ECOLOGICAL ENGINEERING

& ENVIRONMENTAL TECHNOLOGY

Ecological Engineering & Environmental Technology, 2025, 26(2), 407–419 Received: 2024.11.19
https://doi.org/10.12912/27197050/199448 Accepted: 2024.12.21
ISSN 2719–7050, License CC-BY 4.0 Published: 2025.01.01

Optimizing the performance of rainflow-runoff models

using multi-source satellite data-application to the region
of Beni-Bahdel (Algeria)
Hiba Chahrouri1 , Abdelkader Bemmoussat1* , Khaled Korichi1
1
Civil and Environmental Engineering Laboratory, University of Djillali Liabes, BP 89. DZ-22000, Sidi Bel
Abbés, Algeria
* Corresponding author’s e-mail: bemmoussaek@yahoo.fr

ABSTRACT
The integration of satellite data in rainfall-runoff simulation is of paramount importance in regions where data are
limited and not easily available. The data used in this study, concerning the daily precipitation and daily runoff
during the period extending from 2007 to 2015, were measured and recorded in situ. This study aims primarily to
optimize the performance of the modeling carried out, using data from various satellite sources in order to com-
plete the missing data, and consequently to predict the liquid flow in the Beni Bahdel watershed, in the northwest-
ern region of Algeria, by applying the long short-term memory (LSTM) learning model. It is important to know
that the optimization of rainfall-runoff modeling is based on the use of satellite data relating to evapotranspira-
tion, mean temperatures, minimum and maximum temperatures, net radiation, wind speed, and relative humidity.
These data come from the NOAA CPC, ERA, ERA5_AG, ERA5-Land, GLDAS, CFSR and MERRA2 satellites.
In addition, two statistical indicators were calculated to perform this optimization that is based on the LSTM ap-
proach that integrates remote sensing data, the coefficient of determination (R²), and the Nash-Sutcliffe efficiency
(NSE). Thus, a performance difference of about 0.30 was observed, for the NSE and R² coefficients, between the
CFSR temperature data (NSE of 0.61and R² of 0.61) and the maximum and minimum temperature data from the
ERA5-LAND and ERA5 satellite sources (0.92 for NSE and 0.93 for R²), for the validation period. This significant
difference suggests that the use of the minimum and maximum temperature data from the ERA5-LAND source
allows achieving a rainfall-runoff modeling wih optimal performance. Indeed, the findings showed that quite high
performances were achieved for the calibration period (0.93 for NSE and 0.95 for R²) and for the validation period
(0.92 for NSE and 0.93 for R²).

Keywords: optimization, modeling, satellite data, long short-term memory, Beni Bahdel, Algeria

INTRODUCTION be highly important, particularly in rainfall-runoff

modeling. The application of remote sensing al-
It is widely acknowledged that remote sens- lows converting multiple data collected by remote
ing is an innovative information technology that sensors into information that is exploited to better
plays a vital role in different fields (Huang et al., understand and predict hydrological flows. In par-
2024). Remote sensing techniques, as indicated in ticular, precipitation data are viewed as the essen-
a number of studies, are often used in agriculture tial tool that generally affects several disciplines,
(Khanal et al., 2020), environmental monitoring such as water resources management, planning,
(Vasiliou & Economides, 2006), urban planning modeling and planning (Duan et al., 2019; Shekar
(Netzband et al., 2007), mining exploration (Ci- et al., 2024; Ur Rahman et al., 2020).
ampalini et al., 2013), as well as salinity-affected A large number of studies, carried out
soil management (Dehni & Lounis, 2012). Like- throughout the world, have used remote sensing
wise, in hydrology, remote sensing turns out to for the purpose of analyzing various hydrological

407
Ecological Engineering & Environmental Technology 2025, 26(2), 407–419

processes. For example, (Yu et al., 2024) con- system (GLDAS). Moreover, other studies were
ducted a study, in northern China, to develop a conducted by a number of researchers who used
runoff simulation model using multi-source satel- precipitation, wind, relative humidity, solar radia-
lite data. For this, these researchers relied on the tion data in the field of hydrogeology (Derdour et
precipitation data from the precipitation estima- al., 2022; Inan et al., 2024). Nevertheless, to the
tion from remotely sensed information using arti- best of our knowledge, the comparison of multi-
ficial neural networks – climate data record (PER- source satellite data in rainfall-runoff modeling
SIANN-CDR), surface temperature data from has not been sufficiently investigated in Algeria,
MOD11C1, a product developed by moderate which constitutes a good research opportunity for
resolution imaging spectroradiometer (MODIS), those working in this field.
the evapotranspiration data, and soil moisture It should be pointed out that, in the study area
data from Global land evaporation amsterdam under consideration, we only have observed precipi-
model (GLEAM). Similarly, in South Korea (Cho tation and flow data. It turns out that using neural
& Kim, 2022) conducted a study in which they ex- networks for rainfall-runoff modeling represents a
ploited the local data assimilation and prediction solution to this problem. In this case, only precipi-
system (LDAPS) meteorological data, while in- tation data are considered as input variable. How-
tegrating precipitation, temperature data, surface ever, it should be emphasized that the present work
pressure data, and some surface wind elements, aims primarily to optimize the performance of flow
with specific humidity and radiation data. Like- simulation, while integrating input parameters from
wise, in Egypt (Abdelmoneim et al., 2020) used different satellite data sources, in addition to pre-
data from the tropical rainfall measuring mission cipitation data measured and recorded in situ. These
(TRMM 3B42V7) and climate hazards group parameters were selected based on their relationship
infra-red precipitation with station (CHIRPS) with the estimation of evapotranspiration.
data satellite precipitation products as inputs to
establish an appropriate hydrological model. As
for Darand et al. (2017), they carried out a study MATERIALS AND METHODS
in Iran with a view to assessing the tropical rain-
fall measuring mission (TRMM) multi-satellite Presentation of the study region
precipitation analysis (TMPA) precipitation prod-
ucts. In addition, a particularly interesting study In this article, the Beni Bahdel watershed was
was conducted in Ethiopia by (Gebremicael et al., chosen and studied (Figure 1). It is located in the
2019) with the purpose of analyzing satellite pre- northwest of Algeria, between longitudes 1°8’ W
cipitation data from satellite-based rainfall prod- and 1°47’ W, and latitudes 34°33’ N and 34°41’
ucts CHIRPS, RFEv2 (Rainfall Estimates version N. The Beni Bahdel watershed covers an area of
2), TRMM and PERSIANN. 989 Km2 and crosses five rural municipalities, i.e.
It is important to point out that several Alge- Ain Ghoraba, Beni Bahdel, Beni Snous, Azail
rian researchers have used satellite-based rainfall and Sebdou, where the main activity is agricul-
products due to the scarcity of data in some ar- ture. In the year 1930, the Beni Bahdel concrete
eas. In this regard, (Lazri et al., 2020) performed dam, with a capacity of 63 million m3, was built
a study that focused on improving precipitation at the outlet of the watershed where the two main
estimation in northern Algeria. The remote sens- Wadis, namely Wadi Khemis and Wadi Sebdou,
ing data used include multispectral images from meet. It should be noted that the climate in this re-
the spinning enhanced visible and infrared im- gion is semi-arid. This area is characterized by a
ager (SEVIRI) radiometer as well as satellite pre- mountainous relief, with a predominance of karst
cipitation data obtained from the climate predic- geological formations. This favors the infiltration
tion center morphing technique (CMORPH) and of precipitation (Bouguerra & Mansour, 2023).
TRMM 3B42 products. In particular (Bemmous-
sat et al., 2021) benefited from the contribution Data used
of satellite precipitation for the rainfall-runoff
Observed data
hydrological modeling in the region of Ham-
mam Boughrara (Algeria), using precipitation The observed data, which were used in this
data from TRMM 3B43V7 and evapotranspi- study, include information on daily precipita-
ration data from Global land data assimilation tion (mm), and daily runoff (mm), from January

408
 Ecological Engineering & Environmental Technology 2025, 26(2), 407–419

Figure 1. The study area map

1, 2007 to December 31, 2015. These data were • Climate Forecast System Reanalysis (CFSR)
provided by the National Agency for Hydrau- produced by National Centers for Environ-
lic Resources (ANRH – Agence Nationale des mental Prediction (NCEP); this system is
Ressources Hydrauliques). In addition, the managed by NOAA.
arithmetic means from the three rainfall sta-
tions, i.e. Khemis, Beni Bahdel and Sebdou Hydrological modeling
stations, as well as the sum of the runoffs from
the hydrometric stations, i.e. Wadi Khemis and Long-short term memory
Wadi Sebdou, were used for the rainfall-runoff Long short-term memory (LSTM) is a type of
modeling (Figure 1). recurrent neural network (RNN) that was initially
designed by (Hochreiter & Schmidhuber, 1997)
Satellite data and was then developed by many other research-
It is worth emphasizing that several satellite data ers over time. It has been revealed that these neu-
were used as input parameters in the model, along ral networks can be used in modeling the depen-
with precipitation data, for the purpose of optimizing dencies of long time series data with high accura-
the performance of the flow simulation. These data cy (Mao et al., 2021). In addition, LSTM can also
include information on evapotranspiration, mean, be adopted for rainfall-runoff prediction because
maximum, and minimum temperature, radiation, this approach represents a major advance to vari-
wind speed, and humidity. All these data were down- ous fundamental problems in the hydrological do-
loaded for free from the Climate Engine, POWER, main (Hashemi et al., 2022). In this regard, many
and Giovanni websites. These sites provide weather studies have been carried out by (Cho & Kim,
and climate data from the following sources: 2022; Lees et al., 2021; Sabzipour et al., 2023)
• National Oceanic and Atmospheric Administra- with the aim of comparing the performances of
tion Climate Prediction Center (NOAA CPC) LSTM networks and classical hydrological mod-
• European Reanalysis Analysis 5th Genera- els in the simulation of flows. The results proved
tion (ERA5) that neural predictions are more efficient.
• European Reanalysis Analysis 5th Generation These positive results regarding neural net-
Agroclimatology (ERA5_AG) works motivated our choice to adopt the deep
• European Reanalysis Analysis 5th Generation learning (DP) technique in our investigation.
land (ERA5 Land) With regard to modeling, it was deemed interest-
• Global Land Data Assimilation System ing to use LSTMs as a black box, while respect-
(GLDAS). ing the following tasks:
• Modern-Era Retrospective analysis for Re- Provide the model with the necessary input
search and Applications, Version 2 (MERRA2) data such as the observed precipitation and other

409
Ecological Engineering & Environmental Technology 2025, 26(2), 407–419

complementary parameters from satellite sources. Modeling steps

In addition, observed flow data were introduced
Figure 2 presents the modeling steps that are
to allow the model to train and evaluate the per-
given below:
formance of the output flows for future periods.
• Introduce the in situ measured data, namely
Segment the study period into two phases.
precipitation and runoff. Precipitation should be
The first phase (2007–2013) for training and the
used as input data for the LSTM model, while
second (2014–2015) for testing.Assign random
values ​​to the training parameters of the LSTM runoff is used for training the model and evalu-
model. This operation is then repeated, while ating the performance of simulated runoffs.
performing several tests, in order to improve • Choose the input data complementary to pre-
the performance. cipitation. These data are divided into five
groups, and each group can contain one or
The LSTM optimization function more parameters. The data for these param-
eters come from several satellite sources.
The performance of flow simulation mainly
• Introduce the initial hyperparameters.
depends on the input data used. However, it can
• Run the LSTM model.
also be influenced by the optimization function
• Evaluate the performance using the NSE and
of the LSTM model. This optimization function
R² indices.
contains several hyperparameters whose values
• Tune the hyperparameters by testing several
can impact the modeling performance. Before
possible combinations, and then stop when the
running the LSTM algorithm, it is essential to
best values of R² and NSE are obtained.
tune the values of all hyperparameters by testing
• The hyperparameters that give the optimal
several combinations because this allows obtain-
ing the best possible performance in a reasonable performance are then used as initial hyperpa-
time. This process is called hyperparameter opti- rameters for the next modeling group.
mization (Wu et al., 2019). • Compare and interpret NSE and R² values.
The hyperparameter optimization process can
be approached using two different methods, i.e. a Evaluation of hydrological modeling
manual method or an automatic method. The manu- performance
al approach has been adopted in this work. It is use-
ful to know that this method is difficult to handle Nash-Sutcliffe efficiency
due to the number of parameters to be optimized and The Nash-Sutcliffe efficiency (NSE) is a sys-
also to their value interval (Wu et al., 2019). Thus, tem of measurement that is widely used by re-
to make the tuning of hyperparameters easier, it is searchers to assess the accuracy of models in flow
highly recommended to set typical starting values simulation. It was first introduced by Nash and
(Table 1) in order to train and evaluate the model. Sutcliffe in 1970 (Nash & Sutcliffe, 1970). It var-
A large number of research studies have ies from minus infinity to 1. The value 1 indicates
shown that the number of epochs is one of the a perfect match between the observed and simu-
most important parameters. According to (Greff lated values. The NSE can be calculated using the
et al., 2016), this parameter accounts for up to following formula:
two-thirds of the overall performance. For this
reason, special attention has been paid to this pa- (1)
rameter during the optimization step.
where: is the daily observed runoff,
is the daily simulated runoff, the av-
Table 1. Typical parameters of LSTM (Ravindra, 2018) erage of observed runoff, and N the num-
Parameters Typical starting values ber of observations.
Max epochs 250
Furthermore, the NSE performance ranges
Num hidden units 200
are presented as follows (Ritter & Muñoz-Carpe-
Gradient threshold 1
na,2013): for NSE ≥ 0.90, the model is excellent;
Initial learn rate 0.005
for 0.80 < NSE < 0.90, the model is very satisfac-
Learn rate drop factor 0.2
tory; for 0.60 < NSE < 0.80, the model is satisfac-
Learn rate drop period 125
tory; for NSE < 0.60: the model is poor.

410
 Ecological Engineering & Environmental Technology 2025, 26(2), 407–419

Figure 2. Modeling steps

Coefficient of determination line, with a coefficient of determination almost

equal to zero (R2 = 0.08), which indicates a non-
The coefficient of linear determination (R2) is
linearity between precipitation and measured
a qualitative indicator of a simple linear regres-
runoff. These observations confirm that the re-
sion. It is commonly used in statistical studies;
lationship between these two variables is much
it allows evaluating the correlation between ob-
more complex, which requires the use of a more
served and simulated data (Renaud & Victoria-
sophisticated model that can be utilized to find a
Feser, 2010). This coefficient can be calculated
solution to this problem.
by means of the following formula:

(2) Study of the correlation between rainfall and

measured runoff using the LSTM model
where: is the average of simulated runoff.
Analysis of the linear correlation between pre-
It is worth pointing out that R varies from 0
2 cipitation and runoff gave rather poor results. It
to 1. The fit is considered perfect for R2 values​​ was therefore deemed necessary to test the prob-
equal to 1, and values ​​greater than 0.5 are gen- able existence of a relationship with LSTM while
erally accepted (Santhi et al., 2001; Van Liew et considering only rainfall as input data. Conse-
al., 2003). With regard to the performance ranges quently, very good results were obtained for both
of the R² coefficient, it was decided to adopt the the training and test periods (training period: R2
same thresholds as for the NSE. = 0.92 and NSE = 0.9; test period: R2 = 0.82 and
NSE = 0.79). The graphical results of the test pe-
riod, which obviously are the most interesting, are
presented below. Thus, (Figure 4) clearly shows
RESULTS AND DISCUSSIONS
that the points are grouped around the regres-
Study of the linear correlation between sion line, which confirms the existence of a sig-
rainfall and measured runoff nificant correlation. No apparent shift is observed
when the two curves depicted in Figure 5 are su-
The graph depicted in Figure 3 shows a sig- perimposed, which indicates a good fit between
nificant dispersion of points around the regression the observed and simulated data. These results

411
Ecological Engineering & Environmental Technology 2025, 26(2), 407–419

Figure 3. Linear correlation between precipitation and measured runoff, during the period between 2007 and 2015)

Figure 4. Linear regression between measured and simulated runoff, using the LSTM model,
for the test period between 2014 and 2015

confirm that it is more advantageous to use neural between different variables, using only a limited
networks than classical models which include at number of data (Yu et al., 2024)
least evapotranspiration, in addition to rainfall as
input data, and sometimes even other parameters, Rainfall-runoff modeling
such as pedology, land cover, slope, and others.
The development and adoption of deep learning The modeling, in which only precipitation
(DL) is highly beneficial for hydrological model- was used as input, gave quite satisfactory results.
ing as it helps to detect the non-linear relationship However, these results were further improved

Figure 5. Observed daily runoff and simulated runoff, using only measured rainfall
and observed runoff as inputs in the LSTM model

412
 Ecological Engineering & Environmental Technology 2025, 26(2), 407–419

by adding other parameters (Table 2). It should systems. On the other hand, the results ob-
also be noted that the performances are gener- tained from the ERA5-Land satellite are the
ally good, for training and prediction, with NSE weakest in comparison with the other sources.
values ranging from satisfactory to excellent, and The high performances from NOAA CPC and
R² values generally quite high. With regard to in- MERRA-2 are certainly due to the quality
terpretation of the results obtained, it was agreed and accuracy of the data that these satellites
to focus mainly on the prediction ones which are provide. Observations from both systems and
considered essential for the evaluation, compari- ground stations are also integrated (Okirya et
son and ranking of the models. These outcomes al., 2024; Yakubailik et al., 2020; Chaudhari et
allow a better understanding of the effectiveness al., 2004). The method adopted here allows to
of the different models tested. better estimate the variations of evapotranspi-
ration (Núñez et al., 2021; Olioso et al., 2022).
Interpretation of modeling results by input group It is noteworthy that evapotranspiration is a
crucial factor in a region like northwestern Al-
The main purpose of this section is to evaluate geria where climatic conditions vary consider-
the performance of each dataset in order to analyze ably due to the proximity to the Mediterranean
and compare the quality of modeling using input Sea and the Atlas Mountains (Born & Bachner,
parameters from different satellite data sources. 2003; Meddi, 2015). In addition, the NOAA
• Evapotranspiration – the impact of different CPC and MERRA-2 data are distinguished by
evapotranspiration data sources on rainfall- their high spatial and temporal resolution, their
runoff modeling reveals variations in per- advanced data assimilation techniques (Khat-
formance, with very satisfactory results ob- ibi & Krauter, 2021), and their adaptability to
tained data from NOAA CPC and MERRA-2 the specific conditions of regions with sparse

Table 2. Performance results and training parameters, using different satellite data
Training Prediction
Group Parameters Satellite sources
NSE R 2
NSE R2
NOAA CPC 0.92 0.94 0.84 0.85
ERA5_AG 0.91 0.93 0.81 0.87
GLDAS 0.88 0.89 0.73 0.78
1 Evapotranspiration MERRA2 0.92 0.92 0.87 0.89
CFSR 0.91 0.93 0.76 0.78
ERA5 0.88 0.89 0.79 0.79
ERA5 LAND 0.79 0.82 0.76 0.76
ERA5_AG 0.95 0.96 0.91 0.91
MERRA2 0.95 0.96 0.86 0.86
2 Temperature CFSR 0.80 0.81 0.61 0.61
ERA5 0.89 0.91 0.92 0.93
ERA5 LAND 0.89 0.90 0.82 0.82

Radiation, wind speed, MERRA2 0.79 0.81 0.75 0.79

3
humidity CFSR 0.96 0.98 0.81 0.83
GLDAS 0.91 0.93 0.78 0.87
4 Radiation MERRA2 0.76 0.78 0.77 0.78
CFSR 0.89 0.91 0.82 0.83
NOAA CPC 0.84 0.87 0.71 0.86
ERA5_AG 0.92 0.94 0.84 0.89
MERRA2 0.90 0.91 0.88 0.90
5 Temperature min-max
CFSR 0.83 0.86 0.75 0.75
ERA5 0.90 0.91 0.88 0.89
ERA5 LAND 0.93 0.95 0.92 0.93

413
Ecological Engineering & Environmental Technology 2025, 26(2), 407–419

monitoring network, such as northwestern semi-arid environments where they can have
Algeria (El & Kerroumi, 2022; White et al., a significant impact (Levesque, 2022). In ad-
2008)where the climatic conditions are quite dition, ERA5-Land adapts mainly to semi-arid
complex and can vary significantly. Such an regions where temperature directly affects hy-
approach allows providing more accurate and drological processes, which helps to achieve
reliable evapotranspiration estimates. more accurate runoff modeling. Finally, the
• Temperature – as concerns temperatures, the extended temporal coverage and data consis-
ERA5_AG data evidently achieve an excellent tency provided by ERA5-Land facilitate more
performance level. The very good outcomes efficient calibration of LSTM models (Muñoz-
are obviously attributed to the high resolution Sabater et al., 2021), which allows for more
of this data source, the quality of reanalyses, accurate simulations. All these observations
the combination of multiple data sources, the lead to the conclusion that ERA5-Land out-
long-term consistency of this source, its adapt- performs other data sources.
ability to various regions, as well as the con- • Radiation – the solar radiation results ob-
tinuous updates that ensure the accuracy and tained from the CFSR reanalysis product
relevance of temperature data (Kassem et al., show slightly better performance than other
2024). In compliance with these results, Ma sources. The Nash and R² values obtained​​
et al. (2008) also found out that data from the suggest that the model used simulates the flow
global reanalysis dataset ERA-40 provided an with a fairly high accuracy using the CFSR
excellent match with the temperature mea- radiation data. Moreover, the achieved per-
surements that were recorded in China, al- formance was found quite consistent with the
though ERA-40 is a rather old version. These findings of (Khaled et al., 2014) who evalu-
findings then suggest that the global weather ated the accuracy of CFSR solar radiation data
dataset ERA5_AG is the most reliable data and compared it with that of the ground-based
source because it offers better robustness and measurements made in the MENA region. The
precision for temperatures compared to other study found that CFSR data generally show
sources. These findings corroborate the excel- good agreement with ground-based observa-
lent NSE and R² results obtained. tions, especially in the regions of Egypt and
• Min-max temperature – the minimum and Northwest Africa where the normalized mean
maximum temperatures from the reanalysis bias coefficients were low and the correlation
dataset ERA5-Land allowed achieving an coefficients high, reaching the value of 0.982.
optimal flow simulation performance, which • Radiation, wind speed, humidity – the results
means that using min-max temperature data of the modeling integrating radiation, wind
from the ERA5-Land source leads to superi- speed and humidity, show that CFSR out-
or performances compared to those obtained performs MERRA2 as input data source in
from the other data sources. The good results the LSTM model. This is due to several fac-
obtained are due to several reasons. First, the tors. First, the spatial and temporal resolution
ERA5-Land source which offers a precise spa- of CFSR is slightly more precise than that of
tial resolution of 9 km (Arsenault et al., 2020) MERRA. Indeed, CFSR has a resolution of
is particularly advantageous for detecting lo- 0.5° x 0.5° (Saha et al., 2010), while MERRA-2
cal climatic variations. These variations are has a slightly lower resolution, namely 0.5° x
crucial in a semi-arid region like northwestern 0.625° (Khatibi & Krauter, 2021). It should be
Algeria where microclimatic variations are emphasized that this accuracy is particularly
important due to the diversity of terrain and important in regions, such as northwestern Al-
vegetation, and surface characteristics. In- geria, where specific climatic conditions can
deed, according to (Aimé & Remaoun, 1988), have a significant impact on the results (Irvem
these circumstances can change rapidly over & Ozbuldu, 2018; Khaled et al., 2014).
short distances. Next, the climate data pro-
vided by ERA5-Land are regularly updated Furthermore, the diversity of data types in-
and adjusted from multiple observations, tegrated in CFSR is remarkable, including direct
which ensures greater reliability of extreme measurements of wind speed and humidity, from
temperatures that are highly important in various sources, such as ocean buoys, weather

414
 Ecological Engineering & Environmental Technology 2025, 26(2), 407–419

stations and aircraft (Essou et al., 2016; Saha et the training phase is integrated into the evalua-
al., 2010). This multitude of data sources pro- tion, the ERA5-Land dataset stands out with an
motes a more robust modeling of dynamic param- excellent performance. It was indeed found that
eters, unlike MERRA-2 which could have a data the Nash coefficient was equal to 0.93 and R²
weighting and assimilation less adapted to these equal to 0.95 in the training phase, while the Nash
specific parameters (McClean et al., 2023). coefficient was equal to 0.92 and R² to 0.93 in the
validation period. Furthermore, (Figure 6) clearly
Global interpretation by comparison of the results shows that, for the test period, the data points are
of the modeling of the different input parameters tightly clustered around the regression line, con-
in the LSTM model from various satellite sources firming an almost perfect correlation between the
The results obtained with the LSTM model, observed and simulated data.
with the integration of different input parameters It should also be noted that the superposition of
from various satellite data sources, exhibit a vari- the two curves (Figure 7) shows that there is almost
ability that affects the flow simulation perfor- no mismatch between them, proving the strong fit
mance. Indeed, a difference of 0.31 was recorded between the observed and simulated results.
for NSE and 0.32 for R², between the temperature The excellent performance obtained with
data from the CSFR (NSE = 0.61 and R² = 0.61) ERA5-Land data can be explained as follows. It is
and the maximum and minimum temperature data observed that minimum and maximum tempera-
from ERA5-Land and ERA5 (NSE =0.92 and R² tures have a direct impact on several hydrologi-
= 0.93), which represents a very significant gap. cal processes, such as snowmelt, evaporation and
In addition, the performances of the minimum evapotranspiration, which significantly affect the
and maximum temperatures from the ERA5- river runoff (Wang et al., 2020). In addition, these
Land source and those of the temperature data temperatures reveal a wide range of climatic con-
from the ERA5 source are ranked equal, for the ditions, which allows the model to better detect and
considered validation periods. However, when predict watershed responses in extreme situations.

Figure 6. Linear regression between measured and simulated runoff, using the ERA5-Land dataset,
for the test period from 2014 to 2015

Figure 7. Observed and simulated daily runoffs, using the ERA5-Land dataset

415
Ecological Engineering & Environmental Technology 2025, 26(2), 407–419

This should increase and improve the accuracy of with recurrent neural networks. In the future, it
forecasts (Zhang et al., 2023). Likewise, the inte- would be interesting to explore other data sources
gration of minimum and maximum temperatures to further improve the accuracy of hydrological
from remote sensing allows the LSTM model to forecasts in various contexts.
better understand the complex dynamics of runoff
in watersheds, hence maximizing its performance
(Wang et al., 2024). In the same context, (Boul- REFERENCES
maiz et al., 2020) have also consistently found
that using these temperature data as inputs to the 1. Abdelmoneim, H., Soliman, M.R., Moghazy, H.M.
LSTM model contributes to significantly improv- 2020. Evaluation of TRMM 3B42V7 and CHIRPS
ing the modeling of the rainfall-runoff relation- Satellite Precipitation Products as an Input for Hy-
ship. Furthermore, these temperature data are gen- drological Model over Eastern Nile Basin. Earth
erally considered more reliable and less subject Systems and Environment 4(4), 685–698. https://
doi.org/10.1007/s41748-020-00185-3
to measurement or estimation errors compared to
other parameters, such as evapotranspiration or ra- 2. Aimé, S., Remaoun, K. 1988. Climatic variability
diation (Yilmaz, 2023). and steppisation in the Tafna basin (western Oranie)
(in french). Méditerranée 63(1), 43–51. https://doi.
org/10.3406/medit.1988.2528
3. Arsenault, R., Brissette, F., Martel, J.L., Troin,
CONCLUSIONS M., Lévesque, G., Davidson-Chaput, J., Gonzalez,
M.C., Ameli, A., Poulin, A. 2020. A comprehen-
When satellite data as well as rainfall and run-
sive, multisource database for hydrometeorologi-
off are used, the model shows generally good per- cal modeling of 14,425 North American watersheds.
formance for training and prediction, with NSE Scientific Data 7(1), 1–12. https://doi.org/10.1038/
and R² values ranging from satisfactory to excel- s41597-020-00583-2
lent. The findings indicate that the temperature 4. Bemmoussat, A., Korichi, K., Baahmed, D., Maref,
parameter presents the best performance, which N., Djoukbala, O., Kalantari, Z., Bateni, S.M. 2021.
corroborates its significant influence on evapo- Contribution of Satellite-Based Precipitation in Hy-
transpiration, a key factor in the water balance. drological Rainfall–Runoff Modeling: Case Study
In addition, the minimum and maximum tem- of the Hammam Boughrara Region in Algeria. Earth
peratures from the ERA5-Land source showed Systems and Environment 5(4), 873–881. https://
excellent performance during the calibration and doi.org/10.1007/s41748-021-00256-z
validation periods. They seem particularly well 5. Born, K., Bachner, S. 2003. Cyclogenesis and
suited to capture the hydrological dynamics in severe weather in the vicinity of the Atlas Moun-
our study area, outperforming all other param- tains: studies using a nonhydrostatic mesoscale
eters, such as evapotranspiration, mean tempera- atmospheric model, in: EGS-AGU-EUG Joint As-
ture, radiation, as well as the combination of radi- sembly. p. 11689. https://ui.adsabs.harvard.edu/
ation, wind speed, and humidity, whose impact is abs/2003EAEJA....11689B/abstract
more moderate. Based on these findings, it can be 6. Bouguerra, S.A., Mansour, B. 2023. Rainfall-flow
concluded that minimum and maximum temper- modeling using a global conceptual model: Case of
atures are the most influential input parameters, the beni bahdel watershed (northwest of algeria).
Journal of Water Management Modeling. https://
and that the ERA5-Land data source is particu-
doi.org/10.14796/JWMM.C500
larly effective for optimizing rainfall-runoff mod-
7. Boulmaiz, T., Guermoui, M., Boutaghane, H. 2020.
eling with the LSTM model. However, it should
Impact of training data size on the LSTM perfor-
be noted that the runoff simulation performance
mances for rainfall–runoff modeling. Modeling
varies in accordance with the satellite data source Earth Systems and Environment 6(4), 2153–2164.
used, because the choice of sources is highly cru- https://doi.org/10.1007/s40808-020-00830-w
cial in modeling optimization. Moreover, a sig-
8. Chaudhari, K.N., Sarkar, C., Patel, N.K., Parihar,
nificant gap of nearly 0.30 was observed between J.S. 2006. an Inter-Comparison of Satellite Based
the CSFR and ERA5-Land sources. The results Noaa Cpc Rainfall Estimates and Gauge Observa-
of the present study highlight the effectiveness tions Over Selected Stations in India. Proc of ISPRS
of satellite data in the LSTM model. They also symposium on Geospatial databases for Sustain-
emphasize the importance of correctly choos- able Development 27–30. https://www.isprs.org/
ing the input data for rainfall-runoff modeling proceedings/xxxvi/part4/RS-A-3.pdf

416
 Ecological Engineering & Environmental Technology 2025, 26(2), 407–419

9. Cho, K., Kim, Y. 2022. Improving streamflow pre- P. 2022. How can we benefit from regime informa-
diction in the WRF-Hydro model with LSTM net- tion to make more effective use of long short-term
works. Journal of Hydrology 605, 127297. https:// memory (LSTM) runoff models? Hydrology and
doi.org/10.1016/j.jhydrol.2021.127297 Earth System Sciences 26(22), 5793–5816. https://
10. Ciampalini, A., Garfagnoli, F., Del Ventisette, C., doi.org/10.5194/hess-26-5793-2022
Moretti, S. 2013. Potential Use of Remote Sens- 20. Hochreiter, S., Schmidhuber, J. 1997. Long
ing Techniques for Exploration of Iron Deposits in Short-Term Memory. Pakistan Journal of Zool-
Western Sahara and Southwest of Algeria. Natural ogy 9(8), 1735–1780. https://doi.org/10.1162/
Resources Research 22(3), 179–190. https://doi. neco.1997.9.8.1735
org/10.1007/s11053-013-9209-5 21. Huang, X., Lu, K., Wang, S., Lu, J., Li, X., Zhang, R.
11. Darand, M., Amanollahi, J., Zandkarimi, S. 2017. 2024. Understanding remote sensing imagery like
Evaluation of the performance of TRMM Multi- reading a text document: What can remote sensing
satellite Precipitation Analysis (TMPA) estimation image captioning offer? International Journal of Ap-
over Iran. Atmospheric Research 190, 121–127. plied Earth Observation and Geoinformation 131,
https://doi.org/10.1016/j.atmosres.2017.02.011 103939. https://doi.org/10.1016/j.jag.2024.103939
12. Dehni, A., Lounis, M. 2012. Remote sensing tech- 22. Inan, C.A., Maoui, A., Lucas, Y., Duplay, J. 2024.
niques for salt affected soil mapping: Applica- Multi-Station Hydrological Modelling to Assess
tion to the Oran region of Algeria. Procedia En- Groundwater Recharge of a Vast Semi-Arid Ba-
gineering 33, 188–198.https://doi.org/10.1016/j. sin Considering the Problem of Lack of Data: A
proeng.2012.01.1193 Case Study in Seybouse Basin, Algeria. Water
13. Derdour, A., Benkaddour, Y., Bendahou, B. 2022. (Switzerland) 16(1), 160. https://doi.org/10.3390/
Application of remote sensing and GIS to assess w16010160
groundwater potential in the transboundary water- 23. Irvem, A., Ozbuldu, M. 2018. Accuracy of satellite-
shed of the Chott-El-Gharbi (Algerian–Moroccan based solar data to estimate solar energy potential
border). Applied Water Science 12(6), 136. https:// for Hatay province, Turkey. Bitlis Eren Üniversi-
doi.org/10.1007/s13201-022-01663-x tesi Fen Bilimleri Dergisi 7(2), 361–369. https://doi.
14. Duan, Z., Tuo, Y., Liu, J., Gao, H., Song, X., Zhang, org/10.17798/bitlisfen.428757
Z., Yang, L., Mekonnen, D.F. 2019. Hydrological 24. Kassem, Y., Camur, H., Preala, T.A. 2024. Assess-
evaluation of open-access precipitation and air tem- ment of Wind Energy Potential for achieving Sus-
perature datasets using SWAT in a poorly gauged ba- tainable Development Goal 7 in the Rural Region of
sin in Ethiopia. Journal of Hydrology 569, 612–626. Jeje, Nigeria. Engineering, Technology & Applied
https://doi.org/10.1016/j.jhydrol.2018.12.026 Science Research 14(4), 14977–14987. https://doi.
15. El, N., Kerroumi, I. 2022. Study of the interan- org/10.48084/etasr.7311
nual variability of temperature and precipitation 25. Khaled, W., Authority, E.M., Afandi, G.S. El.
in the high plateaus (in french) 6, 65–77. https:// 2014. Evaluation of Ncep / Cfsr Solar Data Against
onm-blog.meteo.dz/wp-content/uploads/2024/08/ Ground Observation Over MENA. Open Journal of
Article_Kerroumi_2.pdf Atmospheric and Climate Change 1, 1–12. https://
16. Essou, G.R.C., Sabarly, F., Lucas-Picher, P., Bris- doi.org/10.15764/ACC.2014.02001
sette, F., Poulin, A. 2016. Can precipitation and 26. Khanal, S., Kushal, K.C., Fulton, J.P., Shearer, S.,
temperature from meteorological reanalyses be used Ozkan, E. 2020. Remote sensing in agriculture—ac-
for hydrological modeling? Journal of Hydrometeo- complishments, limitations, and opportunities. Re-
rology 17(7), 1929–1950. https://doi.org/10.1175/ mote Sensing 12(22), 1–29. https://doi.org/10.3390/
JHM-D-15-0138.1 rs12223783
17. Gebremicael, T.G., Mohamed, Y.A., Zaag, P. van 27. Khatibi, A., Krauter, S. 2021. Validation and perfor-
der, Gebremedhin, A., Gebremeskel, G., Yazew, mance of satellite meteorological dataset merra-2
E., Kifle, M. 2019. Evaluation of multiple satellite for solar and wind applications. Energies 14(4), 882.
rainfall products over the rugged topography of https://doi.org/10.3390/en14040882
the Tekeze-Atbara basin in Ethiopia. International 28. Lazri, M., Labadi, K., Brucker, J.M., Ameur, S.
Journal of Remote Sensing 40(11), 4326–4345. 2020. Improving satellite rainfall estimation from
https://doi.org/10.1080/01431161.2018.1562585 MSG data in Northern Algeria by using a multi-clas-
18. Greff, K., Srivastava, R.K., Koutník, J., Steunebrink, sifier model based on machine learning. Journal of
B.R., Schmidhuber, J. 2016. LSTM: A search space Hydrology 584, 124705. https://doi.org/10.1016/j.
odyssey. IEEE transactions on neural networks and jhydrol.2020.124705
learning systems 28(10), 2222–2232. https://doi. 29. Lees, T., Buechel, M., Anderson, B., Slater, L.,
org/10.48550/arXiv.1503.04069 Reece, S., Coxon, G., Dadson, S.J. 2021. Bench-
19. Hashemi, R., Brigode, P., Garambois, P.A., Javelle, marking data-driven rainfall-runoff models in Great

417
Ecological Engineering & Environmental Technology 2025, 26(2), 407–419

Britain: A comparison of long short-term memory Analysis of Annual Maximum Rainfall Using Ob-
(LSTM)-based models with four lumped concep- served and Remotely Sensed Data in the Tropical
tual models. Hydrology and Earth System Sci- Climate Zones of Uganda. Sustainability 16, 6081.
ences 25(10), 5517–5534. https://doi.org/10.5194/ https://doi.org/10.3390/su16146081
hess-25-5517-2021 40. Olioso, A., Jacob, F., Olioso, A., Jacob, F., De,
30. Levesque, G. 2022. Evaluation of the use of ERA5 E., Olioso, A. 2002. Estimation of evapotrans-
and ERA5-Land reanalysis data for evapotranspira- piration from remote sensing measurements (in
tion modelling (in french). Doctoral dissertation, french). la houille blanche 88(1), 62–67. https://
École de technologie supérieure. https://espace. doi.org/10.1051/lhb/2002008
etsmtl.ca/id/eprint/3085 41. Ravindra, B. 2018. Forecasting solar radiation
31. Ma, L., Zhang, T., Li, Q., Frauenfeld, O.W., Qin, D. during dust storms using deep learning. arXiv pre-
2008. Evaluation of ERA-40, NCEP-1, and NCEP- print arXiv:1808.10854. https://doi.org/10.48550/
2 reanalysis air temperatures with ground-based arXiv.1808.10854
measurements in China. Journal of Geophysical 42. Renaud, O., Victoria-Feser, M.P. 2010. A robust
Research Atmospheres 113, D15115. https://doi. coefficient of determination for regression. Journal
org/10.1029/2007JD009549 of Statistical Planning and Inference 140(7), 1852–
32. Mao, G., Wang, M., Liu, J., Wang, Z., Wang, K., 1862. https://doi.org/10.1016/j.jspi.2010.01.008
Meng, Y., Zhong, R., Wang, H., Li, Y. 2021. Com- 43. Ritter, A., Muñoz-Carpena, R .2013. Performance
prehensive comparison of artificial neural net- evaluation of hydrological models: Statistical sig-
works and long short-term memory networks for nificance for reducing subjectivity in goodness-of-
rainfall-runoff simulation. Physics and Chemistry fit assessments. Journal of Hydrology 480, 33-45.
of the Earth 123, 103026. https://doi.org/10.1016/j. https://doi.org/10.1016/j.jhydrol.2012.12.004
pce.2021.103026
44. Sabzipour, B., Arsenault, R., Troin, M., Martel, J.L.,
33. McClean, F., Dawson, R., Kilsby, C. 2023. Inter-
Brissette, F., Brunet, F., Mai, J. 2023. Comparing a
comparison of global reanalysis precipitation for
long short-term memory (LSTM) neural network
flood risk modelling. Hydrology and Earth System
with a physically-based hydrological model for
Sciences 27(2), 331–347. https://doi.org/10.5194/
streamflow forecasting over a Canadian catchment.
hess-27-331-2023
Journal of Hydrology 627. https://doi.org/10.1016/j.
34. Meddi, H. 2015. Variability of annual rainfall jhydrol.2023.130380
in north-western Algeria (in french). Science et
45. Saha, S., Moorthi, S., Pan, H.L., Wu, X., Wang,
changements planétaires/Sécheresse 20(1), 57–65.
Jiande, Nadiga, S., Tripp, P., Kistler, R., Woollen,
https://doi.org/10.1684/sec.2009.0169
J., Behringer, D., Liu, H., Stokes, D., Grumbine,
35. Muñoz-Sabater, J., Dutra, E., Agustí-Panareda, R., Gayno, G., Wang, Jun, Hou, Y.T., Chuang,
A., Albergel, C., Arduini, G., Balsamo, G., Bous- H.Y., Juang, H.M.H., Sela, J., Iredell, M., Treadon,
setta, S., Choulga, M., Harrigan, S., Hersbach, H., R., Kleist, D., Van Delst, P., Keyser, D., Derber,
Martens, B., Miralles, D.G., Piles, M., Rodríguez- J., Ek, M., Meng, J., Wei, H., Yang, R., Lord, S.,
Fernández, N.J., Zsoter, E., Buontempo, C., Thé- Van Den Dool, H., Kumar, A., Wang, W., Long, C.,
paut, J.N. 2021. ERA5-Land: A state-of-the-art Chelliah, M., Xue, Y., Huang, B., Schemm, J.K.,
global reanalysis dataset for land applications. Earth Ebisuzaki, W., Lin, R., Xie, P., Chen, M., Zhou,
System Science Data 13(9), 4349–4383. https://doi. S., Higgins, W., Zou, C.Z., Liu, Q., Chen, Y., Han,
org/10.5194/essd-13-4349-2021 Y., Cucurull, L., Reynolds, R.W., Rutledge, G.,
36. Nash, J.E., Sutcliffe, J. V. 1970. River Flow Forecast- Goldberg, M. 2010. The NCEP climate forecast
ing Through Conceptual Models - Part I - A Discus- system reanalysis. Bulletin of the American Me-
sion of Principles. Journal of Hydrology 10, 282–290. teorological Society 91(8), 1015–1057. https://doi.
https://doi.org/10.1016/0022-1694(70)90255-6 org/10.1175/2010BAMS3001.1
37. Netzband, M., Stefanov, W.L., Redman, C. 2007. Ap- 46. Santhi, C., Arnold, J.G., Williams, J.R., Dugas,
plied remote sensing for urban planning, governance W.A., Srinivasan, R., Hauck, L.M. 2001. Validation
and sustainability. Springer Science & Business Me- of the SWAT model on a large river basin with point
dia.https://doi.org/10.1007/978-3-540-68009-3 and nonpoint sources. Journal of the American Wa-
38. Núñez, P.Á., Silva, B., Schulz, M., Rollenbeck, R., ter Resources Association 37(5), 1169–1188. https://
Bendix, J. 2021. Evapotranspiration estimates for doi.org/10.1111/j.1752-1688.2001.tb03630.x
two tropical mountain forest using high spatial reso- 47. Shekar, P.R., Mathew, A., Yeswanth, P. V., Deival-
lution satellite data. International Journal of Remote akshmi, S. 2024. A combined deep CNN-RNN net-
Sensing 42(8), 2940–2962. https://doi.org/10.1080/ work for rainfall-runoff modelling in Bardha Water-
01431161.2020.1864058 shed, India. Artificial Intelligence in Geosciences 5,
39. Okirya, M., Du Plessis, J. 2024. Trend and Variability 100073. https://doi.org/10.1016/j.aiig.2024.100073

418
 Ecological Engineering & Environmental Technology 2025, 26(2), 407–419

48. Ur Rahman, K., Shang, S., Shahid, M., Wen, Y. org/10.1016/j.agrformet.2008.05.017

2020. Hydrological evaluation of merged satellite 54. Wu, J., Chen, X.-Y., Zhang, H., Xiong, L.-D., Lei,
precipitation datasets for streamflow simulation us- H., Deng, S.-H. 2019. Hyperparameter optimiza-
ing SWAT: A case study of Potohar Plateau, Paki- tion for machine learning models based on Bayes-
stan. Journal of Hydrology 587, 125040. https://doi. ian optimization. Journal of Electronic Science and
org/10.1016/j.jhydrol.2020.125040 Technology 17(1), 26–40. https://doi.org/10.11989/
49. Van Liew, M.W., Arnold, J.G., Garbrecht, J.D. 2003. JEST.1674-862X.80904120
Hydrologic simulation on agricultural watersheds: 55. Yakubailik, O.E., Yakubalik, T. V. 2020. Analysis
Choosing between two models. Transactions of the of accumulated precipitation based on satellite data
ASAE 46(6), 1539–1551. https://doi.org/https://doi. in Central Siberia. IOP Conference Series: Earth
org/10.13031/2013.15643 and Environmental Science 548, 032025. https://
50. Vasiliou, A., Economides, A.A. 2006. MANETs doi.org/10.1088/1755-1315/548/3/032025
for environmental monitoring. 2006 International 56. Yilmaz, M. 2023. Accuracy assessment of tempera-
Telecommunications Symposium, IEEE 813–818. ture trends from ERA5 and ERA5-Land. Science
https://doi.org/10.1109/ITS.2006.4433383 of the Total Environment 856, 159182. https://doi.
51. Wang, H., Zhang, J., Yang, J. 2024. Time series fore- org/10.1016/j.scitotenv.2022.159182
casting of pedestrian-level urban air temperature by 57. Yu, C., Hu, D., Shao, H., Dai, X., Liu, G., Wu, S.
LSTM: Guidance for practitioners. Urban Climate 56, 2024. Runoff simulation driven by multi-source
102063. https://doi.org/10.1016/j.uclim.2024.102063 satellite data based on hydrological mechanism
52. Wang, K., Xu, Q., Li, T. 2020. Does recent climate algorithm and deep learning network. Journal of
warming drive spatiotemporal shifts in functioning Hydrology: Regional Studies 52, 101720. https://
of high-elevation hydrological systems? Science doi.org/10.1016/j.ejrh.2024.101720
of the Total Environment 719, 137507. https://doi. 58. Zhang, Y.-R., Shang, G.-F., Leng, P., Ma, C., Ma, J.,
org/10.1016/j.scitotenv.2020.137507 Zhang, X., Li, Z.-L. 2023. Estimation of quasi-full
53. White, J.W., Hoogenboom, G., Stackhouse Jr, P.W., spatial coverage soil moisture with fine resolution
Hoell, J.M. 2008. Evaluation of NASA satellite-and in China from the combined use of ERA5-Land re-
assimilation model-derived long-term daily temper- analysis and TRIMS land surface temperature prod-
ature data over the continental US. agricultural and uct. Agricultural Water Management 275, 107990.
forest meteorology 148(10), 1574–1584. https://doi. https://doi.org/10.1016/j.agwat.2022.107990

419