Chin. Geogra. Sci. 2023 Vol. 33 No. 4 pp.

751−763 Springer Science Press

https://doi.org/10.1007/s11769-023-1370-0 www.springerlink.com/content/1002-0063

Estimating Monthly Surface Air Temperature Using MODIS LST Data

and an Artificial Neural Network in the Loess Plateau, China

HE Tian1, 2, LIU Fuyuan2, WANG Ao2, FEI Zhanbo3

(1. School of Water Conservancy Engineering, Zhengzhou University, Zhengzhou 450001, China; 2. School of Geoscience and Techno-
logy, Zhengzhou University, Zhengzhou 450001, China; 3. Suntront Technology Company Limited, Zhengzhou 450052, China)

Abstract: Air temperature (Ta) datasets with high spatial and temporal resolutions are needed in a wide range of applications, such as
hydrology, ecology, agriculture, and climate change studies. Nonetheless, the density of weather station networks is insufficient, espe-
cially in sparsely populated regions, greatly limiting the accuracy of estimates of spatially distributed Ta. Due to their continuous spatial
coverage, remotely sensed land surface temperature (LST) data provide the possibility of exploring spatial estimates of Ta. However, be-
cause of the complex interaction of land and climate, retrieval of Ta from the LST is still far from straightforward. The estimation accur-
acy varies greatly depending on the model, particularly for maximum Ta. This study estimated monthly average daily minimum temper-
ature (Tmin), average daily maximum temperature (Tmax) and average daily mean temperature (Tmean) over the Loess Plateau in
China based on Moderate Resolution Imaging Spectroradiometer (MODIS) LST data (MYD11A2) and some auxiliary data using an ar-
tificial neural network (ANN) model. The data from 2003 to 2010 were used to train the ANN models, while 2011 to 2012 weather sta-
tion temperatures were used to test the trained model. The results showed that the nighttime LST and mean LST provide good estimates
of Tmin and Tmean, with root mean square errors (RMSEs) of 1.04°C and 1.01°C, respectively. Moreover, the best RMSE of Tmax es-
timation was 1.27°C. Compared with the other two published Ta gridded datasets, the produced 1 km × 1 km dataset accurately cap-
tured both the temporal and spatial patterns of Ta. The RMSE of Tmin estimation was more sensitive to elevation, while that of Tmax
was more sensitive to month. Except for land cover type as the input variable, which reduced the RMSE by approximately 0.01°C, the
other vegetation-related variables did not improve the performance of the model. The results of this study indicated that ANN, a type of
machine learning method, is effective for long-term and large-scale Ta estimation.
Keywords: air temperature; land surface temperature (LST); artificial neural network (ANN); remote sensing; climate change; Loess
Plateau, China

Citation: HE Tian, LIU Fuyuan, WANG Ao, FEI Zhanbo, 2023. Estimating Monthly Surface Air Temperature Using MODIS LST
Data and an Artificial Neural Network in the Loess Plateau, China. Chinese Geographical Science, 33(4): 751−763. https://doi.org/
10.1007/s11769-023-1370-0

1 Introduction ous scientific research fields (Benali et al., 2012). Ta

data are generally recorded as point samples. In global
Near-surface air temperature (Ta) is measured at 1.5–2.0 m change studies, many models, such as gross primary
above the land surface at weather stations (Jin and production models (Zhang Y et al., 2016), hydrological
Dickinson, 2010), and it is an important variable in vari- models (Zhang et al., 2009), and land surface models

Received date: 2022-12-15; accepted date: 2023-03-10

Foundation item: Under the auspices of the ‘Beautiful China’ Ecological Civilization Construction Science and Technology Project (No.
XDA23100203), National Natural Science Foundation of China (No. 42071289), Henan Postdoctoral Foundation (No. 20180087)
Corresponding author: FEI Zhanbo. E-mail: feizhanb@gmail.com
© Science Press, Northeast Institute of Geography and Agroecology, CAS and Springer-Verlag GmbH Germany, part of Springer
Nature 2023
752 Chinese Geographical Science 2023 Vol. 33 No. 4

(Cooke et al., 2012), require spatially distributed sur- used for regional or local land-climate interaction re-
face Ta (Didari and Zand-Parsa, 2018). As the density search (Hooker et al., 2018). Furthermore, reanalysis
of weather station networks is insufficient, especially in data have some other drawbacks, including the absence
sparsely populated and underdeveloped regions, accur- of clouds and surface moisture, changes in atmospheric
ate estimation of spatially distributed Ta is still a chal- composition, and biases resulting from surface heat in-
lenge (Vancutsem et al., 2010). Recently, numerous come and expense balances (Vose et al., 2004; Li et al.,
studies have adopted space-for-time substation methods 2019). Satellite remote sensing can provide global cov-
to evaluate the impact of land cover change on global erage of LST and vegetation indices. These parameters
climate change (Peng et al., 2014; Zhao and Jackson, are measured continuously in both space and time. Be-
2014; Li et al., 2015; Alkama and Cescatti, 2016; cause LST has a strong correlation with Ta, satellite re-
Forzieri et al., 2017). Due to the lack of reliable spa- mote sensing provides a possible method for improving
tially distributed near-surface Ta datasets, the conclu- the spatiotemporal accuracy of Ta estimation.
sions were mainly based on readily available satellite- Four primary methods are used for estimating Ta
based land surface temperature (LST) data. However, based on LST. 1) Statistical methods based on regres-
near-surface Ta and LST have different physical mean- sion techniques. Statistical regression is the most com-
ings, although they are closely coupled (Jin and Dickin- monly used classic method. As the relationship between
son, 2010). Therefore, the results based on LST are LST and Ta is restricted by underlying conditions and
probably different from the results inferred from Ta. weather conditions, a simple linear relationship does not
Some studies (Lu et al., 2018) found that the spatial pat- fully express the physical connection between the two
terns and warming trends of LST and Ta showed signi- (Recondo et al., 2013; Ho et al., 2014; Zhu et al., 2017).
ficant distinctions, implying that they can not be substi- Therefore, the evaluation accuracy of the traditional re-
tuted for one another when assessing climate change. gression method, whether a simple linear regression or a
Therefore, producing Ta grid data with spatial resolu- multiple linear regression, needs to be improved. Re-
tions fine enough to capture regional or local variation is cently, some new regression techniques, such as geo-
crucial (Oyler et al., 2016). graphically weighted regression (GWR) (Chen et al.,
To date, various methods have been used to produce 2015; Hooker et al., 2018) and regression kriging (Hengl
spatially continuous Ta data. Geospatial interpolation is et al., 2012; Kilibarda et al., 2014), have been proven to
regarded as a classic and simple method. However, be better estimation methods than traditional regression
many studies have shown that traditional interpolation methods. 2) Temperature vegetation index (TVX) meth-
methods such as inverse distance weighting, kriging and ods. These methods use the normalized difference ve-
spline methods do not provide accurate estimates in getation index (NDVI) as an important input variable
complex terrain because the interpolation accuracy de- and have focused on estimating daily or instantaneous
pends largely on the density of weather station net- Ta (Zhu et al., 2013). However, the significant negative
works and the spatiotemporal variability of the paramet- correlation between LST and the NDVI is not always
ers (Benali et al., 2012; Didari and Zand-Parsa, 2018). applicable and is affected by seasonality and ecosystem
Therefore, some new interpolation methods have been type (Vancutsem et al., 2010). 3) Energy-balance meth-
developed in recent years. The WorldClim gridded cli- ods. These methods have a solid physical basis and
mate dataset, which is produced by the thin-plate sm- equate the sum of net incoming radiation with the sum
oothing spline algorithm implemented in the ANUS- of surface sensible and latent heat fluxes. The disadvant-
PLIN package (McVicar et al., 2007), has high corres- age of these methods is that they depend on information
pondence with observations (Hijmans et al., 2005; Fick that is not easily available through remote sensing data
and Hijmans, 2017). Meteorological reanalysis data (e.g., solar radiation, wind speed, and surface rough-
(e.g., ERA-Interim or NCEP data) (Kalnay et al., 1996; ness) (Vancutsem et al. 2010). 4) Machine learning
Dee et al., 2011) provide another source of gridded Ta methods. Recently, machine learning methods such as
data with a high temporal resolution, but these spatial artificial neural networks (ANNs) (Jang et al., 2004;
resolutions are fairly coarse at approximately 30 to 50 km, Şahin, 2012), model trees (Emamifar et al., 2013),
which may mask local variation, and thus they are rarely Bayesian methods (Lu et al., 2018; Zhang and Du,
 HE Tian et al. Estimating Monthly Surface Air Temperature Using MODIS LST Data and an Artificial Neural Network... 753

2019) and random forests (Meyer et al., 2016; Noi et al., in the southeast and by grasslands and unused land in
2017; Otgonbayar et al., 2019) have been applied in Ta the northeast (Fig. 1) (Feng et al., 2012; Li et al., 2012;
estimation, greatly expanding the traditional estimation Xiao, 2014; Qu et al., 2020).
methods. The multilayer perceptron ANN is the classic
and simple neural network, which is the foundation of 2.2 Data
other deep learning algorithm. However, a limited num- Daily Ta data from 2003–2012 were collected from 35
ber of studies estimating Ta based on LST with an ANN weather stations whose locations did not change during
have been reported in the scientific literature. Jang et al. the study period. These data were downloaded from the
(2004) and Şahin (2012) used Advanced Very High China Meteorological Data Service Center (CMDC, ht-
Resolution Radiometer (AVHRR) data to train a neural tp://data.cma.cn). Ta was measured 2 m above the gro-
network model for Ta estimation in Canada and Turkey. und, and the data were all subjected to quality control.
However, these studies focused only on monthly aver- Two types of data were used in this study. First, mon-
age daily mean temperature (Tmean), and many studies thly Tmax, Tmin and Tmean were calculated from daily
have shown that the estimation of monthly average daily values for training and validation. The other dataset was
maximum temperature (Tmax) is a challenge compared a 0.5° × 0.5° gridded surface monthly temperature in
with that of Tmean (Vancutsem et al., 2010; Benali et China (SURF_CLI_CHN_TEM_MON_GRID_0.5), wh-
al., 2012; Zhu et al., 2013). Are ANN models still use- ich contains monthly Tmax, Tmin, and Tmean. The
ful when estimating Tmax? In addition, can the per- dataset comprised temperature data collected by high-
formance of the models be improved if vegetation-re- density Chinese ground stations (approximately 2400)
lated variables are added as input parameters? as part of the National Meteorological Observatory and
To answer these questions, we used an ANN model to obtained using the ANUSPLIN package (Table 1). Be-
estimate Tmin, Tmax and Tmean based on MODIS LST cause the data cover from 1961 only through 2012, we
data and some auxiliary data on the Loess Plateau in aggregated the 0.5° × 0.5° gridded monthly Ta from
China. Data from the period of 2003–2010 were used to 2003 to 2012 to the seasonal scale for comparison with
train the ANN, and weather station observations from the spatial pattern of our simulated gridded Ta.
2011–2012 were used to test the accuracy of the trained MODIS LST data were used in this study. These data
network. Finally, we used the trained neural network to were downloaded from the Land Processes Distributed
produce a gridded Ta dataset at 1 km spatial resolution Active Archive Center (LP DAAC, https://lpdaac.usgs.
for the 2003–2012 period. This study, especially the gov/). The MODIS instrument operates on both the
dataset and methods, will be valuable for various applic- Terra and Aqua spacecraft. The Terra overpass times are
ations involving agriculture, hydrology and climate approximately 10:30 am (local time) and 10:30 pm. The
change in areas with low weather station density. Aqua overpass times are approximately 1:30 pm and
1:30 am. Although some studies indicated that overpass
2 Materials and Methods
100°E 105°E 110°E 115°E
2.1 Study area
40°N
This study covers the entire Loess Plateau region, which 40°N
Legend Inner Mongolia
includes seven provinces in the northwestern China. The Weather station
Provincial boundary
Loess Plateau covers 6.4 × 105 km2 of land in the
middle reaches of the Yellow River, and the region is Ningxia Shanxi
Shaanxi
characterized by a temperate continental monsoon cli- Qinghai
Gansu
mate with an annual mean temperature ranging from 35°N
35°N
Land use Henan
4°C to 14°C and mean annual precipitation varying Cropland Forest
from 200 mm in the northwest to approximately 800 mm Grassland Unused land 0 150 300 km
Water Built-up land
in the southeast. The northwest-southeast precipitation 100°E 105°E 110°E 115°E
gradient strongly influences vegetation distribution pat- Fig. 1 Location of Loess Plateau region, China and distribution
terns. The plateau is dominated by croplands and forests of the weather stations
754 Chinese Geographical Science 2023 Vol. 33 No. 4

time did not have much of an effect on the accuracy of data from China’s National Land Use and Cover Ch-
Ta estimation (Mostovoy et al., 2006; Didari and Zand- ange (CNLUCC, https://www.resdc.cn/DOI/DOI.aspx?
Parsa, 2018), Aqua data (MYD11A2) are used more of- DOIID=54) dataset were used in this study. This data-
ten than Terra data (MOD11A2) because daytime LST set was primarily produced by interpreting Landsat
reflects the temperature maxima and nighttime LST re- Thematic Mapper (TM) and Enhanced Thematic Map-
flects the temperature minima. In MYD11A2 Collection- per Plus (ETM +) images. The classification scheme in-
6 production, the LST errors are within ± 1 K in most cludes six first-class land use types: cropland, wood-
cases (Wan et al., 2004; Wan, 2014). We only used good- land, grassland, water body, built-up land, and unused
quality pixels (QA flags = 00) for this study to minim- land. Field survey-based validation indicated that the
ize the effects of cloud contamination. Because the classification accuracy for each land use type was over
MYD11A2 data start in July 2002, we used 2003–2012 90% (Liu et al., 2014).
data in this study. The data from 2003–2010 were used
to train the ANN, and the data from 2011–2012 were 2.3 Methods
used to test the accuracy of the trained network. The In this study, we used the multilayer perceptron ANN,
MYD11A2 are 8-d composite images with a total of 46 which consists of one input layer, one hidden layer and
images per year. These 8-d data were aggregated to a one output layer, to estimate Ta. Backpropagation, a su-
monthly mean between 2003 and 2012 for each of the pervised learning algorithm, was used to train multilay-
weather stations. er perception. The goal of any supervised learning al-
The NASA Shuttle Radar Topographic Mission (SR- gorithm is to find a function that best maps a set of in-
TM, https://srtm.csi.cgiar.org/) provides elevation data puts to their correct output. The neural network works
for the study area. In this study, the SRTM data were like a black box without detailed information about the
resampled from 90 m to 1 km using ArcGIS to make system. In the beginning, the network initializes weights
them consistent with the other data in terms of resolu- with some random values for input variables, such as
tion when producing the map of Ta. LST, elevation, and month (Fig. 2). The neuron then
Vegetation products were used as input parameters in learns from the training samples. The initial network
the neural network model, including the NDVI and land computes an output that is likely to differ from the tar-
cover type. Because vegetation affects temperature get outcome. A loss function is used for measuring the
through a series of biophysical effects, such as albedo discrepancy (error) between the target output and the
and evapotranspiration, it has often been used in regres- computed output. The neural network needs to be trained
sion prediction models for Ta. Because we focused on to adjust the weights, which in turn reduces the error.
monthly temperature prediction, monthly normalized The backpropagation algorithm looks for the minimum
difference vegetation index (NDVI) products at a 1-km of the error function in the weight space using a tech-
spatial resolution (MOD13A3) were used in the study nique called gradient descent. The weights that minim-
(https://lpdaac.usgs.gov/). In addition, the 2010 land use ize the error function are then considered as a solution

Table 1 The description of the data used in the study

Dataset Type of data Variable Resolution Use
CMDC weather station data CSV Air temperature – Training and validation
CMDC gridded data ASCII Air temperature 0.5° Spatial comparison

Hooker’ s gridded data NetCDF Air temperature 0.05° Spatial comparison

MODIS MYD11A2 HDF Land surface temperature 1000 m Model input

MODIS MOD13A3 HDF NDVI 1000 m Model input

CNLUCC data RASTER Land cover type 1000 m Model input

SRTM RASTER Elevation 90 m Model input
Notes: CMDC: China Meteorological Data Service Center; CNLUCC: China’s National Land Use and Cover Change; SRTM: Shuttle Radar Topographic Mission;
NDVI: normalized difference vegetation index. – means nodata
 HE Tian et al. Estimating Monthly Surface Air Temperature Using MODIS LST Data and an Artificial Neural Network... 755

to the learning problem (Rumelhart et al., 1986). study compared with the other alternatives. No theoret-
In this study, MODIS monthly LSTs, elevation, ical method can yet determine the number of neurons in
months, latitude, longitude and vegetation-related vari- the hidden layer of the ANN. The number of neurons in
ables (NDVI and LC type) were employed in the input the hidden layer was determined based on the result
layer to obtain monthly Ta in the output layer. As eleva- (high R2 and small RMSE value) of network training.
tion has a strong correlation with temperature and digit- Calculations were performed separately for between 1
al elevation models (DEMs) are available worldwide, and 40 neurons to determine the optimum number of
elevation is commonly used as a predictor variable neurons in the hidden layer according to the minimum
(Oyler et al., 2016). Some studies (Jang et al. 2004; RMSE values. At the same time, because the initial
Benali et al. 2012) have shown that Julian day is anoth- weights of the neural network are random, the results of
er significant parameter, and it is used as a proxy for the each simulation may be different, even if the number of
variability of the surface energy balance. Because we est- neurons in the hidden layer is constant. Therefore, the
imated monthly Ta, the months from January to Decem- average temperature obtained after 30 replicates for
ber were used as input variables. As the geographical each number of neurons (from 1 to 40) in the hidden
location was linked with the solar zenith angle, longit- layer was regarded as the final result of the neural net-
ude and latitude were used as two other auxiliary vari- work.
ables. Vegetation indices such as NDVI have been wi-
dely applied in previous statistical models, but these 3 Results and Analyses
variables are still less used in neural network models
(Jang et al. 2004; Şahin 2012). We used the NDVI and 3.1 LST-observed Ta comparison
LC types as input variables in this study to see whether The comparison between the monthly MODIS LST and
these variables could improve the performance of the the corresponding weather station-based Ta is shown in
ANN. Fig. 3 for 2011–2012. The monthly nighttime LST and
In this study, we used the Deep Learning Toolbox™ mean LST were good proxies for Tmin and Tmean, re-
provided by MATLAB for estimation. The backpro- spectively (R2 = 0.97), while the correlation between
pagation network in MATLAB has many settings. The monthly daytime LST and Tmax was slightly lower
transfer function ‘tansig’ was used for the hidden layer, (R2 = 0.91).
and ‘purelin’ was used for the output layer. The Leven- Generally, the daytime LST was higher than the
berg-Marquardt backpropagation algorithm ‘trainlm’ was Tmax at most sites. For example, Tmax was always be-
used as a network training function. The mean squared low 33°C, whereas daytime LST could exceed 43°C
error (MSE) was used to evaluate network performance. during these two years (Fig. 3a). Overall, the daytime
We found that these settings performed the best in the LST was (6.87 ± 3.90)°C higher than the corresponding
Tmax on average. Large differences, reaching up to
Input layer Hidden layer Output layer 10°C in spring (March to May), were observed.
The nighttime LST was lower than Tmin, but the dif-
Monthly LST
ference was smaller than the difference between the
Elevation
daytime LST and Tmax (Fig. 3b). Overall, the night-
time LST was (1.42 ± 1.64)°C lower than the corres-
Months Monthly air ponding Tmin, and the largest difference (only 5.82°C)
temperature

Latitude
was observed in winter (December to February). Al-
.
.
though the correlations between daytime/nighttime LST
Longitude .
and Ta were strong, the RMSEs differed. The RMSE
Vegetation variable
(e.g., NDVI)
(2.18°C) between nighttime LST and Tmin was smaller
than that between daytime LST and Tmax (RMSE of
Fig. 2 The artificial neural network (ANN) model used in the 7.78°C). As shown inFig. 3d, the proposed ANNs in the
study. LST, land surface temperature; NDVI, normalized differ- study effectively converged the data pairs, especially for
ence vegetation index Tmax estimation.
756 Chinese Geographical Science 2023 Vol. 33 No. 4

45 45
40 a 40 d
35 35

MODIS daytime LST / °C

30 30

Estimated Tmax / °C
25 25
20 20
15 15
10 10
5 5
y = 1.089x + 5.435 y = 0.952x + 1.031
0 0
R2 = 0.91 R2 = 0.99
−5 RMSE = 7.78 −5 RMSE =1.27
−10 −10
−15 −15
−15 −10 −5 0 5 10 15 20 25 30 35 40 45 −15 −10 −5 0 5 10 15 20 25 30 35 40 45
Measured Tmax / °C Measured Tmax / °C
30 30
25 b 25 e
20 20
MODIS nighttime LST / °C

15 15

Estimated Tmin / °C
10 10
5 5
0 0
−5 −5
−10 y = 1.015x − 1.472 −10
y = 0.989x + 0.092
−15 R2 = 0.97 −15
R2 = 0.99
−20 RMSE = 2.18 −20 RMSE = 1.04
−25 −25
−30 −30
−30 −25 −20 −15 −10 5 0 5 10 15 20 25 30 −30 −25 −20 −15 −10 5 0 5 10 15 20 25 30
Measured Tmin / °C Measured Tmin / °C
35 35
30 c 30 f
25 25
MODIS daily LST / °C

20
Estimated Tmean / °C

20
15 15
10 10
5 5
0 0
−5 y = 1.026x + 3.085 −5 y = 0.977x + 0.308
R2 = 0.97 R2 = 0.99
−10 RMSE = 3.88 −10 RMSE = 1.01
−15 −15
−20 −20
−20 −15 −10 −5 0 5 10 15 20 25 30 35 −20 −15 −10 −5 0 5 10 15 20 25 30 35
Measured Tmean / °C Measured Tmean / °C

Fig. 3 Scatter plots of measured Ta versus MODIS (Moderate Resolution Imaging Spectroradiometer) LST (a–c) and versus Ta estim-
ated by the ANN model (d–f). ANN, artificial neural network. N = 840, P < 0.05

3.2 Overall performance formance was found. The average RMSE of model 6
In this study, a backpropagation neural network was was 0.01°C larger than that of model 8 for Tmin and
used to predict monthly Ta. Because we did not know Tmax. That is, when using five input variables, we ob-
which variables would more strongly affect the perform- tained adequate performance, while the data processing
ance of the ANN model, several different combinations workload was minimal.
of variables (models 1–9) were tested (Table 2). An interesting finding was that elevation, which is
Overall, good agreement was found between the used by most researchers, was closely associated with
measured and estimated Ta by ANN models in this Tmin only and was not as useful when estimating Tmax.
study (Fig. 3 and Table 2). The ANN models presented Table 2 showed that the RMSE value of Tmin in model
the best performance for Tmin and Tmax estimation 1, which included elevation as a parameter, was smaller
when using six input variables (model 8) (Table 2). The than the RMSE of model 2 (1.19°C vs 1.44°C), and for
average RMSE was 1.04°C and 1.27°C for Tmin and Tmax, the opposite was true (2.89°C vs 2.28°C). The
Tmax, respectively. When we used only five variables result of Tmin being more sensitive to elevation is in ac-
as input neurons (model 6), very slight decrease in per- cordance with the findings of Rangwala and Miller
 HE Tian et al. Estimating Monthly Surface Air Temperature Using MODIS LST Data and an Artificial Neural Network... 757

(2012) and Beniston et al. (1997), which are detailed in results showed that for Tmin approximately 86% of the
the discussion section. stations had an RMSE lower than 1.1°C (Fig. 4a). The
Table 3 lists the R2 and RMSE for Ta estimation of 35 stations’ RMSEs ranged between 0.75°C and 1.27°C.
the monthly Tmin, Tmean and Tmax. The model per- The top 3 stations with the highest RMSE values were
formed well during the summer, with an R2 > 0.90 and a mostly located in the southeast of the Loess Plateau
small RMSE (usually below 1.00°C). The model had the (Fig. 5a). Similar to the pattern observed for Tmin, ap-
best performance in July when estimating Tmin and proximately 80% of the stations had a Tmean RMSE
Tmax. In contrast, the performance of the model was lower than 1.1°C, and the stations’ RMSEs ranged
poor during the cold winter. When evaluating Tmax and between 0.77°C and 1.39°C (Fig. 4b). However, in con-
Tmean, the maximum RMSE appeared in January. For trast to the spatial distribution observed for Tmin, the
Tmin, the largest RMSE was also found during the cold top five stations with the highest RMSEs were located in
month of November. This means that the ANN model the northern Loess Plateau (Fig. 5b). Regarding Tmax,
performs well under warm conditions and poorly under only approximately 30% of the stations had an RMSE
cold conditions. lower than 1.1°C, but almost 71% of the stations’
RMSEs were below 1.3°C. The RMSE of Tmax ranged
3.3 Station-level performance between 0.87°C and 1.79°C (Fig. 4c). Precipitation in-
Performance was analyzed for each weather station. The fluences the accuracy of the Tmax and Tmean estima-

Table 2 The RMSE of monthly average Ta estimation compared with weather station observations in Loess Plateau, China / °C
Model Input variables Tmin Tmean Tmax
1 LST, Elevation 1.19 1.69 2.89
2 LST, Month 1.44 1.51 2.28
3 LST, Elevation, Month 1.15 1.11 1.51
4 LST, Elevation, Latitude, Longitude 1.08 1.66 2.75
5 LST, Month, Latitude, Longitude 1.07 1.02 1.32
6 LST, Elevation, Month, Latitude, Longitude 1.05 1.01 1.28
7 LST, Elevation, Month, Latitude, Longitude, NDVI 1.06 1.03 1.28
8 LST, Elevation, Month, Latitude, Longitude, LC 1.04 1.03 1.27
9 LST, Elevation, Month, Latitude, Longitude, NDVI, LC 1.04 1.03 1.28
Notes: Each value is the average of 30 simulations. LST: land surface temperature; LC: land cover type.

Table 3 The coefficient of determination and RMSE values of monthly Ta estimation in Loess Plateau, China / °C
Tmin Tmean Tmax
Month
2 2 2
R RMSE R RMSE R RMSE
January 0.94 1.11 0.89 1.62 0.72 2.34
February 0.91 1.12 0.85 1.25 0.78 1.51
March 0.89 1.49 0.90 1.50 0.87 1.72
April 0.95 0.69 0.96 0.67 0.91 1.26
May 0.95 0.70 0.94 0.75 0.91 1.02
June 0.97 0.59 0.97 0.59 0.96 0.74
July 0.97 0.58 0.98 0.53 0.97 0.58
August 0.93 0.92 0.93 0.76 0.92 0.93
September 0.94 0.97 0.96 0.97 0.92 1.35
October 0.95 0.65 0.97 0.49 0.94 0.76
November 0.87 1.95 0.87 1.43 0.89 0.96
December 0.94 0.92 0.94 0.72 0.91 1.05
758 Chinese Geographical Science 2023 Vol. 33 No. 4

tion (Fig. 5). Four of the five stations with the highest spatial distribution of the simulated temperature, two
RMSEs were located in the northern Loess Plateau other published gridded Ta datasets covering the Loess
along the precipitation gradient, which was similar to Plateau are also illustrated in Fig. 6. The first 0.05° ×
the pattern observed for Tmean. The site with the 0.05° resolution Ta dataset comes from Hooker et al.
highest RMSE was located in arid areas with precipita- (2018), and the second coarse 0.5° × 0.5° gridded Ta
tion below 200 mm (Figs. 5b–5c). However, in contrast dataset was downloaded from the CMDC. Because
to the patterns observed for Tmean and Tmax, the three Hooker provided only Tmean data, we took the season-
lowest performances in terms of Tmin were found in al Tmean maps as examples. Spatially, the simulated
high-precipitation areas. We found that two sites are Tmean showed similar regional variations compared
located near rivers (Fenhe River and Zhuozhang River), with those from the two other datasets (Fig. 6). Some
with a distance of less than 2 km. According to previ- local differences also occurred among the datasets. For
ous studies (Vancutsem et al., 2010), the presence of example, compared with the Tmean from Hooker’s
water decreases the accuracy of estimating Tmin. A dis- dataset, that from the ANN showed a cold pattern on the
tance of 1 to 2 km is not sufficient to avoid the effect of western edge of the Loess Plateau in DJF (Figs. 6a, 6e)
water (Vancutsem et al., 2010). The other site is located and a warm pattern in the southwest of the Loess Plat-
in Huashan Mountain, and its elevation is approxim- eau in MAM (Figs. 6b, 6f). Because of the spatial resol-
ately 2000 m. More complex mountain topography may ution, the ANN-simulated data and Hooker’s data led to
be a factor affecting the accuracy of Ta estimates (Van- a more heterogeneous Ta pattern and showed more spa-
cutsem et al., 2010; Fick and Hijmans, 2017). tial details than the CMDC data.
Averaged over the Loess Plateau, the simulated
3.4 Mapping of Ta and comparison with other Tmean was –5.34°C, 9.54°C, 20.24°C, and 8.17°C for
gridded datasets the four seasons (DJF, MAM, JJA, and SON), respect-
According to the trained neural network, we produced ively, which is very close to the Tmean obtained from
seasonal Ta at spatial resolution of 1 km across the the other two datasets (Table 4). Overall, Hooker’s
Loess Plateau from 2003 to 2012. To understand the Tmean is slightly higher than those of ANN and CMDC.

16 12 10
a. Tmin b. Tmean c. Tmax
14
10 8
12
Number of stations

Number of stations

Number of stations

8
10 6
8 6
6 4
4
4
2 2
2
0 0 0
0.6 0.8 1.0 1.2 1.4 0.6 0.8 1.0 1.2 1.4 1.6 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
RMSE / °C RMSE / °C RMSE / °C

Fig. 4 Distribution of model performance (RMSE) for 35 weather stations in Loess Plateau, China for 2011–2012

880
a. Tmin b. Tmean c. Tmax
Precipitation / mm

RMSE RMSE RMSE

0.76−0.86 0.77−0.82 0.87−1.02
0.87−1.00 0.83−0.88 1.03−1.11
1.01−1.07 0.89−1.03 1.12−1.28
1.08−1.14 0 150 300 km 1.04−1.16 0 150 300 km 1.29−1.52 0 150 300 km
1.15−1.27 1.17−1.39 1.53−1.79 140

Fig. 5 RMSE of Tmin, Tmean and Tmax estimation for all of the weather stations in the Loess Plateau, China for 2011–2012
 HE Tian et al. Estimating Monthly Surface Air Temperature Using MODIS LST Data and an Artificial Neural Network... 759

ANN ANN ANN ANN

a b c d
Hooker Hooker Hooker Hooker

e f g h
CMDC CMDC CMDC CMDC

i j k l
DJF MAM JJA SON

−15 Temperature / °C 28

Fig. 6 Spatial pattern of seasonal Tmean from the ANN model and two other datasets averaged during 2003–2012. Hooker and CM-
DC are two datasets from Hooker et al. (2018) and CMDC, respectively. DJF, MAM, JJA, and SON means winter, spring, summer and
autumn, respectively

Table 4 Seasonal Tmean from different sources for 2003−2012 have been applied to estimate daily and monthly Ta val-
/ °C ues in recent years. Although the backpropagation ANN
Dataset DJF MAM JJA SON is a classic, simple model that was developed by Rumel-
This study –5.34 9.54 20.24 8.17 hart et al. (1986) almost 30 yr ago, we found its per-
Hooker et al. (2018) –5.12 9.71 20.95 8.84 formance in Ta estimation to be surprisingly high. The
CMDC –5.68 9.43 20.28 7.50 error was relatively high when retrieving Tmax from
Notes: Seasons are defined as winter (December to February, DJF), spring daytime LSTs compared with retrieving Tmin from
(March to May, MAM), summer (June to August, JJA) and autumn
nighttime LSTs. As Table 5 shows, the RMSEs of Tmax
(September to November, SON). CMDC: China Meteorological Data Service
estimated by various statistical methods were between
Center
1.27°C and 2.15°C. GWR has recently been widely used
to estimate monthly Ta. Chen et al. (2015) and Hooker
4 Discussion et al. (2018) used this method in China and globally, re-
spectively. The RMSEs of Tmin, Tmean and Tmax ob-
4.1 Comparison with other studies tained in Chen et al.’s (2015) study were 1.24°C, 1.08°C
In this study, we show that by using MODIS LST data and 1.29°C, respectively, which were very close to the
and some simple auxiliary data, one can reliably estim- RMSEs from Hooker’s global estimation. The GWR
ate Ta across large regions with a classic backpropaga- method produced RMSEs that were 25%–26% smaller
tion ANN model. The ANN model produced excellent than those from other regression methods, such as mul-
results, with RMSE values for simulated Tmax, Tmin tiple linear regression and regression kriging (Chen et
and Tmean of 1.27°C, 1.04°C and 1.01°C in the Loess al., 2015). In this study, the accuracy of the estimated
Plateau, respectively. As mentioned in the introduction, Tmax and Tmean was very close to that obtained with
regression methods and the TVX method are the two ap- estimation by GWR methods, but the RMSE of estim-
proaches used most widely in previous studies. In addi- ated Tmin was only 1.04°C, which was lower than that
tion, machine learning algorithm models, such as ran- obtained with GWR. Unlike advanced machine learn-
dom forests, model trees and support vector machines, ing algorithms such as random forest (Li and Zha, 2019)
760 Chinese Geographical Science 2023 Vol. 33 No. 4

Table 5 RMSE reported in some monthly Ta estimation studies

Studies Method RMSE (monthly) / °C Study area
This study ANN Tmin (1.04), Tmean (1.01), Tmax (1.27) Loess Plateau, China
Şahin (2012) ANN Tmean (1.25) Turkey

Chen et al. (2015) GWR Tmin (1.24), Tmean (1.08), Tmax (1.29) China

Hooker et al. (2018) GWR & CSWR Tmean (1.14–1.58) Global

Lu et al. (2018) Bayesian Tmin (1.97), Tmax (2.15) Qinghai, China

Li et al. (2018) Random forest Tmin (1.57–1.99) China

Qin et al. (2023) Stacking ensemble learning Tmean (1.61) Tibetan Plateau, China

Notes: GWR, geographically weighted regression; CSWR, climate space weighted regression

and Bayesian (Lu et al., 2018) methods, ANN methods um temperatures are often more sensitive to elevation
are seldom used to predict Ta from LST. Şahin (Şahin, than are daily maximum temperatures in mountain areas
2012) forecasted Tmean based on AVHRR data and an (Beniston et al., 1997; Rangwala and Miller, 2012). This
ANN in Turkey. In addition to Tmean, our study estim- may be because the solar radiation during the day is
ated Tmin and Tmax, which extended Şahin’s research strong, the LST is much higher than the Ta, and the ra-
and demonstrated the ability of ANNs to predict Ta diation warming effect is greater than the cooling effect
based on LST. caused by the elevation change. However, no radiant
heating occurs at night, and the effect of elevation
4.2 Relationship between Ta and the chosen auxili- changes on temperature changes is more obvious than
ary data during the day.
Elevation is widely applied to develop high-spatiotem- As previously mentioned, vegetation plays an import-
poral resolution Ta datasets because elevation has a ant biophysical role in surface-atmosphere interactions
strong correlation with Ta. Generally, higher elevations (Mildrexler et al., 2011). The Ta and LST relationships
are associated with a lower Ta. For example, ANUS- vary among LC types and vegetation structures and
PLIN, an interpolation method accounting for elevation, compositions is widely acknowledged (Mildrexler et al.,
is considered better for converting point data into grid 2011). To improve the performance of the model, we
data than traditional kriging interpolation. Previous added the NDVI and LC type to the input parameter
studies inferring Ta based on LST, whether using statist- combination (Table 2, models 7, 8 and 9). The results
showed that the inclusion of the NDVI did not improve
ical regression or machine learning, also included eleva-
the model estimation performance ability compared
tion as an important variable in the model (Hengl et al.,
with that obtained with the five basic input parameters
2012; Chen et al., 2015; Parmentier et al., 2015; Zhang
(Table 2, model 6). Some previous studies have dis-
H B et al., 2016; Lian et al., 2017; Li et al., 2018). In
cussed this situation. The studies by Lin et al. (2012)
this study, we also used elevation as one of the input
and Benali et al. (2012) indicated that adding the NDVI
variables of the ANN. Elevation is of course a very im-
or enhanced vegetation index (EVI) did not have im-
portant parameter, but we found that the effect of eleva- portant effects for estimating Ta when LST and eleva-
tion was not significant when estimating Tmax com- tion were included as predictors. In addition, although
pared with Tmin. The RMSE of the Tmax predicted with the NDVI shows a strong negative relationship with
LST and elevation as inputs was 2.89°C, which was LST, it has less temporal variation than LST (Li et al.,
0.61°C higher than the RMSE obtained when using LST 2015), which may result in a smaller role of the NDVI
and month as input parameters. Conversely, when eval- in some models. However, we still found that adding LC
uating Tmin, the effect of elevation was significant. types may slightly improve simulation accuracy. The
When using LST and elevation as input parameters, the RMSE of model 8 was 0.01°C lower than that of model
RMSE was 1.19°C, which was 0.29°C lower than the 6 in the estimation of Tmin and Tmax when accounting
RMSE obtained when using LST and month as parameters for LC types. This result is consistent with that from an
(Table 2). Several studies have shown that daily minim- observational study by Mildrexler et al. (2011), who
 HE Tian et al. Estimating Monthly Surface Air Temperature Using MODIS LST Data and an Artificial Neural Network... 761

found that the strength of the LST/Ta relationship was age Tmin, Tmax and Tmean can be accurately estim-
dependent on land cover. Although the decrease in ated using ANNs and remote sensing data. MODIS
RMSE was not very significant, most of the LC types of LST, weather station-measured Ta and some auxiliary
the weather stations located in our study area are grass- data from 2003 to 2010 were used to train the models,
land, cropland and urban, thus excluding forest, bare while 2011 to 2012 weather station observations were
land and water. If the study area covers more LC types, used to test the trained model. The Tmin and Tmean es-
the results may be better. timated by the model agreed well with the station meas-
urements, with R2 values of both 0.99 and RMSE values
4.3 Limitations and uncertainty of 1.04°C and 1.01°C, respectively. For Tmax estima-
In this study, we found that the model had lower per- tion, the R2 and RMSE values were 0.99 and 1.27°C, re-
formance in winter. The better performance in summer spectively. The study found that incorporating NDVI
compared with that in winter is consistent with the res- and land cover type did not have a significant impact on
ults reported by Li et al. (2018). One possible reason for Ta estimation. Therefore, the researchers concluded that
the seasonal variability could be that LST quality is bet- utilizing five easily accessible variables, namely LST,
ter in summer than in winter. In addition, the model elevation, month, latitude, and longitude, in the model's
showed lowest performance for Tmax and Tmean estim- input layer would be sufficient for reliable results. In ad-
ation in regions with rainfall below 200 mm because the dition, large RMSEs occur in arid area and complex ter-
accuracy of the LST algorithm was lower in arid and rain (e.g., mountain or riverside), may due to the satel-
semiarid regions than in other regions due to large vari- lite data quality. The generated gridded Ta dataset ex-
ations in surface emissivity (Wan et al., 2004; Li and hibited similar spatial variations to two other gridded Ta
Zha, 2019). Similar to some other studies, our study datasets, indicating that ANN is an effective method for
showed that adding vegetation-related variables such as estimating Ta from MODIS LST.
NDVI and LC type does not have a significant impact
on monthly Ta estimation. Therefore, five basic and eas- References
ily accessible variables (LST, elevation, month, latitude
and longitude) are sufficient in the input layer for ob- Alkama R, Cescatti A, 2016. Biophysical climate impacts of re-
taining reliable results. cent changes in global forest cover. Science, 351(6273):
Although the ANN model with the backpropagation 600–604. doi: 10.1126/science.aac8083
learning algorithm performed well in this study, some Benali A, Carvalho A C, Nunes J P et al., 2012. Estimating air
limitations to this method exist. First, the gradient des- surface temperature in Portugal using MODIS LST data. Re-
mote Sensing of Environment, 124: 108–121. doi: 10.1016/j.rse.
cent algorithm is generally slow, and the backpropaga-
2012.04.024
tion algorithm may not always find a solution, meaning Beniston M, Diaz H F, Bradley R S, 1997. Climatic change at
convergence to local minimum points. Second, the mod- high elevation sites: an overview. Climatic Change, 36(3–4):
el is sensitive to the number of neurons in the hidden 233–251. doi: 10.1023/A:1005380714349
layer. Too few neurons can lead to underfitting, and too Chen F R, Liu Y, Liu Q et al., 2015. A statistical method based
many neurons can contribute to overfitting. In addition, on remote sensing for the estimation of air temperature in
China. International Journal of Climatology, 35(8):
the appropriate number of neurons in the hidden layer
2131–2143. doi: 10.1002/joc.4113
can not easily be determined, which makes the network Cooke W H, Mostovoy G V, Anantharaj V G et al., 2012. Wild-
slightly less efficient. Nevertheless, ANNs are still a fire potential mapping over the state of Mississippi: a land sur-
promising tool and should be applied in more scientific face modeling approach. GIScience & Remote Sensing, 49(4):
research areas. 492–509. doi: 10.2747/1548-1603.49.4.492
Dee D P, Uppala S M, Simmons A J et al., 2011. The ERA-Inter-
im reanalysis: configuration and performance of the data as-
5 Conclusions
similation system. Quarterly Journal of the Royal Meteorolo-
gical Society, 137(656): 553–597. doi: 10.1002/qj.828
Accurate estimates of spatially continuous near-surface Didari S, Zand-Parsa S, 2018. Enhancing estimation accuracy of
air temperatures are helpful for a wide range of earth daily maximum, minimum, and mean air temperature using
science studies. This study showed that monthly aver- spatio-temporal ground-based and remote-sensing data in
762 Chinese Geographical Science 2023 Vol. 33 No. 4

southern Iran. International Journal of Remote Sensing, ing areas in the conterminous United States. Remote Sensing of
39(19): 6316–6339. doi: 10.1080/01431161.2018.1460500 Environment, 215: 74–84. doi: 10.1016/j.rse.2018.05.034
Emamifar S, Rahimikhoob A, Noroozi A A, 2013. Daily mean air Li Y, Zhao M S, Motesharrei S et al., 2015. Local cooling and
temperature estimation from MODIS land surface temperature warming effects of forests based on satellite observations.
products based on M5 model tree. International Journal of Cli- Nature Communications, 6: 6603. doi: 10.1038/ncomms7603
matology, 33(15): 3174–3181. doi: 10.1002/joc.3655 Li Y Z, Wang L, Zhou H X et al., 2019. Urbanization effects on
Feng X M, Sun G, Fu B J et al., 2012. Regional effects of vegeta- changes in the observed air temperatures during 1977–2014 in
tion restoration on water yield across the Loess Plateau, China. China. International Journal of Climatology, 39(1): 251–265.
Hydrology and Earth System Sciences, 16(8): 2617–2628. doi: doi: 10.1002/joc.5802
10.5194/hess-16-2617-2012 Li Z, Zheng F L, Liu W Z, 2012. Spatiotemporal characteristics
Fick S E, Hijmans R J, 2017. WorldClim 2: new 1-km spatial res- of reference evapotranspiration during 1961−2009 and its pro-
olution climate surfaces for global land areas. International jected changes during 2011−2099 on the Loess Plateau of
Journal of Climatology, 37(12): 4302–4315. doi: 10.1002/joc. China. Agricultural and Forest Meteorology, 154−155:
5086 147–155. doi: 10.1016/j.agrformet.2011.10.019
Forzieri G, Alkama R, Miralles D G et al., 2017. Satellites reveal Lian X, Zeng Z Z, Yao Y T et al., 2017. Spatiotemporal vari-
contrasting responses of regional climate to the widespread ations in the difference between satellite-observed daily max-
greening of Earth. Science, 356(6343): 1180–1184. doi: 10. imum land surface temperature and station-based daily maxim-
1126/science.aal1727 um near-surface air temperature. Journal of Geophysical Re-
Hengl T, Heuvelink G B M, Perčec Tadić M et al., 2012. Spatio- search:Atmospheres, 122(4): 2254–2268. doi: 10.1002/
temporal prediction of daily temperatures using time-series of 2016JD025366
MODIS LST images. Theoretical and Applied Climatology, Lin S P, Moore N J, Messina J P et al., 2012. Evaluation of estim-
107(1–2): 265–277. doi: 10.1007/s00704-011-0464-2 ating daily maximum and minimum air temperature with
Hijmans R J, Cameron S E, Parra J L et al., 2005. Very high res- MODIS data in east Africa. International Journal of Applied
olution interpolated climate surfaces for global land areas. In- Earth Observation and Geoinformation, 18: 128–140. doi: 10.
ternational Journal of Climatology, 25(15): 1965–1978. doi: 1016/j.jag.2012.01.004
10.1002/joc.1276 Liu J Y, Kuang W H, Zhang Z X et al., 2014. Spatiotemporal
Ho H C, Knudby A, Sirovyak P et al., 2014. Mapping maximum characteristics, patterns, and causes of land-use changes in
urban air temperature on hot summer days. Remote Sensing of China since the late 1980s. Journal of Geographical Sciences,
Environment, 154: 38–45. doi: 10.1016/j.rse.2014.08.012 24(2): 195–210. doi: 10.1007/s11442-014-1082-6
Hooker J, Duveiller G, Cescatti A, 2018. A global dataset of air Lu N, Liang S L, Huang G H et al., 2018. Hierarchical Bayesian
temperature derived from satellite remote sensing and weather space-time estimation of monthly maximum and minimum sur-
stations. Scientific Data, 5: 180246. doi: 10.1038/sdata.2018. face air temperature. Remote Sensing of Environment, 211:
246 48–58. doi: 10.1016/j.rse.2018.04.006
Jang J D, Viau A A, Anctil F, 2004. Neural network estimation of McVicar T R, Van Niel T G, Li L T et al., 2007. Spatially distrib-
air temperatures from AVHRR data. International Journal of uting monthly reference evapotranspiration and pan evapora-
Remote Sensing, 25(21): 4541–4554. doi: 10.1080/ tion considering topographic influences. Journal of Hydrology,
01431160310001657533 338(3–4): 196–220. doi: 10.1016/j.jhydrol.2007.02.018
Jin M L, Dickinson R E, 2010. Land surface skin temperature cli- Meyer H, Katurji M, Appelhans T et al., 2016. Mapping daily air
matology: benefitting from the strengths of satellite observa- temperature for antarctica based on MODIS LST. Remote
tions. Environmental Research Letters, 5(4): 044004. doi: 10. Sensing, 8(9): 732. doi: 10.3390/rs8090732
1088/1748-9326/5/4/044004 Mildrexler D J, Zhao M S, Running S W, 2011. A global compar-
Kalnay E, Kanamitsu M, Kistler R et al., 1996. The NCEP/NCAR ison between station air temperatures and MODIS land surface
40-year reanalysis project. Bulletin of the American Meteorolo- temperatures reveals the cooling role of forests. Journal of
gical Society, 77(3): 437–471. doi: 10.1175/1520-0477(1996) Geophysical Research:Biogeosciences, 116(G3): G03025. doi:
077<0437:TNYRP>2.0.CO;2 10.1029/2010JG001486
Kilibarda M, Hengl T, Heuvelink G B M et al., 2014. Spatio-tem- Mostovoy G V, King R L, Reddy K R et al., 2006. Statistical es-
poral interpolation of daily temperatures for global land areas timation of daily maximum and minimum air temperatures
at 1 km resolution. Journal of Geophysical Research:Atmo- from MODIS LST data over the state of Mississippi. GIS-
spheres, 119(5): 2294–2313. doi: 10.1002/2013JD020803 cience & Remote Sensing, 43(1): 78–110. doi: 10.2747/1548-
Li L, Zha Y, 2019. Estimating monthly average temperature by 1603.43.1.78
remote sensing in China. Advances in Space Research, 63(8): Noi P, Degener J, Kappas M, 2017. Comparison of multiple lin-
2345–2357. doi: 10.1016/j.asr.2018.12.039 ear regression, cubist regression, and random forest algorithms
Li X M, Zhou Y Y, Asrar G R et al., 2018. Developing a 1 km to estimate daily air surface temperature from dynamic com-
resolution daily air temperature dataset for urban and surround- binations of MODIS LST data. Remote Sensing, 9(5): 398. doi:
 HE Tian et al. Estimating Monthly Surface Air Temperature Using MODIS LST Data and an Artificial Neural Network... 763

10.3390/rs9050398 ure in different ecosystems over Africa. Remote Sensing of En-

Otgonbayar M, Atzberger C, Mattiuzzi M et al., 2019. Estima- vironment, 114(2): 449–465. doi: 10.1016/j.rse.2009.10.002
tion of climatologies of average monthly air temperature over Vose R S, Karl T R, Easterling D R et al., 2004. Impact of land-
mongolia using MODIS land surface temperature (LST) time use change on climate. Nature, 427(6971): 213–214. doi: 10.
series and machine learning techniques. Remote Sensing, 1038/427213b
11(21): 2588. doi: 10.3390/rs11212588 Wan Z, Zhang Y, Zhang Q et al., 2004. Quality assessment and
Oyler J W, Dobrowski S Z, Holden Z A et al., 2016. Remotely validation of the MODIS global land surface temperature. In-
sensed land skin temperature as a spatial predictor of air tem- ternational Journal of Remote Sensing, 25(1): 261–274. doi: 10.
perature across the conterminous United States. Journal of Ap- 1080/0143116031000116417
plied Meteorology and Climatology, 55(7): 1441–1457. doi: 10. Wan Z M, 2014. New refinements and validation of the collection-
1175/JAMC-D-15-0276.1 6 MODIS land-surface temperature/emissivity product. Re-
Parmentier B, McGill B J, Wilson A M et al., 2015. Using multi- mote Sensing of Environment, 140: 36–45. doi: 10.1016/j.rse.
timescale methods and satellite-derived land surface temperat- 2013.08.027
ure for the interpolation of daily maximum air temperature in Xiao J F, 2014. Satellite evidence for significant biophysical con-
Oregon. International Journal of Climatology, 35(13): sequences of the ‘Grain for Green’ Program on the Loess Plat-
3862–3878. doi: 10.1002/joc.4251 eau in China. Journal of Geophysical Research:
Peng S S, Piao S, Zeng Z Z et al., 2014. Afforestation in China Biogeosciences, 119(12): 2261–2275. doi: 10.1002/
cools local land surface temperature. Proceedings of the Na-
2014JG002820
tional Academy of Sciences of the United States of America,
Zhang H B, Zhang F, Ye M et al., 2016. Estimating daily air tem-
111(8): 2915–2919. doi: 10.1073/pnas.1315126111
peratures over the Tibetan Plateau by dynamically integrating
Qin J, Pan W H, He M et al., 2023. A long-term 1 km monthly
MODIS LST data. Journal of Geophysical Research: Atmo-
near-surface air temperature dataset over the Tibetan glaciers
spheres, 121(19): 11425–11441. doi: 10.1002/2016JD025154
by fusion of station and satellite observations. Earth System
Zhang K, Kimball J S, Mu Q Z et al., 2009. Satellite based ana-
Science Data, 15(1): 331–344. doi: 10.5194/essd-15-331-2023
lysis of northern ET trends and associated changes in the re-
QU Lulu, HUANG Yunxin, YANG Lingfan et al., 2020. Vegeta-
gional water balance from 1983 to 2005. Journal of Hydrology,
tion restoration in response to climatic and anthropogenic
379(1–2): 92–110. doi: 10.1016/j.jhydrol.2009.09.047
changes in the Loess Plateau, China. Chinese Geographical
Zhang Y, Xiao X M, Jin C et al., 2016. Consistency between sun-
Science, 30(1): 89–100. doi: 10.1007/s11769-020-1093-4
induced chlorophyll fluorescence and gross primary produc-
Rangwala I, Miller J R, 2012. Climate change in mountains: a re-
view of elevation-dependent warming and its possible causes. tion of vegetation in North America. Remote Sensing of Envir-
Climatic Change, 114(3): 527–547. doi: 10.1007/s10584-012- onment, 183: 154–169. doi: 10.1016/j.rse.2016.05.015
0419-3 Zhang Z W, Du Q Y, 2019. A Bayesian kriging regression meth-
Recondo C, Peón J J, Zapico E et al., 2013. Empirical models for od to estimate air temperature using remote sensing data. Re-
estimating daily surface water vapour pressure, air temperature, mote Sensing, 11(7): 767. doi: 10.3390/rs11070767
and humidity using MODIS and spatiotemporal variables. Ap- Zhao K G, Jackson R B, 2014. Biophysical forcings of land-use
plications to peninsular Spain. International Journal of Re- changes from potential forestry activities in North America.
mote Sensing, 34(22): 8051–8080. doi: 10.1080/01431161. Ecological Monographs, 84(2): 329–353. doi: 10.1890/12-
2013.828185 1705.1
Rumelhart D E, Hinton G E, Williams R J, 1986. Learning rep- Zhu W B, Lü A, Jia S F, 2013. Estimation of daily maximum and
resentations by back-propagating errors. Nature, 323(6088): minimum air temperature using MODIS land surface temperat-
533–536. doi: 10.1038/323533a0 ure products. Remote Sensing of Environment, 130: 62–73. doi:
Şahin M, 2012. Modelling of air temperature using remote sens- 10.1016/j.rse.2012.10.034
ing and artificial neural network in Turkey. Advances in Space Zhu W B, Lü A, Jia S F et al., 2017. Retrievals of all-weather
Research, 50(7): 973–985. doi: 10.1016/j.asr.2012.06.021 daytime air temperature from MODIS products. Remote Sens-
Vancutsem C, Ceccato P, Dinku T et al., 2010. Evaluation of ing of Environment, 189: 152–163. doi: 10.1016/j.rse.2016.11.
MODIS land surface temperature data to estimate air temperat- 011