Professional Documents
Culture Documents
Research Article
Keywords: Daily reference evapotranspiration, Arti cial neural network, North of Algeria, Standardized
Penman-Monteith equation, Hargreaves-Samani equation,the POWER-NASA
DOI: https://doi.org/10.21203/rs.3.rs-2585756/v1
License: This work is licensed under a Creative Commons Attribution 4.0 International License.
Read Full License
4 1
Ecole nationale Supérieure de l’Hydraulique. LGEE laboratory. Blida. Algeria. k.gherbi@ensh.dz.
5 Corresponding author.
6 2
Ecole nationale Supérieure de l’Hydraulique. LGEE laboratory. Blida. Algeria. a.ammari@ensh.dz.
7 Corresponding author.
8 3
Ecole nationale Supérieure de l’Hydraulique. LGEE laboratory. Blida. Algeria. m.meddi@ensh.dz
9 Abstract
10 The estimation of reference evapotranspiration ET0 is required to know the water needs of crops in
11 advance and to plan and manage water resources. However, to obtain reliable irrigation management results,
12 simplified artificial neural networks models and prediction algorithms must be applied and evaluated. These
13 models were examined by comparing the multilayer perceptron network, and the radial basis function network, to
14 estimate the reference evapotranspiration for planning tool irrigation, The Standardized Penman-Monteith
15 evapotranspiration equation has been recommended by the American Society of Civil Engineers and Hargreaves-
16 Samani were used to calculate the daily reference evapotranspiration values. Due to a lack of meteorological
17 dataset, NASA Prediction of Worldwide Energy Resources NASA POWER database was used and validated by
18 comparing it with observed values. These daily sets of meteorological data included the minimum and maximum
19 air temperature, the temperature of the dew point, solar radiation, and the wind speed of three stations Mouzaia,
20 Soumaa, and Boukourdane in the Mitidja plain, Algeria. The validity of 4 combinations scenarios was investigated,
21 and the performance of the models was assessed with the coefficient of determination, the average quadratic error,
22 and the mean absolute percentage error, between the estimated and calculated reference evapotranspiration values.
23 The results confirm the feasibility of obtaining appropriate level estimates of ET0 based on the agroclimatological
24 POWER-NASA archive, even for some Mediterranean countries where climate variables are not available at a
25 large scale, which can help us understand the mechanism of reference evapotranspiration and use it for water
27 Keywords: Daily reference evapotranspiration, Artificial neural network, North of Algeria, Standardized
1
29 1. Introduction
30 Water resources in the Mediterranean region are scarce and unbalanced, particularly in North Africa (le
31 Page et al., 2020) . The primary high-water demand in the study region concerns agriculture, which uses between
32 70 and 80 percent of global water resources (Ouassanouan et al., 2022) and (Crovella et al., 2022).Most farmers
33 use more or less irregular irrigation schedules that are not based on weather and soil conditions (Maselli et al.,
34 2020) .One of the most crucial components of the management and planning of water resources is evaporation,
35 which is a strong process (PENMAN, 1948) and (Stewart, 1984). Because of this, farmers can increase the
36 effectiveness of irrigation planning and water resource management by precisely estimating the amount of water
37 needed for each crop and anticipating the various agronomic and weather conditions that will affect that crop's
38 growth (Mohammadi & Mehdizadeh, 2020).As a result, accurate estimates of the water requirements of crops
40 The simulation of the reference evapotranspiration ET0 research has made use of several models based
41 on meteorological data(Izadifar & Elshorbagy, 2010) . Several researchers have developed many mathematical
42 models and empirical equations to estimate the ET0 for instance in (Celestin et al., 2020) and (Ahooghalandari et
43 al., 2016). As well as the Penman-Monteith equation (Allen et al., 2006) which was proposed by the Organization
44 of United Nations for Food and Agriculture FAO (FAO56-PM) as a standard method for estimating the ET0 and
45 calibrating other empirical and semi-empirical models. Among the many forms of the PM equations, the most
46 widely used equations for the tall reference crop are the full version Penman-Monteith equation (Allen et al., 2005)
48 calculate daily reference evapotranspiration using daily air temperature data (Hargreaves & Samani, 1982) .
49 A more direct and effective response may be provided by machine learning techniques (ML). These
50 approaches are useful tools for creating hydrological quantity forecasting models. They have been used in some
51 studies within the framework of ET0 evaluation models in recent years (Dou & Yang, 2018; Granata, 2019;
52 Reis et al., 2019) In addition to using usual algorithms, artificial neural network (ANN) techniques have been
53 used to predict both reference evapotranspiration (Huo et al., 2012; Laqui et al., 2019; Özgür & Yamaç, 2020;
55 The climate data for three stations in the current study come from NASA's POWER project (Prediction
56 of Worldwide Energy Resources) compared with observed data from ANRH of Blida (National Hydraulic
2
57 Resources Agency). Statistical analysis was done with the observed data. The POWER NASA base was chosen
58 because the data are generally contemporaneous and global and help to generate very large data archives. These
59 data include average long-term climatological estimates of meteorological quantities and surface solar energy
60 fluxes where surface measurements are infrequent or non-existent. Numerous studies have been carried out using
61 NASA POWER data and agree with an appropriate level of precision. In the continental United States(White et
62 al., 2011).They have concluded that the NASA POWER solar data are useful for investigations that call for a
63 thorough explanation of historical variation, In Italy (Negm et al., 2017a) findings showed that the archive
64 POWER-NASA is suitable for estimating ET0 in locations where the vast majority of climate variables. In Oman
65 (Marzouk, 2021) has been found that according to the statistical analysis and visual inspection, the NASA POWER
66 measurements for the air temperature at 2 m are valid. Adding that this approach has been used by (Jiménez-
68 The aim of this work is to estimate the daily reference evapotranspiration ET0 using minimal of
69 parameters by developing a neural networks models ANN-MLP and ANN-RBF based on the Standardized PM
70 and HS equations, However the NASA POWER data have been statistically analysed and validate.
73 The Mitidja plain is located in the northwest of Algeria. Covers an area of 1,700 km² and extends over
74 the territory of Algiers, Boumerdes, Tipaza, and Blida. It is located between latitudes 36 36 ° 36 ° 25′N and 36°
75 48′N, and longitudes 2 ° 32′E and 3 ° 20′E in the south of the wilaya of Algiers with an average elevation of 300
76 m over the sea level with an agricultural area ranging from 120,000 to 130,000 ha (Bouziane et al., 2021) .
3
77
80 The climate data for the current study for different periods for three meteorological stations of Mouzaia,
81 Soumaa, and Boukourdane which are located in the Mitidja plain, comes from:
85 The POWER NASA database is complete and provides daily records of air temperature (Tmin, Tmax) (°C)
86 dew point temperature ( Td°C), wind speed (U₂ m/s) at 2 m height, and solar radiation(Rs MJ/m2), All comparisons
87 between the two datasets were limited to days when all relevant were available ,for the periods used in this study.
88 The reference evapotranspiration ET0 was calculated using equations (3) and (4). The details of the data sets in
4
89 Table 1, name, elevation, geographical coordinates, and period used in this study of the POWER-NASA, and
92
93 3. Methods
94 The models based on the meteorological parameters from two databases reel and POWER NASA for
95 calculating the ET0 by the HS equation and Standardized PM equation were estimated by ANN-MLP and ANN-
96 RBF models. The optimal model and the evaluation of its performance were performed using the R², RMSE, and
97 MAPE. A detail of the process for estimation the reference evapotranspiration ET0 is illustrated in Fig.2.
5
98
100 The Environmental Water Resources Institute-American Society of Civil Engineers has adopted the
101 following steps required to estimate reference evapotranspiration ET0 for a 0.12 m tall reference surface using
102 daily weather data (Allen et al., 2005b). Where (Forster et al., 2022) indicate that the Penman-Monteith ASCE-
103 EWRI model had the highest predictive power. Based on the Penman-Monteith equation PM-ET0 has been used
104 as a shortcut for the symbols in this study, The equation for estimating ET0 from (George H. Hargreaves &
105 Zohrab A. Samani, 1985) is also presented, and the symbol used is HS-ET0.
107 The Penman-Monteith approach is used in the FAO (Food and Agriculture Organization of the United
108 Nations) study, often referred to as the combined method since it combines the energy balance and mass transfer
109 method, is used to estimate the reference evapotranspiration ET0 (PENMAN, 1948) (Monteith, J.L.1965). Daily
110 meteorological data such as air temperature, solar radiation, relative humidity, wind speed, and other estimated
6
113 Where ET0 is the reference evapotranspiration (mm/day), Rs is the net solar radiation (MJ/m2/day), G is
114 the soil heat flux in (MJ/m2/day), T is the mean air temperature in (°C), U2 is the mean wind speed in (m/s), (es –
117 The site characteristics, including latitude and elevation (m) above sea level, must be entered. The
118 necessary weather data include solar radiation (MJ /m²day), maximum and minimum air temperatures (°C),
119 average wind speed (m/ s) and average dew point temperatures (°C). The temperature of the air and w point should
120 be measured between 1.5 and 2 m and the wind speed at 2 m (Snyder et al., n.d.).
121 The dew point temperatures were used to estimate standardized PM-ET0, using the following formula
124 Where:
125 T mean: Mean temperature °C
126 Rh: Relative humidity %
127 In this research the reference evapotranspiration ET0 was calculate by using the ASCE-EWRI (Walter et al.,
128 2004) standardized equation as follows:
129 ET0 = R0 + A0 (3)
130 Where:
131 R0 is the radiation term of the Penman-Monteith equation for short canopy reference evapotranspiration with
132 U₂ the wind speed at 2 m height and 0.408 is the units from MJ m²/ day to mm/day, and A0 is the aerodynamic
133 term of the Penman-Monteith equation for short canopy reference evapotranspiration.
134 3.3 Hargreaves-Samani
135 If the data on solar radiation, relative humidity and wind speed are missing, they should be estimated by
136 the procedures provided in this section. As an alternative, ET0 can be estimated using the Hargreaves- Samani
137 equation (George H. Hargreaves & Zohrab A. Samani, 1985) . The Hargreaves-Samani method is suggested
138 by FAO as a different method for estimating ET0 when Penman-Monteith meteorological data are
139 insufficient(Allen et al., 1998). However, the FAO propose that the use of the Penman-Monteith method with
140 estimated solar radiation, vapor pressure, and wind speed generally provides more exacte estimates than the
7
143 Where ET0-HS is the daily reference evapotranspiration of grass (mm/d), Tmax, Tmin, and Tmean are the
144 maximum, minimum, and mean temperature (°C), respectively. Extraterritorial radiation (Ra) can be estimated
145 from the solar constant, the solar declination, and the time of the year by the following:
24(60)
146 𝑅𝑎 = 𝐺𝑠𝑐 𝑑𝑟(𝜔𝑠 sin(𝜑) sin ( 𝛿) + cos(𝛿) cos(𝜑) sin ( 𝜔𝑠)) (5)
𝜋
147 Where:
148 Ra extraterrestrial radiation (MJ/m²/day), Gsc solar constant = 0.0820 MJ m²/min,dr inverse relative
149 distance Earth-Sun,ωs sunset hour angle (rad),φ latitude (rad), 𝛿 solar decimation (rad) (FAO, s.d.).
150 In general, the FAO will provide the Penman-Monteith equation with somewhat more accurate estimates
151 than estimating ET0 directly using the Hargreaves-Samani equation. This is due to the ability of the estimating
152 equations to incorporate general climatic characteristics such as high or low wind speed or high or low relative
153 humidity into the ET0 estimate. Equation (4) tends to underestimate in high wind conditions (U₂ > 3 m/s) and
155
157 Artificial neural networks (ANNs) have long been used in modelling. Since their inception in
158 1943(McCulloch & Pitts, 1943) . Specifically, to solve problems that include stochastic variables, nonlinear or
159 unknown variations, or those that need to be determined from a less controlled environment (Moustris et al.,
160 2011).Artificial neural networks (ANN) are prediction tools that are universally recognized as a technique to
161 address complex problems and systems to ensure predictions and generalizations(Afram et al., 2017).They can
162 solve difficult problems that cannot be stated using mathematical procedures. Different neural networks can be
163 trained based on the arrangement of neurons and the patterns of layer connections. MLP and RBF neural networks
164 are two of the best-known neural networks with a wide range of applications in problem-solving(Hashemi Fath et
167 The multilayer perceptron network MLP, which consists of one or more hidden layers with a variable
168 number of neurons, and the choice of the activation function depending on the dataset, is one of several forms of
169 neural networks (Dikshit et al., 2022)The multilayer perceptron network MLP consists of three layers: the input
170 layer, the hidden layers, and the output layer Fig.3. An N-neuron can be mathematically characterized by the
171 following two equations:
8
172 𝑦𝑁 =f (𝑢𝑁 +𝑏𝑁 ) (6)
173 𝑢𝑁 =∑𝑛𝑖=1 𝑤𝑁𝑖 𝑥𝑖 (7)
174 Where:
175 x1, x2, x3,..., xn are the input signals to the neuron,
176 wN1, wN2, wN3,..., wNn are the connection weights of the neuron,
177 uN is the linear output of the linear combination of the weighted inputs,
178 bN is the bias term,
179 f is the activation function and yN is the output of the neuron.
180
181
182 Fig.3: Structure of a general regression neural network
183 3.4.2 Radial basis function RBF neural networks
184 The Radial Basis Function RBF is a mathematical function used for neural networks that have three
185 completely related layers (input layer, hidden layer, and output layer), such as a multilayer perceptron MLP, but
186 its manifold and activator functions have been substituted for each other. The activators of the hidden layer neurons
187 are based on radial functions (Gaussian functions), while the functions of the output layer neurons are linear.
188 Compared to other networks, the RBF is an effective feed forward neural network, it has the advantage of having
189 local minimum problems and a faster learning approach (Radial), it usually consists of an input layer, hidden layer
190 and output layer (Deng et al., 2021).The activation function of the hidden layer neurons is given by the Gaussian
191 function φ i (x):
⃦𝑥−𝑐𝑖 ⃦
( )
192 ∅𝑖 (𝑥) = −𝑒 2𝜕2 (8)
9
193 𝑦𝑛 = ∑𝑛𝑖=0 ∅𝑖 𝑤𝑗 (9)
194 Where:
195 φ i is the center, i is the propagation of node i, are important parameters in the performance of these
196 networks, The weighted sum of the hidden layer outputs is used to produce the network output y n , and ordinary
197 least squares are used to calculate the weight vector w j .
198 3.4.3 Assessment of model performance
199 As performance indicators for the proposed fits to the ANN model, statistical measures such as RMSE
200 (Root Mean Square Error), MAPE (Mean absolute percentage error), and R² (Coefficient of Determination)
201 between actual and expected values of ET0 were used. The standard deviation of the difference between
202 expected and actual values is used to calculate the goodness of fit measure, RMSE, which is provided by (10)
203 :
𝑇𝑥 1
204 𝑅𝑀𝑆𝐸 = √ ∑𝑖=1 (𝐸𝑇𝑒𝑠𝑡 − 𝐸𝑇𝑐𝑎𝑙 )2 (10)
𝑇𝑥
205 The coefficient of determination (R²) is a measure of the degree of linearity between the predicted and
206 actual values of an output variable, and is given by (11) :
(∑𝑛 ̅̅̅̅̅̅̅̅ ̅̅̅̅̅̅̅̅
𝑖=1(𝐸𝑇𝑐𝑎𝑙 −𝐸𝑇𝑒𝑠𝑡 )(𝐸𝑇𝑐𝑎𝑙 −𝐸𝑇𝑒𝑠𝑡 ))
207 𝑅2 = ∑𝑛 (𝐸𝑇 ̅̅̅̅̅̅̅̅ )2 ∑𝑛 (𝐸𝑇 ̅̅̅̅̅̅̅̅ 2 (11)
𝑖=1 𝑐𝑎𝑙 −𝐸𝑇 𝑒𝑠𝑡 𝑖=1 𝑐𝑎𝑙 −𝐸𝑇 𝑒𝑠𝑡 )
208 The mean absolute percentage error (MAPE) method was used to calculate the difference between the
209 simulated and calculated values. It is determined as follows (12):
∑𝑛
𝑖=1|𝐸𝑇𝑐𝑎𝑙 −𝐸𝑇𝑒𝑠𝑡 |
210 MAPE = 100 % (12)
𝑛
10
Adjusted R
0,48 0,29 0,01 0,42
Square
Standard Error 5,63 7,24 1,99 4,15
Significance F 0 0 0 0
P-value 0 0 0 0
Soumaa
Multiple R 0,43 0,4 0,04 0,33
R Square 0,18 0,16 0 0,11
Adjusted R
0,18 0,16 0 0,11
Square
Standard Error 7 6,26 3,09 5,43
Significance F 0 0 0.09 0
P-value 0 0 0.09 0
223
224 The results found show that the multiple R is higher than 0.50 for the max temperature the dew point
225 temperature and the min temperature (0.64, 0.66,0.58) (0.69, 0.54, 0.58) for Mouzaia, Boukourdane stations, this
226 represents the multiple correlation between the real observations data and POWER NASA data is bit close, but it
227 is low for the wind speed (0.012, 0.11) in this case the data are poorly fitted.
228 Additionally, the description of the standard deviation of the normal distribution (standard error) is small
229 for the wind speed (1.029, 1.99); in another case, the standard error of max, min, and dew temperatures is between
230 3.5 and 7 for the three stations means that the fidelity of the statistical regression is too wide for our prediction
231 interval.
232 The following figures (04,05,06) show the NASA POWER data and observed measurements for the daily
233 maximum temperature in °C, the daily minimum temperature in °C, the daily wind speed in m/s, (at 2m above
234 ground) based on meteorological stations.
235
11
236
237 Fig.4: Daily maximum air temperature (° C) (at 2 m above ground level) from NASA POWER data and
238 measurements.
239
240 The maximum air temperature graphs show the differences between the two data sets, but are
241 not significant, and generally follow the same trend, indicating that NASA POWER data are valid.
242 According to (White et al., 2008) topography and changes in land use or localized effects that could be
243 sources of errors in adaptation models that could explain the variability of the data.
244
245
246 Fig.5: Daily minimum air temperature (at 2 m above ground level) from NASA POWER data and
247 measurements.
248
249 For each weather station, the minimum air temperature graph shows a degree of respect for both
250 data sets, the measured data and those estimated by NASA POWER vary considerably, and these
251 differences have changed over the days. It is a favourable result and a reliable NASA POWER data; these
252 results line up with those attained by (Duarte & Sentelhas, 2020) in Brazil.
12
253
254 Fig.6: Daily wind speed (at 2 m above ground level) from NASA POWER data and measurements
255 The wind speed graph shows significant changes, for all stations there is a largest difference between
256 NASA POWER data and real meteorological data. NASA POWER values are significantly above reality, but
257 unlike temperatures, there is no general agreement between NASA POWER data points and measurement data
258 points. Both curves are the same measurements, but the trend of change is slightly different (Najmaddin et al.,
259 2017) indicate the same results.
260 The comparisons presented earlier between the NASA POWER data and the observed meteorological
261 data for three meteorological stations of variables: maximum and minimum air temperature at 2 m and wind speed
262 at 2 m, there is a good agreement between NASA POWER reanalysis and observed data for all parameters of max,
263 min, and dew temperatures, except for wind speed, this behaviour is reported by(Rodrigues & Braga, 2021) .
264 Moreover, according to (Negm et al., 2018) the downscaled POWER-NASA air temperature improves the ET0
265 estimation, which is more noticeable at higher altitudes.
266 Table 03: Comparison of statistical data from Mouzaia soumaa and Boukourdane meteorological stations
267 (observed value) and NASA POWER (NASA), analyzing daily data.
T max(°C) NASA- 46,50 7,76 22,93 42,24 1,00 22,86 40,46 4,42 22,56
POWER
Observed 40,53 1,90 25,24 41,50 2,86 21,51 49,00 2,00 27,25
T min(°C) NASA- 33,00 2,64 13,87 27,54 -0,74 12,44 27,01 0,14 13,49
POWER
Observed 27,19 -1,00 13,49 35,00 -1,50 15,88 34,00 -12,00 11,30
13
Rs(MJ/m2) NASA- 31,24 9,39 20,63 32,26 8,82 20,44 32,26 8,82 20,44
POWER
U₂(m/s) NASA- 9,64 0,91 2,57 6,74 0,75 2,34 7,69 0,93 2,60
POWER
Observed 62,06 0,00 1,90 10,90 0,02 3,26 19,44 0,74 5,29
Tr (°C) NASA- 20,28 -0,20 10,57 20,53 -1,31 9,62 21,13 -0,82 10,43
POWER
Observed 40,30 1,58 19,94 31,03 -0,12 15,03 30,02 0,49 15,36
PM NASA- 9,88 0,51 4,02 10,08 0,69 4,02 9,37 0,71 3,94
ET0(mm/day) POWER
Observed 7,31 0,03 3,01 8,63 0,31 3,52 12,49 0,82 4,69
HS ET0 NASA- 8,21 0,45 3,20 8,27 0,57 3,44 7,24 0,51 3,16
(mm/day) POWER
Observed 9,79 0,24 3,71 6,99 0,28 2,46 10,77 0,00 4,00
268
269 Table 03 compares 3 statistical values for each of the four climate variables, calculated from NASA
270 POWER and observed from the three weather stations using the Standardized PM and HS equations. These
271 statistical values are the maximum, minimum, and arithmetic mean. The air temperature shows small numerical
272 differences for all three statistical values, which makes the NASA POWER data reliable for this particular variable.
273 For wind speed shows, there is a difference between the NASA POWER data and the observed data and a minor
274 difference between PM-ET0 and HS-ET0 for all stations. Without exception, each of the years considered shows
275 that it is not possible to obtain a realistic wind speed and that the real situation is much more reliable than suggested
276 by NASA POWER due to the simulation of Standardized PM and HS results.
277 Table 04: Numerical level of absolute daily deviation for three meteorological properties between the observed
278 value and the NASA POWER value.
Stations Parameter The arithmetic mean of Root mean square of
(NASA – observed) (NASA – observed)
Mouzaia T max 2.31 6.56
T min 0.38 4.96
Td 9.36 10.93
U₂ 0.67 3.16
Soumaa T max 1.35 8.80
T min 3.44 8.00
Td 5.40 8.00
U₂ 0.91 3.37
Boukourdane T max 4.69 8.91
T min 2.19 7.68
Td 4.92 6.50
U₂ 2.68 3.44
279
280 Table 04 describes the quantity of the mismatch between the NASA POWER data and the observed data
281 for the four weather variables. The table gives two figures for each variable, the arithmetic mean of the absolute
282 difference between the values based on NASA POWER data and the values based on observed data for the daily
283 weather variable is a whole number. The other value is the root mean square of the daily difference records, which
284 is the square root of the daily mean differences squared. The two numbers measure the disparity between the two
285 data sets. They become not equal in the special case where the differences are identical on all days. For
286 temperatures, the two figures expressing the levels of disparity are within. This is not a huge difference and is
14
287 consistent with the discrepancy observed visually earlier when the data were plotted. The daily absolute differences
288 in the wind speed variable also show a much larger RMS than the mean, indicating an irregularity in these daily
289 differences. The degree of variation between the two data sets, with the observed data serving as the true reference.
290 4.2 Model Building
291 Determining the combination of input, model structure, important parameters, and performance
292 objectives, as well as the successful implementation of ANNs models, are all problems(Yu et al., 2020) .The
293 input variables that can be used to create the model in the current study are the maximum and minimum
294 temperature and dew point, solar radiation, wind speed, and relative humidity. Of these parameters, the
295 variables that are best correlated with the ET0 output were chosen for the build of the model. Table 05 shows
296 that solar radiation is the most correlated with ET0, with correlation values of 0.903 in the meteorological
297 station, Maximum and dew point temperatures also have a strong relationship with ET0, with values of 0.902
298 and 0.901, Consequently, 04 sets of reduced input combinations and one set of all available inputs were used
299 in this study, as shown in Table 06.
300 Table 05: Correlation of meteorological variables with ET0.
Mouzaia Rs Tx Tn U2 Td ET0
Rs 1
Tx 0,664 1
Tn 0,617 0,938 1
U2 -0,183 -0,325 -0,234 1
Td 0,654 0,9888 0,9789 -0,29 1
Et0 0,903 0,902 0,8687 -0,219 0,901 1
301
302 Table 06: Combinations of input variables.
303 Input combinations Input variables
304
1 Rs,Tmax,Tmin
305
2 Rs,Tmax,Tmin,Td,U₂
306
3 Tmax,Tmin,Td,U₂
307
4 Td,Tmax,Tmin
308
309 4.3 Daily estimation of reference evapotranspiration ET0
310 In this study, various inputs were used and one output was to estimate the reference evapotranspiration ET0
311 to develop the ANNs models by the MLP and RBF algorithms. It was estimated in different ways each case used
312 a method for estimating Standardized PM or HS and a different meteorological data source measured stations and
313 POWER-NASA data .For the Boukourdane and Soumaa stations, the model (ANN-MLP) where ET0 calculated
314 by Standardized PM equation and with three hidden layers (8-4-1), in terms of the lowest RMSE values (0,17 0,10
315 ), MAPE (2,10 2,43), the model that includes five inputs (Rs, Tmax, Tmin, Td, U₂) was in a good range based on the
316 R² test (0,995) and the model that includes 3 inputs (Rs, Tmax, Tmin) has all so a low RMSE (0.0204), MAE (0.1037)
317 and a good R² (0.9940), For the station of Boukourdane the model (ANN-MLP) where ET0 calculated by the HS
318 equation in terms of the lowest RMSE values (0,654), MAPE (9,5573), the model which includes 4 inputs (Tmax,
319 Tmin, Td, U₂) was in a good range based on the R² test (0,9386) and the model that includes 3 inputs (Td, Tmax, Tmin)
320 based on the R² test ,has also a good range.
321 In the Boukourdane station the model (ANN-RBF) with 300 neurons, ET0 was calculated by the PM method,
322 the algorithm with three inputs ( Rs, Tmax, Tmin) of the POWER-NASA data had the highest accuracy determined
323 based on the smallest RMSE (0,7352), MAPE (13,8386) and the largest R² (0,8947). when ET0 was calculated
324 by the HS method, the algorithm with three inputs ( T max, Tmin, Td) of observed data had the highest accuracy
325 determined based on the smallest RMSE (1,1178), MAPE (13,6999) and the largest R² (0,8234).The number of
326 neurons in the hidden layer is an important factor in the structure of ANNs that depends on the type and number
327 of inputs. In this study, to achieve the highest accuracy in ANN, the number of neurons in the hidden layer was
328 determined by trial and error. This process was also repeated for MLP and RBF, but the number of inputs in ANN-
329 MLP and ANN-RBF was changed due to the scenario and the number of fixed neurons. Structure 8-4-1 represents
15
330 the hidden layers in ANN-MLP with different inputs and one output (ET0). The ANN inputs (except for the ANN
331 MLP) are in the right range according to R². ANN- RBF had a large number of input neurons (300 neurons) in
332 the hidden layer, resulting in an acceptable ET0 estimation accuracy. Table 07and Table 08 shows the results of
333 the ANN-MLP and ANN-RBF models.
334 Table 07: Performance of the ANN MLP and ANN RBF models during the two periods of NASA-POWER data
335 and observed data using PM-ET0 method.
16
NASA-POWER 0,7629 24,7193 0,9735 0,5084 31,3194 1,3168
Td, Tmax, Tmin Observed 0,7237 21,4839 0,7128 0,4066 32,9697 0,9207
NASA-POWER 0,7266 24,9764 1,0416 0,5542 29,625 1,2582
Boukourdane Tmax,Tmin,Td,u2 Observed 0,95 8,8096 0,59 0,6875 22,3219 1,6656
NASA-POWER 0,9004 14,5271 0,5923 0,6803 24,1508 1,0694
Td, Tmax, Tmin Observed 0,9386 9,5573 0,654 0,8234 13,6999 1,1178
NASA-POWER 0,9049 14,1207 0,5777 0,779 18,5096 0,8835
339
340
341 In general, Tables 07,08 shows that model accuracy depends on the number of input variables and
342 combinations of variables, the data set, and ET0 calculation methods. It was also found that neural networks
343 performance varied with the number of inputs as well as the predicted time steps. Combining with more than three
344 parameters showed better performance. The results demonstrated the typical ability of ANNs models to handle
345 complex tasks, as reported in previous studies for instance in (Antonopoulos & Antonopoulos, 2017).In
346 combinations (1,2) the statistics presented in Table 07 directly indicate the discrete effect of Rs on the ET0
347 estimate. Combinations (3,4) clearly fail to improve the performance of ANNs models when incorporated into the
348 input sets, showing that wind speed U₂ has only a small effect on ET0. Therefore, it is redundant to add the dew
349 point temperature Td in the input sets of the ANNs models for the estimation of ET0. To summarise the above
350 discussion, the 3 combinations (1,2) can perfectly estimate ET0 when input to the ANN-MLP models with
351 observed data based on the PM method. The effectiveness of each combination has also been proven by other
352 studies(Yu et al., 2020) .In addition, it seems that all positive combinations contain specific temperature data
353 (Tmax, Tmin) have indicated that Tmax and Tmin are the minimum data required to apply the Standardized PM
354 equation. The results of this study showed obviously the significant role of temperature in the calculation of ET0,
355 such as the combination 04 with the HS method, which is consistent with the results of (Jain et al., 2008).
356 According to this tables, ANN- MLP with 5 inputs combination 02 and ANN-MLP with 3 inputs (Rs,
357 Tmax, Tmin) combination 01 and ANN-RBF with three inputs (Rs, Tmax, Tmin) combination 04 provided the best ET0
358 estimates.
359 Figure 7 shows the comparisons between PM-ET0 and HS-ET0 modeled by the ANN-MLP models with
360 observed and NASA data at three stations ,(a) Mouzia, (b) Soumaa,(c) Boukourdane.
17
361
362 Fig.7: Comparisons between the daily ET0 estimated and calculated from the 04 proposed input combinations
364 This figure shows the comparisons of ET0 estimate by ANN-MLP models using the NASA POWER and
365 observed datasets. As can be observed for each stations the climatic parameters are the most important factors that
366 affect to ET0. Specifically, solar radiation and temperatures. This finding directly accounts for the implicit capacity
18
367 of the ANNs models for physical interpretation,the results showed that the input combination (Tmax, Tmin, U2,
368 Rh, and Rs) provided the best estimates.
369 Figure 8 shows the scatter plots of the estimated and calculated PM-ET0 and HS-ET0 values for the best
370 combinations (ANN-MLP) with observed and NASA datasets for three stations (a) Boukourdane ,(b) Soumaa, (c)
371 Mouzaia.
372
19
373
374 Fig.8: Scatter plots in daily PM-ET0 and HS-ET0 ANN-MLP estimation for three stations (a)
375 Boukourdane ,(b) Soumaa, (c) Mouzaia.
376 What can be seen from the figure 8 is that there is still a small error in PM-ET0 and HS-ET0, even though
377 the estimates derived are extraordinarily good. In fact, what can be inferred from the four satisfactory combinations
378 is that the promising results derived from the ANN-MLP models can be interpreted by the weather combination
379 model rather than by the number of inputs.
380 The ET0 provides an index of the evapotranspiration rate of a reference area, not of a lack of water,
381 independently of crop type, crop development, and management practices (Allen et al., 1998).
382 As finely result the NASA POWER data can be used to estimate ET0, as well as this result obtained by
383 (Jiménez-Jiménez et al., 2021b) and include that the temperature, solar radiation, and relative humidity variables
384 can be obtained from NASA POWER data, but wind speed cannot. The measured ET0 is overestimated by the
385 NASA POWER data, with daily and decadal periods having RMSE of 1.15 and 0.89 mm/day,In Italy ,the results
386 presented by (Negm et al., 2017b) who using other systems with a similar approach,they also obtained the
387 possibility of estimating the daily ET0 on the basis of the POWER-NASA agro-climatology even in other
388 Mediterranean countries where most climate variables are not measured, However in Egypt (Aboelkhair et al.,
389 2019) it appears that NASA POWER reanalysis data can be used when Egypt has few observations. Although this,
390 it needs to be improved to take into account the influence of the Mediterranean location on relative humidity and
391 temperature. Nevertheless, the uncertainty of ET0 by the ANNs models and the sensitivity of the input variables
392 must be assessed in the future.
393 5. Conclusion
394 This research examines the effectiveness of artificial neural network models (ANNs) for estimating the
395 baseline ET0 evapotranspiration, calculated by Standardized Penman-Monteith equation PM-ET0 and Hargreaves-
396 Samani HS-ET0 equations, the POWER-NASA weather variables were validated with real data by statistical
397 analysis linear regression of the daily variables, the analysis shows that in general, the POWER-NASA database
398 tends to slightly overestimate the different climate variables measured on the ground, The comparison between
399 the daily values of four meteorological properties from POWER-NASA and the measurements of the
400 meteorological stations over three stations (Mouzaia, Soumaa, Boukourdane) of Mitidja Algeria with deferent
401 periods, suggests that the air temperatures data (at 2 m above ground) are reliable. In other words, five variants of
402 the ANN-MLP and ANN-RBF models with 04 combinations of input parameters and one ET0 output were
403 developed. The results show that the performance of the five-input ANN-MLP is superior to that of the other model
404 variants in terms of RMSE, MAPE, and R². In developing countries, weather stations are scarce, which implies
405 the difficulty in collecting meteorological parameters such as wind speed, relative humidity, and solar radiation.
406 The results showed that the performance of the three input ANN-MLP models has the same predictive power as
407 the five inputs such as ANN-MLP model combination 01 of three easy-to-find parameters (Rs, Tmax, Tmin) of
408 observed data using the Standardized PM method and the ANN-MLP combination 04 (Td, Tmax, Tmin) of real
409 data using the HS equation. The reference prediction models (ANN-MLP and ANN-RBF) with only three available
410 parameters like temperatures and solar radiation are also adopted, and the use of the Hargreaves-Samani equation
411 is as effective with minimum parameters for the estimation of daily ET0.
412 This approach is useful to estimate ET0 when there is a luck of observed data and where the NASA POWER
413 data can be used to estimate irrigation needs with an acceptable accuracy to improve the irrigation plans and to
414 preserve water and so enlarge the irrigated areas.
415 ACKNOWLEDGEMENTS
416 The study is carried out within the framework of the SWATCH project (Prima project) funded by the
417 DGRSDT, Algeria.The POWER-NASA database was downloaded from the Langley Research Center
418 Atmospheric Science Data Center Surface meteorological and Solar Energy (SSE) web portal supported by the
419 NASA LaRC POWER Project.
420
421
422
423
20
424 References
425 Aboelkhair, H., Morsy, M., & el Afandi, G. (2019). Assessment of agroclimatology NASA POWER
426 reanalysis datasets for temperature types and relative humidity at 2 m against ground
428 https://doi.org/10.1016/J.ASR.2019.03.032
429 Afram, A., Janabi-Sharifi, F., Fung, A. S., & Raahemifar, K. (2017). Artificial neural network (ANN)
430 based model predictive control (MPC) and optimization of HVAC systems: A state of the art
431 review and case study of a residential HVAC system. Energy and Buildings, 141, 96–113.
432 https://doi.org/10.1016/J.ENBUILD.2017.02.012
433 Ahooghalandari, M., Khiadani, M., & Jahromi, M. E. (2016). Developing Equations for Estimating
435 https://doi.org/10.1007/S11269-016-1386-7/METRICS
436 Allen, R. G., Pereira, L. S., Raes, D., & Smith, M. (1998). Crop evapotranspiration-Guidelines for
437 computing crop water requirements-FAO Irrigation and drainage paper 56.
438 Allen, R. G., Pruitt, W. O., Wright, J. L., Howell, T. A., Ventura, F., Snyder, R., Itenfisu, D., Steduto, P.,
439 Berengena, J., Yrisarry, J. B., Smith, M., Pereira, L. S., Raes, D., Perrier, A., Alves, I., Walter, I., &
440 Elliott, R. (2006). A recommendation on standardized surface resistance for hourly calculation
441 of reference ETo by the FAO56 Penman-Monteith method. Agricultural Water Management,
443 Allen, R. G., Walter, I. A., Elliott, R. L., Howell, T. A., Itenfisu, D., Jensen, M. E., & Snyder, R. L. (2005).
444 The ASCE Standardized Reference Evapotranspiration Equation. ASCE Standardized Reference
446 Antonopoulos, V. Z., & Antonopoulos, A. v. (2017). Daily reference evapotranspiration estimates by
447 artificial neural networks technique and empirical equations using limited input climate
21
448 variables. Computers and Electronics in Agriculture, 132, 86–96.
449 https://doi.org/10.1016/J.COMPAG.2016.11.011
450 Bouziane, O., Meddi, M., & Cardena, J. R. (2021). Characterization of the Water Holding Capacity of
451 the Soils in the Mitidja Plain (Algeria) as a Basis for the Development of the Irrigation Water
453 https://doi.org/10.6937/TWC.202112/PP_69(4).0007
454 Celestin, S., Qi, F., Li, R., Yu, T., & Cheng, W. (2020). Evaluation of 32 Simple Equations against the
455 Penman–Monteith Method to Estimate the Reference Evapotranspiration in the Hexi Corridor,
456 Northwest China. Water 2020, Vol. 12, Page 2772, 12(10), 2772.
457 https://doi.org/10.3390/W12102772
458 Crovella, T., Paiano, A., & Lagioia, G. (2022). A meso-level water use assessment in the
459 Mediterranean agriculture. Multiple applications of water footprint for some traditional crops.
461 Deng, Y., Zhou, X., Shen, J., Xiao, G., Hong, H., Lin, H., Wu, F., & Liao, B. Q. (2021). New methods
462 based on back propagation (BP) and radial basis function (RBF) artificial neural networks (ANNs)
463 for predicting the occurrence of haloketones in tap water. The Science of the Total Environment,
465 Dikshit, A., Pradhan, B., & Santosh, M. (2022). Artificial neural networks in drought prediction in the
466 21st century–A scientometric analysis. Applied Soft Computing, 114, 108080.
467 https://doi.org/10.1016/J.ASOC.2021.108080
468 Dou, X., & Yang, Y. (2018). Evapotranspiration estimation using four different machine learning
469 approaches in different terrestrial ecosystems. Computers and Electronics in Agriculture, 148,
22
471 Duarte, Y. C. N., & Sentelhas, P. C. (2020). NASA/POWER and DailyGridded weather datasets—how
472 good they are for estimating maize yields in Brazil? International Journal of Biometeorology,
474 Elemo, E. O., Ogobor, E. A., Ayantunji, B. G., Mangete, O. E., Alagbe, G. A., Abdulkareem, M. L.,
475 Obarolo, A. E., Onuh, B. O., Elemo, E. O., Ogobor, E. A., Ayantunji, B. G., Mangete, O. E., Alagbe,
476 G. A., Abdulkareem, M. L., Obarolo, A. E., & Onuh, B. O. (2021). Relationship between Relative
477 Humidity and the Dew Point Temperature in Abuja, Nigeria. Open Access Library Journal, 8(12),
479 Forster, M. A., Kim, T. D. H., Kunz, S., Abuseif, M., Chulliparambil, V. R., Srichandra, J., & Michael, R.
480 N. (2022). Phenology and canopy conductance limit the accuracy of 20 evapotranspiration
481 models in predicting transpiration. Agricultural and Forest Meteorology, 315, 108824.
482 https://doi.org/10.1016/J.AGRFORMET.2022.108824
483 George H. Hargreaves, & Zohrab A. Samani. (1985). Reference Crop Evapotranspiration from
485 https://doi.org/10.13031/2013.26773
486 Granata, F. (2019). Evapotranspiration evaluation models based on machine learning algorithms—A
488 https://doi.org/10.1016/J.AGWAT.2019.03.015
489 Hargreaves, G. H., & Samani, Z. A. (1982). Estimating Potential Evapotranspiration. Journal of the
491 Hashemi Fath, A., Madanifar, F., & Abbasi, M. (2020). Implementation of multilayer perceptron (MLP)
492 and radial basis function (RBF) neural networks to predict solution gas-oil ratio of crude oil
23
494 Huo, Z., Feng, S., Kang, S., & Dai, X. (2012). Artificial neural network models for reference
495 evapotranspiration in an arid area of northwest China. Journal of Arid Environments, 82, 81–90.
496 https://doi.org/10.1016/J.JARIDENV.2012.01.016
497 Izadifar, Z., & Elshorbagy, A. (2010). Prediction of hourly actual evapotranspiration using neural
498 networks, genetic programming, and statistical models. Hydrological Processes, 24(23), 3413–
500 Jain, S. K., Nayak, P. C., & Sudheer, K. P. (2008). Models for estimating evapotranspiration using
501 artificial neural networks, and their physical interpretation. Hydrological Processes, 22(13),
503 Jiménez-Jiménez, S. I., Ojeda-Bustamante, W., Inzunza-Ibarra, M. A., & Marcial-Pablo, M. de J. (2021).
504 Analysis of the NASA-POWER system for estimating reference evapotranspiration in the
506 https://doi.org/10.5154/R.INAGBI.2021.03.050
507 Laqui, W., Zubieta, R., Rau, P., Mejía, A., Lavado, W., & Ingol, E. (2019). Can artificial neural networks
508 estimate potential evapotranspiration in Peruvian highlands? Modeling Earth Systems and
510 le Page, M., Fakir, Y., & Aouissi, J. (2020). Modeling for integrated water resources management in
511 the Mediterranean region. Water Resources in the Mediterranean Region, 157–190.
512 https://doi.org/10.1016/B978-0-12-818086-0.00007-8
513 Marzouk, O. A. (2021). Assessment of global warming in Al Buraimi, sultanate of Oman based on
514 statistical analysis of NASA POWER data over 39 years, and testing the reliability of NASA
516 https://doi.org/10.1016/J.HELIYON.2021.E06625
24
517 Maselli, F., Battista, P., Chiesi, M., Rapi, B., Angeli, L., Fibbi, L., Magno, R., & Gozzini, B. (2020). Use of
518 Sentinel-2 MSI data to monitor crop irrigation in Mediterranean areas. International Journal of
520 https://doi.org/10.1016/J.JAG.2020.102216
521 McCulloch, W. S., & Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity. The
523 https://doi.org/10.1007/BF02478259/METRICS
524 Mohammadi, B., & Mehdizadeh, S. (2020). Modeling daily reference evapotranspiration via a novel
525 approach based on support vector regression coupled with whale optimization algorithm.
527 Moustris, K. P., Larissi, I. K., Nastos, P. T., & Paliatsos, A. G. (2011). Precipitation Forecast Using
528 Artificial Neural Networks in Specific Regions of Greece. Water Resources Management, 25(8),
530 Najmaddin, P. M., Whelan, M. J., & Balzter, H. (2017). Estimating daily reference evapotranspiration
531 in a semi-arid region using remote sensing data. Remote Sensing, 9(8).
532 https://doi.org/10.3390/RS9080779
533 NASA POWER | Prediction Of Worldwide Energy Resources. (n.d.). Retrieved February 6, 2023, from
534 https://power.larc.nasa.gov/
535 Negm, A., Jabro, J., & Provenzano, G. (2017). Assessing the suitability of American National
536 Aeronautics and Space Administration (NASA) agro-climatology archive to predict daily
537 meteorological variables and reference evapotranspiration in Sicily, Italy. Agricultural and
539 Negm, A., Minacapilli, M., & Provenzano, G. (2018). Downscaling of American National Aeronautics
540 and Space Administration (NASA) daily air temperature in Sicily, Italy, and effects on crop
25
541 reference evapotranspiration. Agricultural Water Management, 209, 151–162.
542 https://doi.org/10.1016/J.AGWAT.2018.07.016
543 Ouassanouan, Y., Fakir, Y., Simonneaux, V., Kharrou, M. H., Bouimouass, H., Najar, I., Benrhanem, M.,
544 Sguir, F., & Chehbouni, A. (2022). Multi-decadal analysis of water resources and agricultural
545 change in a Mediterranean semiarid irrigated piedmont under water scarcity and human
547 https://doi.org/10.1016/J.SCITOTENV.2022.155328
548 Özgür, A., & Yamaç, S. S. (2020). Modelling of daily reference evapotranspiration using deep neural
550 PENMAN, H. L. (1948). Natural evaporation from open water, bare soil and grass. Proceedings of the
551 Royal Society of London. Series A. Mathematical and Physical Sciences, 193(1032), 120–145.
552 https://doi.org/10.1098/RSPA.1948.0037
553 Qian, L., Wu, L., Liu, X., Cui, Y., & Wang, Y. (2022). Comparison of CLDAS and Machine Learning
554 Models for Reference Evapotranspiration Estimation under Limited Meteorological Data.
555 Sustainability 2022, Vol. 14, Page 14577, 14(21), 14577. https://doi.org/10.3390/SU142114577
556 Reis, M. M., da Silva, A. J., Zullo Junior, J., Tuffi Santos, L. D., Azevedo, A. M., & Lopes, É. M. G. (2019).
557 Empirical and learning machine approaches to estimating reference evapotranspiration based
559 https://doi.org/10.1016/J.COMPAG.2019.104937
560 Rodrigues, G. C., & Braga, R. P. (2021). Evaluation of nasa power reanalysis products to estimate daily
562 https://doi.org/10.3390/AGRONOMY11061207
563 Snyder, R. L., Specialist, B., & Eching, S. (n.d.). Penman-Monteith daily (24-hour) Reference
26
565 Sowmya, M. R., Santosh Kumar, M. B., & Ambat, S. K. (2020). Comparison of deep neural networks
566 for reference evapotranspiration prediction using minimal meteorological data. Proceedings -
567 2020 Advanced Computing and Communication Technologies for High Performance
569 Stewart, J. B. (1984). Measurement and Prediction of Evaporation from Forested and Agricultural
571 Walter, I. A., Allen, R. G., Elliott, R., Jensen, M. E., Itenfisu, D., Mecham, B., Howell, T. A., Snyder, R.,
572 Brown, P., Echings, S., Spofford, T., Hattendorf, M., Cuenca, R. H., Wright, J. L., & Martin, D.
575 White, J. W., Hoogenboom, G., Stackhouse, P. W., & Hoell, J. M. (2008). Evaluation of NASA satellite-
576 and assimilation model-derived long-term daily temperature data over the continental US.
578 https://doi.org/10.1016/J.AGRFORMET.2008.05.017
579 White, J. W., Hoogenboom, G., Wilkens, P. W., Stackhouse, P. W., & Hoel, J. M. (2011). Evaluation of
580 Satellite-Based, Modeled-Derived Daily Solar Radiation Data for the Continental United States.
582 Yassin, M. A., Alazba, A. A., & Mattar, M. A. (2016). Modelling daily evapotranspiration using artificial
583 neural networks under hyper arid conditions. Pakistan Journal of Agricultural Sciences, 53(3),
585 Yu, H., Wen, X., Li, B., Yang, Z., Wu, M., & Ma, Y. (2020). Uncertainty analysis of artificial intelligence
586 modeling daily reference evapotranspiration in the northwest end of China. Computers and
27
588
589
28