Estimation of daily reference evapotranspiration implementing satellite image data and strategy of ensemble optimization algorithm of stochastic gradient descent with multilayer perceptron

Water resource management relies on accurate estimations of evapotranspiration (ET0). Moreover, the use of satellite images is often an effective method of compensating for the lack of reliable weather data. It is typically necessary to interpolate remotely sensed vegetation indexes into daily scales when estimating daily ET0 from remote sensing data. Also, it is essential to accurately correct missing data when using remote sensing land surface temperature parameters to estimate daily ET0. The study utilized stochastic gradient descent (SGD) for the optimization of multilayer perceptron (MLP) and developed a hybrid model (SGD-MLP) to estimate daily ET0 in Tabriz, Iran croplands using remote sensing and data from meteorological stations. Using the FAO56 Penman–Monteith equation (FAO56-PM), estimated ET0 values were compared with determined ET0 values from data collected between 2003 and 2021. In order to develop this model, 70% of the data spanning the period was utilized for training, and 30% was used for testing. Moreover, the daily ET0 was estimated based on input variables, such as remote sensing parameters and data from meteorological stations. Also, satellites Aqua (MYD11A1) and Terra (MOD11A1) were used to download LST 1-day using thermal infrared bands. Additionally, the Kalman filter and moving average were implemented to reconstruct missing LST data. Also, data of LAI 4-days and NDVI 16-days were converted into LAI and NDVI 1-days using four different methods of cubic spline, spline, cubic bézier, and fixed intervals of 4 days. The results showed that the NDVI and LAI reconstruction method with cubic bézier and the LST with Kalman filter were the most effective reconstruction methods for both MLP and SGD-MLP models. Moreover, the SGD-MLP model had better performance in all cases compared to the standalone MLP. Conclusively, SGD-MLP-11 with R2 of 0.992, RMSE of 0.165, and MAPE of 11.976 in testing phases, and R2 of 0.992, RMSE of 0.160, and MAPE of 9.232 in training phases produced the most accurate estimates and may be recommended for ET0 estimations in other stations with similar climates.