-Serrano, A., Goy, J.L., Zazo, C., 1996. The high interfluve surfaces
of the rivers Manzanares-Jarama, NE of Madrid, Spain. Geomorphology and soil
characteristics. Estud. Geológicos 52, 231–241.
https://doi.org/10.3989/egeol.96525-6268

Catorci, A., Lulli, R., Malatesta, L., Tavoloni, M., Tardella, F.M., 2021. How the
interplay between management and interannual climatic variability influences the
NDVI variation in a sub-Mediterranean pastoral system: Insight into sustainable
grassland use under climate change. Agric. Ecosyst. Environ. 314, 107372.
https://doi.org/10.1016/j.agee.2021.107372

Chamaille-Jammes, S., Fritz, H., Murindagomo, F., 2006. Spatial patterns of the NDVI-
rainfall relationship at the seasonal and interannual time scales in an African
savanna. Int. J. Remote Sens. 27, 5185–5200.
https://doi.org/10.1080/01431160600702392

Chandrasekar, K., Sesha Sai, M.V.R., Roy, P.S., Dwevedi, R.S., 2010. Land Surface
Water Index (LSWI) response to rainfall and NDVI using the MODIS Vegetation
Index product. Int. J. Remote Sens. 31, 3987–4005.
https://doi.org/10.1080/01431160802575653
Chen, F., Weber, K.T., 2014. Assessing the impact of seasonal precipitation and
temperature on vegetation in a grass-dominated rangeland. Rangel. J. 36, 185.
https://doi.org/10.1071/RJ13098

Chen, Z., Wang, W., Fu, J., 2020. Vegetation response to precipitation anomalies under
different climatic and biogeographical conditions in China. Sci. Rep. 10, 1–16.
https://doi.org/10.1038/s41598-020-57910-1

122
Cheng, W.-C., Chang, J.-C., Chang, C.-P., Su, Y., Tu, T.-M., 2008. A Fixed-Threshold
Approach to Generate High-Resolution Vegetation Maps for IKONOS Imagery.
Sensors 8, 4308–4317. https://doi.org/10.3390/s8074308

Chow, G.C., 1960. Tests of Equality Between Sets of Coefficients in Two Linear
Regressions. Econometrica 28, 591–605. https://doi.org/10.2307/1910133

Chuai, X.W., Huang, X.J., Wang, W.J., Bao, G., 2013. NDVI, temperature and
precipitation changes and their relationships with different vegetation types during
1998-2007 in Inner Mongolia, China. Int. J. Climatol. 33, 1696–1706.
https://doi.org/10.1002/joc.3543

Chuvieco, E., 2002. Principios físicos de la teledetacción, in: Teledeteccion Ambiental.

La Observacion de La Tierra Desde El Espacio. Ariel Ciencia, Barcelona, p. 586.

Clevers, J.G.P.W., De Jong, S.M., Epema, G.F., Van Der Meer, F.D., Bakker, W.H.,
Skidmore, A.K., Scholte, K.H., 2002. Derivation of the red edge index using the
MERIS standard band setting. Int. J. Remote Sens. 23, 3169–3184.
https://doi.org/10.1080/01431160110104647

Coco, M.I., Dale, R., 2014. Cross-recurrence quantification analysis of categorical and
continuous time series: an R package. Front. Psychol. 5, 1–14.
https://doi.org/10.3389/fpsyg.2014.00510

Colaço, A.F., Bramley, R.G.V., 2018. Do crop sensors promote improved nitrogen
management in grain crops? F. Crop. Res. 218, 126–140.
https://doi.org/10.1016/j.fcr.2018.01.007

Cong, N., Shen, M., Yang, W., Yang, Z., Zhang, G., Piao, S., 2017. Varying responses of
vegetation activity to climate changes on the Tibetan Plateau grassland. Int. J.
Biometeorol. 61, 1433–1444. https://doi.org/10.1007/s00484-017-1321-5

Cui, L., Shi, J., 2010. Temporal and spatial response of vegetation NDVI to temperature
and precipitation in eastern China. J. Geogr. Sci. 20, 163–176.
https://doi.org/10.1007/s11442-010-0163-4

Danson, F.M., Plummer, S.E., 1995. Advances in environmental remote sensing,

Advances in environmental remote sensing. John Wiley, Chister.
https://doi.org/10.1016/0277-3791(96)89785-x

Das, A., Agrawal, R., Mohan, S., 2015. Topographic correction of ALOS-PALSAR
images using InSAR-derived DEM. Geocarto Int. 30, 145–153.
https://doi.org/10.1080/10106049.2014.883436

De Castro, A.I., Shi, Y., Maja, J.M., Peña, J.M., 2021. UAVs for Vegetation Monitoring:
Overview and Recent Scientific Contributions. Remote Sens. 13, 2139.
https://doi.org/10.3390/rs13112139

123
Del Cerro, J., Cruz Ulloa, C., Barrientos, A., de León Rivas, J., 2021. Unmanned Aerial
Vehicles in Agriculture: A Survey. Agronomy 11, 203.
https://doi.org/10.3390/agronomy11020203

Dias, D., Pinto, A., Dias, U., Lamparelli, R., Le Maire, G., Torres, R.D.S., 2020. A
Multirepresentational Fusion of Time Series for Pixelwise Classification. IEEE J.
Sel. Top. Appl. Earth Obs. Remote Sens. 13, 4399–4409.
https://doi.org/10.1109/JSTARS.2020.3012117

Dong, X., Zhang, Z., Yu, R., Tian, Q., Zhu, X., 2020. Extraction of information about
individual trees from high-spatial-resolution uav-acquired images of an orchard.
Remote Sens. 12, 1–21. https://doi.org/10.3390/RS12010133

Donner, R. V., Zou, Y., Donges, J.F., Marwan, N., Kurths, J., 2010. Recurrence
networks—a novel paradigm for nonlinear time series analysis. New J. Phys. 12,
033025. https://doi.org/10.1088/1367-2630/12/3/033025

Duan, T., Chapman, S.C., Guo, Y., Zheng, B., 2017. Dynamic monitoring of NDVI in
wheat agronomy and breeding trials using an unmanned aerial vehicle. F. Crop. Res.
210, 71–80. https://doi.org/10.1016/j.fcr.2017.05.025

Duffie, J.A., Beckman, W.A., 2013. Solar Engineering of Thermal Processes, 4th ed.
Wiley, New York.

Dyson, J., Mancini, A., Frontoni, E., Zingaretti, P., 2019. Deep Learning for Soil and
Crop Segmentation from Remotely Sensed Data. Remote Sens. 11, 1859.
https://doi.org/10.3390/rs11161859

Eckmann, J.-P., Kamphorst, S.O., Ruelle, D., 1987. Recurrence Plots of Dynamical
Systems. Europhys. Lett. 4, 973–977. https://doi.org/10.1209/0295-5075/4/9/004

Eisfelder, C., Kuenzer, C., Dech, S., 2012. Derivation of biomass information for semi-
arid areas using remote-sensing data. Int. J. Remote Sens. 33, 2937–2984.
https://doi.org/10.1080/01431161.2011.620034

Estrela, T., Vargas, E., 2012. Drought Management Plans in the European Union. The
Case of Spain. Water Resour. Manag. 26, 1537–1553.
https://doi.org/10.1007/s11269-011-9971-2
Fabricante, I., Oesterheld, M., Paruelo, J.M., 2009. Annual and seasonal variation of
NDVI explained by current and previous precipitation across Northern Patagonia. J.
Arid Environ. 73, 745–753. https://doi.org/10.1016/j.jaridenv.2009.02.006

FAO, 2014. World reference base for soil resources 2014. International soil classification
system for naming soils and creating legends for soil maps, World Soil Resources
Reports No. 106. FAO, Rome.

Feng, W., Zhang, H.-Y., Zhang, Y.-S., Qi, S.-L., Heng, Y.-R., Guo, B.-B., Ma, D.-Y.,
Guo, T.-C., 2016. Remote detection of canopy leaf nitrogen concentration in winter

124
 wheat by using water resistance vegetation indices from in-situ hyperspectral data.
F. Crop. Res. 198, 238–246. https://doi.org/10.1016/j.fcr.2016.08.023

Franklin, O., Harrison, S.P., Dewar, R., Farrior, C.E., Brännström, Å., Dieckmann, U.,
Pietsch, S., Falster, D., Cramer, W., Loreau, M., Wang, H., Mäkelä, A., Rebel, K.T.,
Meron, E., Schymanski, S.J., Rovenskaya, E., Stocker, B.D., Zaehle, S., Manzoni,
S., van Oijen, M., Wright, I.J., Ciais, P., van Bodegom, P.M., Peñuelas, J., Hofhansl,
F., Terrer, C., Soudzilovskaia, N.A., Midgley, G., Prentice, I.C., 2020. Organizing
principles for vegetation dynamics. Nat. Plants 6, 444–453.
https://doi.org/10.1038/s41477-020-0655-x

Fraser, A.M., Swinney, H.L., 1986. Independent coordinates for strange attractors from
mutual information. Phys. Rev. A 33, 1134–1140.
https://doi.org/10.1103/PhysRevA.33.1134

Frilot II, C., Kim, P., Carrubba, S., McCarty, D., Jr, A., Marino, A., 2015. Analysis of
Brain Recurrence, in: Webber, C.L., Marwan, N. (Eds.), Recurrence Quantification
Analysis -- Theory and Best Practices. Springer International Publishing, Cham, pp.
213–251. https://doi.org/10.1007/978-3-319-07155-8

Fu, B., Burgher, I., 2015. Riparian vegetation NDVI dynamics and its relationship with
climate, surface water and groundwater. J. Arid Environ. 113, 59–68.
https://doi.org/10.1016/j.jaridenv.2014.09.010

Fu, Z., Jiang, J., Gao, Y., Krienke, B., Wang, M., Zhong, K., Cao, Q., Tian, Y., Zhu, Y.,
Cao, W., Liu, X., 2020. Wheat Growth Monitoring and Yield Estimation based on
Multi-Rotor Unmanned Aerial Vehicle. Remote Sens. 12, 508.
https://doi.org/10.3390/rs12030508

Gabriel, J.L., Alonso-Ayuso, M., García-González, I., Hontoria, C., Quemada, M., 2016.
Nitrogen use efficiency and fertiliser fate in a long-term experiment with winter
cover crops. Eur. J. Agron. 79, 14–22. https://doi.org/10.1016/j.eja.2016.04.015

Galloway, J.N., Townsend, A.R., Erisman, J.W., Bekunda, M., Cai, Z., Freney, J.R.,
Martinelli, L.A., Seitzinger, S.P., Sutton, M.A., 2008. Transformation of the
Nitrogen Cycle: Recent Trends, Questions, and Potential Solutions. Science (80-. ).
320, 889–892. https://doi.org/10.1126/science.1136674

Gao, Q., Zhu, W., Schwartz, M.W., Ganjurjav, H., Wan, Y., Qin, X., Ma, X., Williamson,
M.A., Li, Y., 2016. Climatic change controls productivity variation in global
grasslands. Sci. Rep. 6, 26958. https://doi.org/10.1038/srep26958

Gessner, U., Naeimi, V., Klein, I., Kuenzer, C., Klein, D., Dech, S., 2013. The
relationship between precipitation anomalies and satellite-derived vegetation
activity in Central Asia. Glob. Planet. Change 110, 74–87.
https://doi.org/10.1016/j.gloplacha.2012.09.007

Gidey, E., Dikinya, O., Sebego, R., Segosebe, E., Zenebe, A., 2018a. Analysis of the
long-term agricultural drought onset, cessation, duration, frequency, severity and

125
 spatial extent using Vegetation Health Index (VHI) in Raya and its environs,
Northern Ethiopia. Environ. Syst. Res. 7, 13. https://doi.org/10.1186/s40068-018-
0115-z

Gidey, E., Dikinya, O., Sebego, R., Segosebe, E., Zenebe, A., 2018b. Using Drought
Indices to Model the Statistical Relationships Between Meteorological and
Agricultural Drought in Raya and Its Environs, Northern Ethiopia. Earth Syst.
Environ. 2, 265–279. https://doi.org/10.1007/s41748-018-0055-9

Gong, Z., Kawamura, K., Ishikawa, N., Goto, M., Wulan, T., Alateng, D., Yin, T., Ito,
Y., 2015. MODIS normalized difference vegetation index (NDVI) and vegetation
phenology dynamics in the Inner Mongolia grassland. Solid Earth 6, 1185–1194.
https://doi.org/10.5194/se-6-1185-2015
Grant, K., Kreyling, J., Dienstbach, L.F.H., Beierkuhnlein, C., Jentsch, A., 2014. Water
stress due to increased intra-annual precipitation variability reduced forage yield but
raised forage quality of a temperate grassland. Agric. Ecosyst. Environ. 186, 11–22.
https://doi.org/10.1016/j.agee.2014.01.013

Guijarro, M., Pajares, G., Riomoros, I., Herrera, P.J., Burgos-Artizzu, X.P., Ribeiro, A.,
2011. Automatic segmentation of relevant textures in agricultural images. Comput.
Electron. Agric. 75, 75–83. https://doi.org/10.1016/j.compag.2010.09.013
Guo, B., Zhou, Y., Wang, S. xin, Tao, H. ping, 2014. The relationship between
normalized difference vegetation index (NDVI) and climate factors in the semiarid
region: A case study in Yalu Tsangpo River basin of Qinghai-Tibet Plateau. J. Mt.
Sci. 11, 926–940. https://doi.org/10.1007/s11629-013-2902-3

Guo, L., Wu, S., Zhao, D., Yin, Y., Leng, G., Zhang, Q., 2014. NDVI-Based Vegetation
Change in Inner Mongolia from 1982 to 2006 and Its Relationship to Climate at the
Biome Scale. Adv. Meteorol. 2014, 1–12. https://doi.org/10.1155/2014/692068
Hadri, A., Saidi, M.E.M., Boudhar, A., 2021. Multiscale drought monitoring and
comparison using remote sensing in a Mediterranean arid region: a case study from
west-central Morocco. Arab. J. Geosci. 14, 118. https://doi.org/10.1007/s12517-
021-06493-w

Hao, F., Zhang, X., Ouyang, W., Skidmore, A.K., Toxopeus, A.G., 2012. Vegetation
NDVI Linked to Temperature and Precipitation in the Upper Catchments of Yellow
River. Environ. Model. Assess. 17, 389–398. https://doi.org/10.1007/s10666-011-
9297-8

Hargreaves, G.H., 1994. Simplified coefficients for estimating monthly solar radiation in
North America and Europe., Dept. of Bio. and Irrig. Engrg. Logan, Utah.

Hargreaves, G.H., Samani, Z.A., 1985. Reference crop evapotranspiration from

temperature. Appl. Eng. Agric. 1, 96–99.

126
Harper, C.W., Blair, J.M., Fay, P.A., Knapp, A.K., Carlisle, J.D., 2005. Increased rainfall
variability and reduced rainfall amount decreases soil CO2 flux in a grassland
ecosystem. Glob. Chang. Biol. 11, 322–334. https://doi.org/10.1111/j.1365-
2486.2005.00899.x
Helman, D., Lensky, I.M., Tessler, N., Osem, Y., 2015. A phenology-based method for
monitoring woody and herbaceous vegetation in mediterranean forests from NDVI
time series. Remote Sens. 7, 12314–12335. https://doi.org/10.3390/rs70912314

Hengl, T., 2006. Finding the right pixel size, Computers and Geosciences.
https://doi.org/10.1016/j.cageo.2005.11.008

Herrick, J.E., Whitford, W.G., 1995. Assessing the quality of rangeland soils: challenges
and opportunities. J. Soil Water Conserv. 50, 237–242.

Hively, W., Lamb, B., Daughtry, C., Shermeyer, J., McCarty, G., Quemada, M., 2018.
Mapping Crop Residue and Tillage Intensity Using WorldView-3 Satellite
Shortwave Infrared Residue Indices. Remote Sens. 10, 1657.
https://doi.org/10.3390/rs10101657

Hobbs, R.J., Yates, S., Mooney, H.A., 2007. Long-term data reveal complex dynamics in
grassland in relation to climate and disturbance. Ecol. Monogr. 77, 545–568.
https://doi.org/10.1890/06-1530.1

Holzapfel, C., Tielbörger, K., Parag, H.A., Kigel, J., Sternberg, M., 2006. Annual plant–
shrub interactions along an aridity gradient. Basic Appl. Ecol. 7, 268–279.
https://doi.org/10.1016/j.baae.2005.08.003

Huete, A., Didan, K., Miura, T., Rodriguez, E.P., Gao, X., Ferreira, L.G., 2002. Overview
of the radiometric and biophysical performance of the MODIS vegetation indices.
Remote Sens. Environ. 83, 195–213. https://doi.org/10.1016/S0034-
4257(02)00096-2

Huete, A.R., Tucker, C.J., 1991. Investigation of soil influences in AVHRR red and near-
infrared vegetation index imagery. Int. J. Remote Sens. 12, 1223–1242.
https://doi.org/10.1080/01431169108929723

Huete, A.R., 1989. Soil influences in remotely sensed vegetation-canopy spectra, in:
Asrar, G. (Ed.), Theory and Applications of Optical Remote Sensing. Wiley, New
York, NY, USA.

Huete, A.R., 1988. A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 25,
295–309. https://doi.org/10.1016/0034-4257(88)90106-X

Huete, A.R., Jackson, R.D., Post, D.F., 1985. Spectral response of a plant canopy with
different soil backgrounds. Remote Sens. Environ. 17, 37–53.
https://doi.org/10.1016/0034-4257(85)90111-7

127
Hui, Q.L., Huete, A., 1995. A feedback based modification of the NDVI to minimize
canopy background and atmospheric noise. IEEE Trans. Geosci. Remote Sens. 33,
457–465. https://doi.org/10.1109/36.377946

Iglesias, A., 2003. Climate Changes in the Mediterranean Region: Physical Aspects and
Effects on Agriculture, in: Bolle, H.-J. (Ed.), Mediterranean Climate. Springer
Berlin Heidelberg, Berlin, Heidelberg, pp. 107–111. https://doi.org/10.1007/978-3-
642-55657-9_4

Jacobs, J.M., Mohanty, B.P., Hsu, E.C., Miller, D., 2004. SMEX02: Field scale
variability, time stability and similarity of soil moisture. Remote Sens. Environ. 92,
436–446. https://doi.org/10.1016/j.rse.2004.02.017

Jafari, R., Bashari, H., Tarkesh, M., 2017. Discriminating and monitoring rangeland
condition classes with MODIS NDVI and EVI indices in Iranian arid and semi-arid
lands. Arid L. Res. Manag. 31, 94–110.
https://doi.org/10.1080/15324982.2016.1224955

Jiang, J., Wang, C., Wang, H., Fu, Z., Cao, Q., Tian, Y., Zhu, Y., Cao, W., Liu, X., 2021.
Evaluation of Three Portable Optical Sensors for Non-Destructive Diagnosis of
Nitrogen Status in Winter Wheat. Sensors 21, 5579.
https://doi.org/10.3390/s21165579
Jiang, W., Wang, L., Feng, L., Zhang, M., Yao, R., 2020. Drought characteristics and its
impact on changes in surface vegetation from 1981 to 2015 in the Yangtze River
Basin, China. Int. J. Climatol. 40, 3380–3397. https://doi.org/10.1002/joc.6403

Jiang, Y., Tao, J., Huang, Y., Zhu, J., Tian, L., Zhang, Y., 2013. The spatial pattern of
grassland aboveground biomass on Xizang Plateau and its climatic controls. J. Plant
Ecol. 8, 30–40. https://doi.org/10.1093/jpe/rtu002

Kang, W., Wang, T., Liu, S., 2018. The Response of Vegetation Phenology and
Productivity to Drought in Semi-Arid Regions of Northern China. Remote Sens. 10,
727. https://doi.org/10.3390/rs10050727

Karnieli, A., Agam, N., Pinker, R.T., Anderson, M., Imhoff, M.L., Gutman, G.G., Panov,
N., Goldberg, A., 2010. Use of NDVI and Land Surface Temperature for Drought
Assessment: Merits and Limitations. J. Clim. 23, 618–633.
https://doi.org/10.1175/2009JCLI2900.1

Khan, S., Gabriel, H.F., Rana, T., 2008. Standard precipitation index to track drought and
assess impact of rainfall on watertables in irrigation areas. Irrig. Drain. Syst. 22,
159–177. https://doi.org/10.1007/s10795-008-9049-3

Kogan, F.N., 1997. Global Drought Watch from Space. Bull. Am. Meteorol. Soc. 78,
621–636. https://doi.org/10.1175/1520-0477(1997)078<0621:GDWFS>2.0.CO;2

128
Kogan, F.N., 1995. Application of vegetation index and brightness temperature for
drought detection. Adv. Sp. Res. 15, 91–100. https://doi.org/10.1016/0273-
1177(95)00079-T

Konings, A.G., Williams, A.P., Gentine, P., 2017. Sensitivity of grassland productivity
to aridity controlled by stomatal and xylem regulation. Nat. Geosci. 10, 284–288.
https://doi.org/10.1038/ngeo2903

Kruger, N.J., 2009. The Bradford Method For Protein Quantitation, in: Walker, J.M.
(Ed.), The Protein Protocols Handbook. Humana Press, Totowa, NJ, pp. 17–24.
https://doi.org/10.1007/978-1-59745-198-7_4

Kurtz, D.B., Schellberg, J., Braun, M., 2010. Ground and satellite based assessment of
rangeland management in sub-tropical Argentina. Appl. Geogr. 30, 210–220.
https://doi.org/10.1016/j.apgeog.2009.01.006

Laio, F., Porporato, A., Fernandez-Illescas, C.., Rodriguez-Iturbe, I., 2001. Plants in
water-controlled ecosystems: active role in hydrologic processes and response to
water stress. Adv. Water Resour. 24, 745–762. https://doi.org/10.1016/S0309-
1708(01)00007-0

Laterra, P., Orúe, M.E., Booman, G.C., 2012. Spatial complexity and ecosystem services
in rural landscapes. Agric. Ecosyst. Environ. 154, 56–67.
https://doi.org/10.1016/j.agee.2011.05.013

Li, B., Tao, S., Dawson, R.W., 2002. Relations between AVHRR NDVI and ecoclimatic
parameters in China. Int. J. Remote Sens. 23, 989–999.
https://doi.org/10.1080/014311602753474192

Li, C., Leal Filho, W., Yin, J., Hu, R., Wang, J., Yang, C., Yin, S., Bao, Y., Ayal, D.Y.,
2018. Assessing vegetation response to multi-time-scale drought across inner
Mongolia plateau. J. Clean. Prod. 179, 210–216.
https://doi.org/10.1016/j.jclepro.2018.01.113

Li, F., Mistele, B., Hu, Y., Chen, X., Schmidhalter, U., 2014. Optimising three-band
spectral indices to assess aerial N concentration, N uptake and aboveground biomass
of winter wheat remotely in China and Germany. ISPRS J. Photogramm. Remote
Sens. 92, 112–123. https://doi.org/10.1016/j.isprsjprs.2014.03.006

Li, G., Wang, J., Wang, Y., Wei, H., Ochir, A., Davaasuren, D., Chonokhuu, S., Nasanbat,
E., 2019. Spatial and Temporal Variations in Grassland Production from 2006 to
2015 in Mongolia Along the China–Mongolia Railway. Sustainability 11, 2177.
https://doi.org/10.3390/su11072177

Li, L., Mu, X., Qi, J., Pisek, J., Roosjen, P., Yan, G., Huang, H., Liu, S., Baret, F., 2021.
Characterizing reflectance anisotropy of background soil in open-canopy plantations
using UAV-based multiangular images. ISPRS J. Photogramm. Remote Sens. 177,
263–278. https://doi.org/10.1016/j.isprsjprs.2021.05.007

129
Li, S.C., Zhao, Z.Q., Liu, F.Y., 2008. Identifying spatial pattern of NDVI series dynamics
using recurrence quantification analysis. Eur. Phys. J. Spec. Top. 164, 127–139.
https://doi.org/10.1140/epjst/e2008-00839-y

Li, W., Fu, R., Juárez, R.I.N., Fernandes, K., 2008. Observed change of the standardized
precipitation index, its potential cause and implications to future climate change in
the Amazon region. Philos. Trans. R. Soc. B Biol. Sci. 363, 1767–1772.
https://doi.org/10.1098/rstb.2007.0022

Liaqat, M.U., Cheema, M.J.M., Huang, W., Mahmood, T., Zaman, M., Khan, M.M.,
2017. Evaluation of MODIS and Landsat multiband vegetation indices used for
wheat yield estimation in irrigated Indus Basin. Comput. Electron. Agric. 138, 39–
47. https://doi.org/10.1016/j.compag.2017.04.006
Lichtenthaler, H.K., Lang, M., Sowinska, M., Heisel, F., Miehé, J.A., 1996. Detection of
Vegetation Stress Via a New High Resolution Fluorescence Imaging System. J.
Plant Physiol. 148, 599–612. https://doi.org/10.1016/S0176-1617(96)80081-2

Liu, L., Wang, J., Bao, Y., Huang, W., Ma, Z., Zhao, C., 2006. Predicting winter wheat
condition, grain yield and protein content using multi‐temporal EnviSat‐ASAR and
Landsat TM satellite images. Int. J. Remote Sens. 27, 737–753.
https://doi.org/10.1080/01431160500296867
Liu, R., Sun, Y., Li, M., Zhang, M., Zhang, Z., Li, H., Yang, W., 2022. Development and
application experiments of a grain yield monitoring system. Comput. Electron.
Agric. 195, 106851. https://doi.org/10.1016/j.compag.2022.106851

Liu, S., Gao, L., Lei, Y., Wang, M., Hu, Q., Ma, X., Zhang, Y., 2021. SAR Speckle
Removal Using Hybrid Frequency Modulations. IEEE Trans. Geosci. Remote Sens.
59, 3956–3966. https://doi.org/10.1109/TGRS.2020.3014130

Liu, S., Peng, D., Zhang, B., Chen, Z., Yu, L., Chen, J., Pan, Y., Zheng, S., Hu, J., Lou,
Z., Chen, Y., Yang, S., 2022. The Accuracy of Winter Wheat Identification at
Different Growth Stages Using Remote Sensing. Remote Sens. 14, 893.
https://doi.org/10.3390/rs14040893

Liu, S., Zhang, Y., Cheng, F., Hou, X., Zhao, S., 2017. Response of Grassland
Degradation to Drought at Different Time-Scales in Qinghai Province: Spatio-
Temporal Characteristics, Correlation, and Implications. Remote Sens. 9, 1329.
https://doi.org/10.3390/rs9121329
Liu, Y., Li, Y., Li, S., Motesharrei, S., 2015. Spatial and temporal patterns of global NDVI
trends: Correlations with climate and human factors. Remote Sens. 7, 13233–13250.
https://doi.org/10.3390/rs71013233

Liu, Y., Zha, Y., Gao, J., Ni, S., 2004. Assessment of grassland degradation near Lake
Qinghai, West China, using Landsat TM and in situ reflectance spectra data. Int. J.
Remote Sens. 25, 4177–4189. https://doi.org/10.1080/01431160410001680419

130
López-Granados, F., 2011. Weed detection for site-specific weed management: Mapping
and real-time approaches. Weed Res. 51, 1–11. https://doi.org/10.1111/j.1365-
3180.2010.00829.x

Loukas, A, Vasiliades, L., Dalezios, N.R., 2003. Intercomparison of Meteorological

Drought Indices for Drought Assessment and Monitoring in Greece. 8th Int. Conf.
Environ. Sci. Technol. 8, 484–491.

Ma, X., Huete, A., Yu, Q., Coupe, N.R., Davies, K., Broich, M., Ratana, P., Beringer, J.,
Hutley, L.B., Cleverly, J., Boulain, N., Eamus, D., 2013. Spatial patterns and
temporal dynamics in savanna vegetation phenology across the north australian
tropical transect. Remote Sens. Environ. 139, 97–115.
https://doi.org/10.1016/j.rse.2013.07.030
Magney, T.S., Eitel, J.U.H., Vierling, L.A., 2017. Mapping wheat nitrogen uptake from
RapidEye vegetation indices. Precis. Agric. 18, 429–451.
https://doi.org/10.1007/s11119-016-9463-8

Magney, T.S., Eitel, J.U.H., Huggins, D.R., Vierling, L.A., 2016. Proximal NDVI derived
phenology improves in-season predictions of wheat quantity and quality. Agric. For.
Meteorol. 217, 46–60. https://doi.org/10.1016/j.agrformet.2015.11.009

Malik, A., Kumar, A., 2021. Application of standardized precipitation index for
monitoring meteorological drought and wet conditions in Garhwal region
(Uttarakhand). Arab. J. Geosci. 14, 800. https://doi.org/10.1007/s12517-021-07158-
4

Mao, D., Wang, Z., Luo, L., Ren, C., 2012. Integrating AVHRR and MODIS data to
monitor NDVI changes and their relationships with climatic parameters in Northeast
China. Int. J. Appl. Earth Obs. Geoinf. 18, 528–536.
https://doi.org/10.1016/j.jag.2011.10.007
Marini, G., Fontana, N., Mishra, A.K., 2019. Investigating drought in Apulia region, Italy
using SPI and RDI. Theor. Appl. Climatol. 137, 383–397.
https://doi.org/10.1007/s00704-018-2604-4

Marszalek, M., Körner, M., Schmidhalter, U., 2022. Prediction of multi-year winter
wheat yields at the field level with satellite and climatological data. Comput.
Electron. Agric. 194. https://doi.org/10.1016/j.compag.2022.106777

Martín-Sotoca, J.J., Saa-Requejo, A., Moratiel, R., Dalezios, N., Faraslis, I., Tarquis,
A.M., 2019. Statistical analysis for satellite-index-based insurance to define
damaged pasture thresholds. Nat. Hazards Earth Syst. Sci. 19, 1685–1702.
https://doi.org/10.5194/nhess-19-1685-2019

Martínez, B., Gilabert, M.A., 2009. Vegetation dynamics from NDVI time series analysis
using the wavelet transform. Remote Sens. Environ. 113, 1823–1842.
https://doi.org/10.1016/j.rse.2009.04.016

131
Marwan, N., Kurths, J., Foerster, S., 2015. Analysing spatially extended high-
dimensional dynamics by recurrence plots. Phys. Lett. A 379, 894–900.
https://doi.org/10.1016/j.physleta.2015.01.013

Marwan, N., Carmen Romano, M., Thiel, M., Kurths, J., 2007. Recurrence plots for the
analysis of complex systems. Phys. Rep. 438, 237–329.
https://doi.org/10.1016/j.physrep.2006.11.001

Marwan, N., 2007. CRP Toolbox 5.22 (R32.4) [WWW Document]. URL http://tocsy.pik-
potsdam.de/CRPtoolbox/ (accessed 6.28.19).

Marwan, N., Kurths, J., 2002. Nonlinear analysis of bivariate data with cross recurrence
plots. Phys. Lett. Sect. A Gen. At. Solid State Phys. 302, 299–307.
https://doi.org/10.1016/S0375-9601(02)01170-2

Marwan, N., Wessel, N., Meyerfeldt, U., Schirdewan, A., Kurths, J., 2002. Recurrence-
plot-based measures of complexity and their application to heart-rate-variability
data. Phys. Rev. E 66, 026702. https://doi.org/10.1103/PhysRevE.66.026702
Meroni, M., Rembold, F., Verstraete, M.M., Gommes, R., Schucknecht, A., Beye, G.,
2014. Investigating the relationship between the inter-annual variability of satellite-
derived vegetation phenology and a proxy of biomass production in the Sahel.
Remote Sens. 6, 5868–5884. https://doi.org/10.3390/rs6065868

Mishra, A.K., Singh, V.P., 2010. A review of drought concepts. J. Hydrol. 391, 202–216.
https://doi.org/10.1016/j.jhydrol.2010.07.012

Mishra, A.K., Singh, V.P., Desai, V.R., 2009. Drought characterization: a probabilistic
approach. Stoch. Environ. Res. Risk Assess. 23, 41–55.
https://doi.org/10.1007/s00477-007-0194-2

Mkhabela, M.S., Bullock, P., Raj, S., Wang, S., Yang, Y., 2011. Crop yield forecasting
on the Canadian Prairies using MODIS NDVI data. Agric. For. Meteorol. 151, 385–
393. https://doi.org/10.1016/j.agrformet.2010.11.012

Mu, S., Yang, H., Li, J., Chen, Y., Gang, C., Zhou, W., Ju, W., 2013. Spatio-temporal
dynamics of vegetation coverage and its relationship with climate factors in Inner
Mongolia, China. J. Geogr. Sci. 23, 231–246. https://doi.org/10.1007/s11442-013-
1006-x

Mulla, D.J., 2013. Twenty five years of remote sensing in precision agriculture: Key
advances and remaining knowledge gaps. Biosyst. Eng. 114, 358–371.
https://doi.org/10.1016/j.biosystemseng.2012.08.009

Nanzad, L., Zhang, J., Tuvdendorj, B., Nabil, M., Zhang, S., Bai, Y., 2019. NDVI
anomaly for drought monitoring and its correlation with climate factors over
Mongolia from 2000 to 2016. J. Arid Environ. 164, 69–77.
https://doi.org/10.1016/j.jaridenv.2019.01.019

132
Olmos-Trujillo, E., González-Trinidad, J., Júnez-Ferreira, H., Pacheco-Guerrero, A.,
Bautista-Capetillo, C., Avila-Sandoval, C., Galván-Tejada, E., 2020. Spatio-
Temporal Response of Vegetation Indices to Rainfall and Temperature in A
Semiarid Region. Sustainability 12, 1939. https://doi.org/10.3390/su12051939
Orlando, F., Marta, A.D., Mancini, M., Motha, R., Qu, J.J., Orlandini, S., 2015.
Integration of Remote Sensing and Crop Modeling for the Early Assessment of
Durum Wheat Harvest at the Field Scale. Crop Sci. 55, 1280–1289.
https://doi.org/10.2135/cropsci2014.07.0479

Orth, R., Seneviratne, S.I., 2012. Analysis of soil moisture memory from observations in
Europe. J. Geophys. Res. Atmos. 117, n/a-n/a.
https://doi.org/10.1029/2011JD017366
Otsu, N., 1979. A Threshold Selection Method from Gray-Level Histograms. IEEE Trans.
Syst. Man. Cybern. 9, 62–66. https://doi.org/10.1109/TSMC.1979.4310076

Pan, Z., Huang, J., Zhou, Q., Wang, L., Cheng, Y., Zhang, H., Blackburn, G.A., Yan, J.,
Liu, J., 2015. Mapping crop phenology using NDVI time-series derived from HJ-1
A/B data. Int. J. Appl. Earth Obs. Geoinf. 34, 188–197.
https://doi.org/10.1016/j.jag.2014.08.011

Pancorbo, J.L., Lamb, B.T., Quemada, M., Hively, W.D., Gonzalez-Fernandez, I.,
Molina, I., 2021a. Sentinel-2 and WorldView-3 atmospheric correction and signal
normalization based on ground-truth spectroradiometric measurements. ISPRS J.
Photogramm. Remote Sens. 173, 166–180.
https://doi.org/10.1016/j.isprsjprs.2021.01.009

Pancorbo, J.L., Camino, C., Alonso-Ayuso, M., Raya-Sereno, M.D., Gonzalez-

Fernandez, I., Gabriel, J.L., Zarco-Tejada, P.J., Quemada, M., 2021b. Simultaneous
assessment of nitrogen and water status in winter wheat using hyperspectral and
thermal sensors. Eur. J. Agron. 127, 126287.
https://doi.org/10.1016/j.eja.2021.126287

Pang, G., Wang, X., Yang, M., 2017. Using the NDVI to identify variations in, and
responses of, vegetation to climate change on the Tibetan Plateau from 1982 to 2012.
Quat. Int. 444, 87–96. https://doi.org/10.1016/j.quaint.2016.08.038

Patro, S.G.K., Sahu, K.K., 2015. Normalization: A Preprocessing Stage. Iarjset 20–22.
https://doi.org/10.17148/iarjset.2015.2305
Peco, B., Sánchez, A.M., Azcárate, F.M., 2006. Abandonment in grazing systems:
Consequences for vegetation and soil. Agric. Ecosyst. Environ. 113, 284–294.
https://doi.org/10.1016/j.agee.2005.09.017

Pei, Z., Fang, S., Wang, L., Yang, W., 2020. Comparative Analysis of Drought Indicated
by the SPI and SPEI at Various Timescales in Inner Mongolia, China. Water 12,
1925. https://doi.org/10.3390/w12071925

133
Pei, Z., Fang, S., Yang, W., Wang, L., Wu, M., Zhang, Q., Han, W., Khoi, D.N., 2019.
The Relationship between NDVI and Climate Factors at Different Monthly Time
Scales: A Case Study of Grasslands in Inner Mongolia, China (1982–2015).
Sustainability 11, 7243. https://doi.org/10.3390/su11247243
Pérez-Hoyos, A., Martínez, B., Gilabert, M.A., García-Haro, F.J., 2010. Multi-temporal
analysis of vegetation dynamics in the Iberian peninsula using MODIS-NDVI data.
EARSeL eProceedings.

Piao, S., Mohammat, A., Fang, J., Cai, Q., Feng, J., 2006. NDVI-based increase in growth
of temperate grasslands and its responses to climate changes in China. Glob.
Environ. Chang. 16, 340–348. https://doi.org/10.1016/j.gloenvcha.2006.02.002

Piedallu, C., Chéret, V., Denux, J.P., Perez, V., Azcona, J.S., Seynave, I., Gégout, J.C.,
2019. Soil and climate differently impact NDVI patterns according to the season and
the stand type. Sci. Total Environ. 651, 2874–2885.
https://doi.org/10.1016/j.scitotenv.2018.10.052

Potgieter, A.B., George-Jaeggli, B., Chapman, S.C., Laws, K., Suárez Cadavid, L.A.,
Wixted, J., Watson, J., Eldridge, M., Jordan, D.R., Hammer, G.L., 2017. Multi-
Spectral Imaging from an Unmanned Aerial Vehicle Enables the Assessment of
Seasonal Leaf Area Dynamics of Sorghum Breeding Lines. Front. Plant Sci. 8, 1–
11. https://doi.org/10.3389/fpls.2017.01532

Prey, L., von Bloh, M., Schmidhalter, U., 2018. Evaluating RGB Imaging and
Multispectral Active and Hyperspectral Passive Sensing for Assessing Early Plant
Vigor in Winter Wheat. Sensors 18, 2931. https://doi.org/10.3390/s18092931

Proulx, R., Parrott, L., Fahrig, L., Currie, D.J., 2015. Long Time-Scale Recurrences in
Ecology: Detecting Relationships Between Climate Dynamics and Biodiversity
Along a Latitudinal Gradient, in: Webber, C.L., Marwan, N. (Eds.), Recurrence
Quantification Analysis -- Theory and Best Practices, Understanding Complex
Systems. Springer International Publishing, Cham, Switzerland, pp. 335–347.
https://doi.org/10.1007/978-3-319-07155-8

Proulx, R., Parrott, L., 2009. Structural complexity in digital images as an ecological
indicator for monitoring forest dynamics across scale, space and time. Ecol. Indic.
9, 1248–1256. https://doi.org/10.1016/j.ecolind.2009.03.015

Proulx, R., Cöté, P., Parrott, L., 2008. Use of recurrence analysis to measure the
dynamical stability of a multi-species community model. Eur. Phys. J. Spec. Top.
164, 117–126. https://doi.org/10.1140/epjst/e2008-00838-0

Qi, J., Chehbouni, A., Huete, A.R., Kerr, Y.H., Sorooshian, S., 1994. A modified soil
adjusted vegetation index. Remote Sens. Environ. 48, 119–126.
https://doi.org/10.1016/0034-4257(94)90134-1

Qiao, L., Tang, W., Gao, D., Zhao, R., An, L., Li, M., Sun, H., Song, D., 2022. UAV-
based chlorophyll content estimation by evaluating vegetation index responses under

134
 different crop coverages. Comput. Electron. Agric. 196, 106775.
https://doi.org/10.1016/j.compag.2022.106775

Qiao, L., Gao, D., Zhang, J., Li, M., Sun, H., Ma, J., 2020. Dynamic Influence Elimination
and Chlorophyll Content Diagnosis of Maize Using UAV Spectral Imagery. Remote
Sens. 12, 2650. https://doi.org/10.3390/rs12162650

Quemada, M., Hively, W.D., Daughtry, C.S.T., Lamb, B.T., Shermeyer, J., 2018.
Improved crop residue cover estimates obtained by coupling spectral indices for
residue and moisture. Remote Sens. Environ. 206, 33–44.
https://doi.org/https://doi.org/10.1016/j.rse.2017.12.012

Quemada, M., Daughtry, C.S.T., 2016. Spectral indices to improve crop residue cover
estimation under varying moisture conditions. Remote Sens. 8, 660.
https://doi.org/10.3390/rs8080660

Quemada, M., 2005. Predicting crop residue decomposition using moisture adjusted time
scales. Nutr. Cycl. Agroecosystems 70, 283–291. https://doi.org/10.1007/s10705-
005-0533-y

R Core Team, 2021. R: A language and environment for statistical computing. R

Foundation for Statistical Computing, Vienna, Austria.

Ramirez-Garcia, J., Almendros, P., Quemada, M., 2012. Ground cover and leaf area index
relationship in a grass, legume and crucifer crop. Plant, Soil Environ. 58, 385–390.
https://doi.org/10.17221/195/2012-PSE

Ramos, M.C., 2001. Rainfall distribution patterns and their change over time in a
Mediterranean area. Theor. Appl. Climatol. 69, 163–170.
https://doi.org/10.1007/s007040170022

Raya-Sereno, M.D., Alonso-Ayuso, M., Pancorbo, J.L., Gabriel, J.L., Camino, C., Zarco-
Tejada, P.J., Quemada, M., 2022. Residual Effect and N Fertilizer Rate Detection
by High-Resolution VNIR-SWIR Hyperspectral Imagery and Solar-Induced
Chlorophyll Fluorescence in Wheat. IEEE Trans. Geosci. Remote Sens. 60, 1–17.
https://doi.org/10.1109/TGRS.2021.3099624

Raya-Sereno, M.D., Ortiz-Monasterio, J.I., Alonso-Ayuso, M., Rodrigues, F.A.,

Rodríguez, A.A., González-Perez, L., Quemada, M., 2021. High-Resolution
Airborne Hyperspectral Imagery for Assessing Yield, Biomass, Grain N
Concentration, and N Output in Spring Wheat. Remote Sens. 13, 1373.
https://doi.org/10.3390/rs13071373

Reinermann, S., Asam, S., Kuenzer, C., 2020. Remote Sensing of Grassland Production
and Management—A Review. Remote Sens. 12, 1949.
https://doi.org/10.3390/rs12121949

Ren, H., Zhou, G., Zhang, F., 2018. Using negative soil adjustment factor in soil-adjusted
vegetation index (SAVI) for aboveground living biomass estimation in arid

135
 grasslands. Remote Sens. Environ. 209, 439–445.
https://doi.org/10.1016/j.rse.2018.02.068

Řezník, T., Pavelka, T., Herman, L., Lukas, V., Širůček, P., Leitgeb, Š., Leitner, F., 2020.
Prediction of Yield Productivity Zones from Landsat 8 and Sentinel-2A/B and Their
Evaluation Using Farm Machinery Measurements. Remote Sens. 12, 1917.
https://doi.org/10.3390/rs12121917

Richard, Y., Poccard, I., 1998. A statistical study of NDVI sensitivity to seasonal and
interannual rainfall variations in Southern Africa. Int. J. Remote Sens. 19, 2907–
2920. https://doi.org/10.1080/014311698214343

Rivas-Tabares, D., Tarquis, A.M., Willaarts, B., De Miguel, Á., 2019. An accurate
evaluation of water availability in sub-arid Mediterranean watersheds through
SWAT: Cega-Eresma-Adaja. Agric. Water Manag. 212, 211–225.
https://doi.org/10.1016/j.agwat.2018.09.012

Rodriguez-Iturbe, I., 2000. Ecohydrology: A hydrologic perspective of climate-soil-

vegetation dynamies. Water Resour. Res. 36, 3–9.
https://doi.org/10.1029/1999WR900210

Roschewitz, I., Thies, C., Tscharntke, T., 2005. Are landscape complexity and farm
specialisation related to land-use intensity of annual crop fields? Agric. Ecosyst.
Environ. 105, 87–99. https://doi.org/10.1016/j.agee.2004.05.010

Rouse J., J.W., Haas, R.H., Schell, J.A., Deering, D.W., 1974. Monitoring Vegetation
Systems in the Great Plains with ERTS, in: NASA Special Publication. p. 309.

Sala, O.E., Parton, W.J., Joyce, L.A., Lauenroth, W.K., 1988. Primary Production of the
Central Grassland Region of the United States. Ecology 69, 40–45.
https://doi.org/10.2307/1943158

Salazar, L., Kogan, F., Roytman, L., 2007. Use of remote sensing data for estimation of
winter wheat yield in the United States. Int. J. Remote Sens. 28, 3795–3811.
https://doi.org/10.1080/01431160601050395

San Miguel-Ayanz, A., Barbeito Sanchez, I., Roig Gomez, S., Rodríguez Rojo, M.P.,
2009. Los pastos en la Comunidad de Madrid. Tipología, Cartografía y Evaluación.
Dirección General de Medio Ambiente. Conserjería de medio ambiente, vivienda y
ordenación del territorio, Madrid, Spain.

Sanz, E., Saa-Requejo, A., Díaz-Ambrona, C.H., Ruiz-Ramos, M., Rodríguez, A.,
Iglesias, E., Esteve, P., Soriano, B., Tarquis, A.M., 2021a. Normalized Difference
Vegetation Index Temporal Responses to Temperature and Precipitation in Arid
Rangelands. Remote Sens. 13, 840. https://doi.org/10.3390/rs13050840

Sanz, E., Saa-Requejo, A., Díaz-Ambrona, C.H., Ruiz-Ramos, M., Rodríguez, A.,
Iglesias, E., Esteve, P., Soriano, B., Tarquis, A.M., 2021b. Generalized Structure
Functions and Multifractal Detrended Fluctuation Analysis Applied to Vegetation

136
 Index Time Series: An Arid Rangeland Study. Entropy 23, 576.
https://doi.org/10.3390/e23050576

Saxton, K.E., Rawls, W.J., 2006. Soil Water Characteristic Estimates by Texture and
Organic Matter for Hydrologic Solutions. Soil Sci. Soc. Am. J. 70, 1569–1578.
https://doi.org/10.2136/sssaj2005.0117

Scanlon, T.M., Caylor, K.K., Manfreda, S., Levin, S.A., Rodriguez-Iturbe, I., 2005.
Dynamic response of grass cover to rainfall variability: implications for the function
and persistence of savanna ecosystems. Adv. Water Resour. 28, 291–302.
https://doi.org/10.1016/j.advwatres.2004.10.014

Scheuring, I., Riedi, R.H., 1994. Application of multifractals to the analysis of vegetation
pattern. J. Veg. Sci. 5, 489–496. https://doi.org/10.2307/3235975

Schmid, T., Millán, R., Lago, C., Trueba, C., 2000. Caracterización edafológica e índices
de vulnerabilidad de la Comunidad Autónoma de Madrid. Escala 1:200.000. Madrid.

Schmidtlein, S., Sassin, J., 2004. Mapping of continuous floristic gradients in grasslands
using hyperspectral imagery. Remote Sens. Environ. 92, 126–138.
https://doi.org/10.1016/j.rse.2004.05.004

Semeraro, T., Luvisi, A., Lillo, A.O., Aretano, R., Buccolieri, R., Marwan, N., 2020.
Recurrence Analysis of Vegetation Indices for Highlighting the Ecosystem
Response to Drought Events: An Application to the Amazon Forest. Remote Sens.
12, 907. https://doi.org/10.3390/rs12060907

Serrano, J., Shahidian, S., Marques da Silva, J., 2019. Evaluation of Normalized
Difference Water Index as a Tool for Monitoring Pasture Seasonal and Inter-Annual
Variability in a Mediterranean Agro-Silvo-Pastoral System. Water 11, 62.
https://doi.org/10.3390/w11010062

Shen, B., Fang, S., Li, G., 2014. Vegetation Coverage Changes and Their Response to
Meteorological Variables from 2000 to 2009 in Naqu, Tibet, China. Can. J. Remote
Sens. 40, 67–74. https://doi.org/10.1080/07038992.2014.917580

Steele-Dunne, S.C., McNairn, H., Monsivais-Huertero, A., Judge, J., Liu, P.W.,
Papathanassiou, K., 2017. Radar Remote Sensing of Agricultural Canopies: A
Review. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 10, 2249–2273.
https://doi.org/10.1109/JSTARS.2016.2639043

Storch, D., Gaston, K.J., 2004. Untangling ecological complexity on different scales of
space and time. Basic Appl. Ecol. 5, 389–400.
https://doi.org/10.1016/j.baae.2004.08.001

Sun, J., Qin, X., 2016. Precipitation and temperature regulate the seasonal changes of
NDVI across the Tibetan Plateau. Environ. Earth Sci. 75, 291.
https://doi.org/10.1007/s12665-015-5177-x

137
Suzuki, R., Masuda, K., G. Dye, D., 2007. Interannual covariability between actual
evapotranspiration and PAL and GIMMS NDVIs of northern Asia. Remote Sens.
Environ. 106, 387–398. https://doi.org/10.1016/j.rse.2006.10.016

Swemmer, A.M., Knapp, A.K., Snyman, H.A., 2007. Intra-seasonal precipitation patterns
and above-ground productivity in three perennial grasslands. J. Ecol. 95, 780–788.
https://doi.org/10.1111/j.1365-2745.2007.01237.x

Syta, A., Grzegorz, L., 2015. Vibration Analysis in Cutting Materials, in: Webber, C.L.,
Marwan, N. (Eds.), Recurrence Quantification Analysis -- Theory and Best
Practices. Springer International Publishing, Cham, pp. 279–290.
https://doi.org/10.1007/978-3-319-07155-8

Takens, F., 1981. Detecting strange attractors in turbulence, in: Lecture Notes in
Mathematics. pp. 366–381. https://doi.org/10.1007/bfb0091924

Tan, C., Yang, J., Li, M., 2015. Temporal-Spatial Variation of Drought Indicated by SPI
and SPEI in Ningxia Hui Autonomous Region, China. Atmosphere (Basel). 6, 1399–
1421. https://doi.org/10.3390/atmos6101399

Tarquis, A.M., Castellanos, M.T., Cartagena, M.C., Arce, A., Ribas, F., Cabello, M.J.,
López De Herrera, J., Bird, N.R.A., 2017. Scale and space dependencies of soil
nitrogen variability. Nonlinear Process. Geophys. 24, 77–87.
https://doi.org/10.5194/npg-24-77-2017

Thiel, M., Romano, M.C., Kurths, J., 2004. How much information is contained in a
recurrence plot? Phys. Lett. A 330, 343–349.
https://doi.org/10.1016/j.physleta.2004.07.050

Thiel, M., Romano, M.C., Kurths, J., Meucci, R., Allaria, E., Arecchi, F.T., 2002.
Influence of observational noise on the recurrence quantification analysis. Phys. D
Nonlinear Phenom. 171, 138–152. https://doi.org/10.1016/S0167-2789(02)00586-9

Thomasson, J.A., Sui, R., Cox, M.S., Al-Rajehy, A., 2001. Soil reflectance sensing for
determining soil properties in precision agriculture. Trans. Am. Soc. Agric. Eng. 44,
1445–1453. https://doi.org/10.13031/2013.7002

Torres-Sánchez, J., López-Granados, F., Peña, J.M., 2015. An automatic object-based

method for optimal thresholding in UAV images: Application for vegetation
detection in herbaceous crops. Comput. Electron. Agric. 114, 43–52.
https://doi.org/10.1016/j.compag.2015.03.019

Tremblay, N., Wang, Z., Ma, B.-L., Belec, C., Vigneault, P., 2009. A comparison of crop
data measured by two commercial sensors for variable-rate nitrogen application.
Precis. Agric. 10, 145–161. https://doi.org/10.1007/s11119-008-9080-2

Vannoppen, A., Gobin, A., 2021. Estimating farm wheat yields from NDVI and
meteorological data. Agronomy 11. https://doi.org/10.3390/agronomy11050946

138
Vaudour, J., Birot, P., Sudries, J., 1979. Etude comparée de la géomorphologie de la
Sierra de Guadarrama (s.l.) et de la Région de Tolède. [Compte rendu de l’Excursion
de la Commission d’Etude des Ensembles cristallins du Comité National de
Géographie (20-28 mai 1978)]. Méditerranée 36, 71–84.
https://doi.org/10.3406/medit.1979.2184

Verbesselt, J., Hyndman, R., Newnham, G., Culvenor, D., 2010. Detecting trend and
seasonal changes in satellite image time series. Remote Sens. Environ. 114, 106–
115. https://doi.org/10.1016/j.rse.2009.08.014

Vermote, E., 2015. MOD09A1 MODIS/Terra Surface Reflectance 8-Day L3 Global

500m SIN Grid V006. [WWW Document].
https://doi.org/https://doi.org/10.5067/MODIS/MOD09A1.006
Vermote, E.F., Saleous, N., 2006. Operational Atmospheric Correction of MODIS
Visible to Middle Infrared Land Surface Data in the Case of an Infinite Lambertian
Target BT, in: Qu, J.J., Gao, W., Kafatos, M., Murphy, R.E., Salomonson, V. V
(Eds.), Earth Science Satellite Remote Sensing: Vol. 1: Science and Instruments.
Springer Berlin Heidelberg, Berlin, Heidelberg, pp. 123–153.
https://doi.org/10.1007/978-3-540-37293-6_8

Vermote, E.F., El Saleous, N.Z., Justice, C.O., 2002. Atmospheric correction of MODIS
data in the visible to middle infrared: first results. Remote Sens. Environ. 83, 97–
111. https://doi.org/https://doi.org/10.1016/S0034-4257(02)00089-5

Vicente-Serrano, S.M., Azorin-Molina, C., Peña-Gallardo, M., Tomas-Burguera, M.,

Domínguez-Castro, F., Martín-Hernández, N., Beguería, S., El Kenawy, A.,
Noguera, I., García, M., 2019. A high-resolution spatial assessment of the impacts
of drought variability on vegetation activity in Spain from 1981 to 2015. Nat.
Hazards Earth Syst. Sci. 19, 1189–1213. https://doi.org/10.5194/nhess-19-1189-
2019

Vicente-Serrano, S.M., Tomas-Burguera, M., Beguería, S., Reig, F., Latorre, B., Peña-
Gallardo, M., Luna, M.Y., Morata, A., González-Hidalgo, J.C., 2017. A High
Resolution Dataset of Drought Indices for Spain. Data 2, 22.
https://doi.org/10.3390/data2030022

Vicente-Serrano, S.M., Beguería, S., López-Moreno, J.I., 2010. A Multiscalar Drought

Index Sensitive to Global Warming: The Standardized Precipitation
Evapotranspiration Index. J. Clim. 23, 1696–1718.
https://doi.org/10.1175/2009JCLI2909.1

Viola, F., Daly, E., Vico, G., Cannarozzo, M., Porporato, A., 2008. Transient soil-
moisture dynamics and climate change in Mediterranean ecosystems. Water Resour.
Res. 44, 1–12. https://doi.org/10.1029/2007WR006371

Wan, Z., Hook, S., Hulley, G., 2015. MOD11A2 MODIS/Terra Land Surface
Temperature/Emissivity 8-Day L3 Global 1km SIN Grid V006. [WWW Document].

139
 NASA EOSDIS L. Process. DAAC. URL
https://doi.org/10.5067/MODIS/MOD11A2.006 (accessed 5.12.21).

Wang, H., Liu, D., Lin, H., Montenegro, A., Zhu, X., 2015. NDVI and vegetation
phenology dynamics under the influence of sunshine duration on the Tibetan plateau.
Int. J. Climatol. 35, 687–698. https://doi.org/10.1002/joc.4013

Wang, J., Rich, P.M., Price, K.P., 2003. Temporal responses of NDVI to precipitation
and temperature in the central Great Plains, USA. Int. J. Remote Sens. 24, 2345–
2364. https://doi.org/10.1080/01431160210154812

Wang, X., Ge, L., Li, X., 2013. Pasture Monitoring Using SAR with COSMO-SkyMed,
ENVISAT ASAR, and ALOS PALSAR in Otway, Australia. Remote Sens. 5, 3611–
3636. https://doi.org/10.3390/rs5073611

Webber, C.L., Zbilut, J., 2005. Recurrence quantification analysis of nonlinear dynamical
systems. Tutorials Contemp. nonlinear methods Behav. Sci. Web B. 26–94.
https://doi.org/10.1007/s00213-012-2719-8
Webber, C.L., Zbilut, J.P., 1994. Dynamical assessment of physiological systems and
states using recurrence plot strategies. J. Appl. Physiol. 76, 965–973.
https://doi.org/10.1152/jappl.1994.76.2.965

Wen, X., Wu, X., Gao, M., 2017. Spatiotemporal variability of temperature and
precipitation in Gansu Province (Northwest China) during 1951–2015. Atmos. Res.
197, 132–149. https://doi.org/10.1016/j.atmosres.2017.07.001

Wu, T., Feng, F., Lin, Q., Bai, H., 2019. Advanced Method to Capture the Time-Lag
Effects between Annual NDVI and Precipitation Variation Using RNN in the Arid
and Semi-Arid Grasslands. Water 11, 1789. https://doi.org/10.3390/w11091789

Xie, Q., Huang, W., Zhang, B., Chen, P., Song, X., Pascucci, S., Pignatti, S., Laneve, G.,
Dong, Y., 2016. Estimating Winter Wheat Leaf Area Index from Ground and
Hyperspectral Observations Using Vegetation Indices. IEEE J. Sel. Top. Appl. Earth
Obs. Remote Sens. 9, 771–780. https://doi.org/10.1109/JSTARS.2015.2489718

Xu, D., Guo, X., 2013. A study of soil line simulation from landsat images in mixed
grassland. Remote Sens. 5, 4533–4550. https://doi.org/10.3390/rs5094533
Xu, M., Eckstein, Y., 1995. Use of Weighted Least‐Squares Method in Evaluation of the
Relationship Between Dispersivity and Field Scale. Groundwater.
https://doi.org/10.1111/j.1745-6584.1995.tb00035.x

Xu, Y., Yang, J., Chen, Y., 2016. NDVI-based vegetation responses to climate change in
an arid area of China. Theor. Appl. Climatol. 126, 213–222.
https://doi.org/10.1007/s00704-015-1572-1

Xue, J., Su, B., 2017. Significant remote sensing vegetation indices: A review of
developments and applications. J. Sensors 2017.
https://doi.org/10.1155/2017/1353691

140
Xue, L.-H., Cao, W.-X., Yang, L.-Z., 2007. Predicting Grain Yield and Protein Content
in Winter Wheat at Different N Supply Levels Using Canopy Reflectance Spectra.
Pedosphere 17, 646–653. https://doi.org/10.1016/S1002-0160(07)60077-0

Yagci, A.L., Liping Di, Meixia Deng, 2014. The influence of land cover-related changes
on the NDVI-based satellite agricultural drought indices, in: 2014 IEEE Geoscience
and Remote Sensing Symposium. IEEE, Québec, QC, Canada, 13-18 July, pp. 2054–
2057. https://doi.org/10.1109/IGARSS.2014.6946868

Yang, L., Wylie, B.K., Tieszen, L.L., Reed, B.C., 1998. An analysis of relationships
among climate forcing and time-integrated NDVI of grasslands over the U.S.
northern and central Great Plains. Remote Sens. Environ. 65, 25–37.
https://doi.org/10.1016/S0034-4257(98)00012-1
Yang, M., Hassan, M.A., Xu, K., Zheng, C., Rasheed, A., Zhang, Y., Jin, X., Xia, X.,
Xiao, Y., He, Z., 2020. Assessment of Water and Nitrogen Use Efficiencies Through
UAV-Based Multispectral Phenotyping in Winter Wheat. Front. Plant Sci. 11, 1–16.
https://doi.org/10.3389/fpls.2020.00927

Ye, X., Li, Y., Li, X., Xu, C., Zhang, Q., 2015. Investigation of the Variability and
Implications of Meteorological Dry/Wet Conditions in the Poyang Lake Catchment,
China, during the Period 1960–2010. Adv. Meteorol. 2015, 1–11.
https://doi.org/10.1155/2015/928534

Zarco-Tejada, P.J., Hornero, A., Beck, P.S.A., Kattenborn, T., Kempeneers, P.,
Hernández-Clemente, R., 2019. Chlorophyll content estimation in an open-canopy
conifer forest with Sentinel-2A and hyperspectral imagery in the context of forest
decline. Remote Sens. Environ. 223, 320–335.
https://doi.org/10.1016/j.rse.2019.01.031

Zarco‐Tejada, P.J., Ustin, S.L., Whiting, M.L., 2005. Temporal and Spatial Relationships
between Within‐Field Yield Variability in Cotton and High‐Spatial Hyperspectral
Remote Sensing Imagery. Agron. J. 97, 641–653.
https://doi.org/10.2134/agronj2003.0257

Zarei, A.R., Shabani, A., Moghimi, M.M., 2021. Accuracy Assessment of the SPEI, RDI
and SPI Drought Indices in Regions of Iran with Different Climate Conditions. Pure
Appl. Geophys. 178, 1387–1403. https://doi.org/10.1007/s00024-021-02704-3

Zhang, J., Xie, T., Yang, C., Song, H., Jiang, Z., Zhou, G., Zhang, D., Feng, H., Xie, J.,
2020. Segmenting purple rapeseed leaves in the field from UAV RGB imagery using
deep learning as an auxiliary means for nitrogen stress detection. Remote Sens. 12.
https://doi.org/10.3390/RS12091403

Zhang, Q., Kong, D., Singh, V.P., Shi, P., 2017. Response of vegetation to different time-
scales drought across China: Spatiotemporal patterns, causes and implications. Glob.
Planet. Change 152, 1–11. https://doi.org/10.1016/j.gloplacha.2017.02.008

141
Zhang, S., Chen, H., Fu, Y., Niu, H., Yang, Y., Zhang, B., 2019. Fractional Vegetation
Cover Estimation of Different Vegetation Types in the Qaidam Basin. Sustainability
11, 864. https://doi.org/10.3390/su11030864

Zhang, Y., Wang, X., Li, C., Cai, Y., Yang, Z., Yi, Y., 2018. NDVI dynamics under
changing meteorological factors in a shallow lake in future metropolitan, semiarid
area in North China. Sci. Rep. 8, 1–13. https://doi.org/10.1038/s41598-018-33968-
w

Zhao, A., Zhang, A., Cao, S., Liu, X., Liu, J., Cheng, D., 2018. Responses of vegetation
productivity to multi-scale drought in Loess Plateau, China. Catena 163, 165–171.
https://doi.org/10.1016/j.catena.2017.12.016

Zhao, J., Huang, S., Huang, Q., Wang, H., Leng, G., Fang, W., 2020. Time-lagged
response of vegetation dynamics to climatic and teleconnection factors. Catena 189,
104474. https://doi.org/10.1016/j.catena.2020.104474

Zhao, Song, Yang, Li, Zhang, Feng, 2019. Monitoring of Nitrogen and Grain Protein
Content in Winter Wheat Based on Sentinel-2A Data. Remote Sens. 11, 1724.
https://doi.org/10.3390/rs11141724

Zhao, Z., Liu, J., Peng, J., Li, S., Wang, Y., 2015. Nonlinear features and complexity
patterns of vegetation dynamics in the transition zone of North China. Ecol. Indic.
49, 237–246. https://doi.org/10.1016/j.ecolind.2014.08.038

Zhao, Z.Q., Li, S.C., Gao, J.B., Wang, Y.L., 2011. Identifying spatial patterns and
dynamics of climate change using recurrence quantification analysis: A case study
of qinghaitibet plateau. Int. J. Bifurc. Chaos 21, 1127–1139.
https://doi.org/10.1142/S0218127411028933

Zhong, L., Ma, Y., Salama, M.S., Su, Z., 2010. Assessment of vegetation dynamics and
their response to variations in precipitation and temperature in the Tibetan Plateau.
Clim. Change 103, 519–535. https://doi.org/10.1007/s10584-009-9787-8

Zhou, H., Fu, L., Sharma, R.P., Lei, Y., Guo, J., 2021. A Hybrid Approach of Combining
Random Forest with Texture Analysis and VDVI for Desert Vegetation Mapping
Based on UAV RGB Data. Remote Sens. 13, 1891.
https://doi.org/10.3390/rs13101891

Zhou, X., Kono, Y., Win, A., Matsui, T., Tanaka, T.S.T., 2021. Predicting within-field
variability in grain yield and protein content of winter wheat using UAV-based
multispectral imagery and machine learning approaches. Plant Prod. Sci. 24, 137–
151. https://doi.org/10.1080/1343943X.2020.1819165

Zhu, H., Liu, H., Xu, Y., Guijun, Y., 2018. UAV-based hyperspectral analysis and
spectral indices constructing for quantitatively monitoring leaf nitrogen content of
winter wheat. Appl. Opt. 57, 7722. https://doi.org/10.1364/AO.57.007722

142
Zolotova, N. V, Ponyavin, D.I., 2007. Synchronization in Sunspot Indices in the Two
Hemispheres. Sol. Phys. 243, 193–203. https://doi.org/10.1007/s11207-007-0405-5

Zurlini, G., Marwan, N., Semeraro, T., Jones, K.B., Aretano, R., Pasimeni, M.R., Valente,
D., Mulder, C., Petrosillo, I., 2018. Investigating landscape phase transitions in
Mediterranean rangelands by recurrence analysis. Landsc. Ecol. 33, 1617–1631.
https://doi.org/10.1007/s10980-018-0693-1

Zurlini, G., Petrosillo, I., Aretano, R., Castorini, I., D’Arpa, S., De Marco, A., Pasimeni,
M.R., Semeraro, T., Zaccarelli, N., 2014. Key fundamental aspects for mapping and
assessing ecosystem services: Predictability of ecosystem service providers at scales
from local to global. Ann. di Bot. 4, 53–63. https://doi.org/10.4462/annbotrm-11754

143
9 ANNEXES

144
 Original series and Anomalies series time series

0.70

0.60
Vegetation Index
0.50
NDVI

0.40

0.30

0.20
NDVI_ZGU
0.10 NDVI_ZSO
1 24 47 70 93 116 139 162 185 208 231 254 277 300 323 346 369 392 415 438 461 484 507 530 553 576 599 622 645 668 691 714 737 760
Time units

0.70
MSAVI_ZGU
0.60
MSAVI_ZSO
Vegetation Index

0.50
MSAVI

0.40

0.30

0.20

0.10
1 24 47 70 93 116 139 162 185 208 231 254 277 300 323 346 369 392 415 438 461 484 507 530 553 576 599 622 645 668 691 714 737 760
Time units

TMIN_ZGU
28.00
TMIN_ZSO
23.00
Temperature (ºC)

18.00

13.00
TMIN

8.00

3.00

-2.00

-7.00
1 24 47 70 93 116 139 162 185 208 231 254 277 300 323 346 369 392 415 438 461 484 507 530 553 576 599 622 645 668 691 714 737 760
Time units

250.00

200.00
Precipitation (mm)

PCP_ZGU
150.00 PCP_ZSO
PCP

100.00

50.00

0.00
1 24 47 70 93 116 139 162 185 208 231 254 277 300 323 346 369 392 415 438 461 484 507 530 553 576 599 622 645 668 691 714 737 760
Time units

145
 0.20
A_NDVI_GU

Vegetation Index Anomalies

0.15
A_NDVI_SO
0.10

0.05
A_NDVI

0.00

-0.05

-0.10

-0.15

-0.20
1 24 47 70 93 116 139 162 185 208 231 254 277 300 323 346 369 392 415 438 461 484 507 530 553 576 599 622 645 668 691 714 737 760
Time units

0.15
A_MSAVI_GU
Vegetation Index Anomalies

0.10 A_MSAVI_SO

0.05
A_MSAVI

0.00

-0.05

-0.10

-0.15
1 24 47 70 93 116 139 162 185 208 231 254 277 300 323 346 369 392 415 438 461 484 507 530 553 576 599 622 645 668 691 714 737 760
Time units

10.00
A_TMIN_GU
8.00
Temperature Anomalies (ºC)

A_TMIN_SO
6.00
4.00
2.00
A_TMIN

0.00
-2.00
-4.00
-6.00
-8.00
-10.00
1 24 47 70 93 116 139 162 185 208 231 254 277 300 323 346 369 392 415 438 461 484 507 530 553 576 599 622 645 668 691 714 737 760
Time units

250.00
Precipitation Anomalies (mm)

200.00

A_PCP_GU
150.00
A_PCP_SO
A_PCP

100.00

50.00

0.00

-50.00
1 24 47 70 93 116 139 162 185 208 231 254 277 300 323 346 369 392 415 438 461 484 507 530 553 576 599 622 645 668 691 714 737 760
Time units

146
 Threshold values of each sequential segmentation between VIs and wheat traits.

Threshold values between VIs and yield in GS32.

Threshold NDVI MSAVI NDRE BRI1
T_1 0.00 0.00 0.00 0.00
T_2 0.04 0.04 0.04 0.25
T_3 0.07 0.07 0.07 0.29
T_4 0.09 0.09 0.10 0.31
T_5 0.11 0.11 0.12 0.34
T_6 0.13 0.13 0.14 0.36
T_7 0.15 0.15 0.16 0.38
T_8 0.16 0.16 0.18 0.41
T_9 0.18 0.17 0.19 0.43
T_10 0.19 0.19 0.21 0.45
T_11 0.20 0.20 0.22 0.48
T_12 0.22 0.21 0.24 0.50
T_13 0.23 0.23 0.26 0.52
T_14 0.25 0.24 0.27 0.53
T_15 0.26 0.26 0.29 0.55
T_16 0.28 0.26 0.31 0.55
T_17 0.30 0.27 0.33 0.56
T_18 0.33 0.28 0.35 0.56
T_19 0.35 0.29 0.37 0.59

Threshold values between VIs and protein content in GS65.

Threshold NDVI MSAVI NDRE BRI1
T_1 0.00 0.00 0.00 0.10
T_2 0.03 0.03 0.02 0.42
T_3 0.05 0.06 0.04 0.46
T_4 0.07 0.08 0.06 0.49
T_5 0.09 0.09 0.07 0.52
T_6 0.10 0.11 0.08 0.54
T_7 0.12 0.12 0.09 0.56
T_8 0.13 0.13 0.10 0.57
T_9 0.14 0.15 0.11 0.59
T_10 0.15 0.16 0.12 0.60
T_11 0.16 0.17 0.13 0.62
T_12 0.17 0.18 0.13 0.64
T_13 0.18 0.19 0.14 0.65
T_14 0.19 0.20 0.15 0.67
T_15 0.20 0.21 0.16 0.68
T_16 0.21 0.22 0.17 0.71
T_17 0.23 0.24 0.19 0.73
T_18 0.25 0.25 0.20 0.75
T_19 0.27 0.27 0.22 0.79

147
Threshold values between VIs and nitrogen output in GS39.
Threshold NDVI MSAVI NDRE BRI1
T_1 0,00 0,00 0,00 0,41
T_2 0,08 0,09 0,06 0,75
T_3 0,13 0,15 0,10 0,81
T_4 0,16 0,19 0,13 0,86
T_5 0,20 0,22 0,16 0,91
T_6 0,22 0,25 0,18 0,95
T_7 0,25 0,28 0,20 0,98
T_8 0,27 0,31 0,22 1,00
T_9 0,29 0,33 0,24 1,04
T_10 0,32 0,35 0,26 1,07
T_11 0,33 0,38 0,27 1,10
T_12 0,36 0,40 0,29 1,12
T_13 0,38 0,42 0,31 1,15
T_14 0,40 0,44 0,32 1,18
T_15 0,42 0,47 0,34 1,21
T_16 0,44 0,49 0,36 1,24
T_17 0,46 0,51 0,38 1,28
T_18 0,49 0,54 0,40 1,33
T_19 0,51 0,57 0,43 1,40

148
 Recurrence plot (RP) and Cross Recurrence plots (CRP) parameters and Recurrence
Quantification Analysis (RQA) using general z-score NDVI series.

RR DET LAM
Zone RPs and CRPs m τ 𝒓 LT ENTR TT
(%) (%) (%)
NDVI 2 8 8.67 5.00 59.19 2.81 1.23 72.32 3.01
TEMP 2 11 8.76 4.99 40.26 2.31 0.72 52.89 2.49
ZAV PCP 2 3 1.35 4.97 14.65 2.10 0.32 33.72 2.35
NDVI-TEMP 2 11 9.96 5.00 48.99 2.44 0.88 58.30 2.60
NDVI-PCP 2 8 6.34 4.99 25.96 2.26 0.64 30.46 2.45
NDVI 2 10 7.80 4.99 63.92 2.99 1.29 77.22 3.26
TEMP 2 11 9.00 4.99 42.02 2.35 0.76 54.35 2.60
ZMA PCP 10 9 13.25 5.00 6.57 2.03 0.14 20.72 2.11
NDVI-TEMP 2 11 10.53 4.99 52.42 2.51 0.96 59.65 2.63
NDVI-PCP 10 10 26.19 4.99 24.26 2.12 0.39 27.36 2.15
ZAV: Tornadizos de Ávila and ZMA: Soto del Real, NDVI: Normalized Difference Vegetation Index, m:
Embedding dimension, τ: Delay, 𝑟: threshold, RR: Recurrence rate, DET: Determinism, LT: Average
length of diagonal structures, ENTR: Shannon Entropy, LAM: Laminarity, TT: Trapping time.

149
 Median values of vegetation indices (Index Value) and average temperature each 8 Days
(TEMP) from 2002 to 2018, in each zone: ZGU = Guadalix de la Sierra, ZSO =Soto del Real. A) NDVI
and B) MSAVI.

A)

B)

150
 Recurrence plots of NDVI and MSAVI for each zone (ZGU and ZSO). Vegetation indices
series are normalized by the z-score method.

NDVI MSAVI
ZSO
ZGU

Time units are represented as the X and Y axis. Each time-unit is a period of 8-days, coincident with 8-
days compose MODIS images during the study period (2002-2018). Embedding dimension m =1, delay τ
= 0 and recurrence rate RR = 1.5%.

151