Adaptive meta-modeling of evapotranspiration in arid agricultural regions of Saudi Arabia using climatic factors, drought indices and MODIS data

Study Region: The research was conducted in three arid agricultural regions of Saudi Arabia: Wadi Ad-Dawasir, Ranya, and Abha. Study Focus: The objective is to develop an intelligent approach that utilizes ET-AI (Release 1, developed by Osama Elsherbiny), an accessible and user-friendly software, to compute actual evapotranspiration (AET) with both speed and accuracy. This can enhance irrigation efficiency and optimize water resource management. The data collected from 2000 to 2023 encompass four environmental factors (EF), two drought indices (DI), and six MODIS spectral indices (SI). Machine learning models, including the backpropagation neural network (BPNN) and XGBoost Regressor (XGB), enhanced with an adaptive meta-model (AMM) strategy, were evaluated for predicting monthly AET. The best-performing model was designated based on statistical metrics, with a focus on minimizing discrepancies between predicted and actual AET values. New Hydrological Insights for the Region: The results revealed robust correlations between AET and the combination of climatic water deficit (Def) with minimum (Tmin) or maximum temperature (Tmax), with R² values ranging from 0.70–0.67 in Wadi Ad-Dawasir, 0.59–0.57 in Ranya, 0.80–0.77 in Abha, and 0.64–0.55 for combined data. These findings highlight the regional sensitivity of AET to temperature and water deficit variations, offering valuable insights for water management strategies. Furthermore, the study reveals distinct spatial patterns of evapotranspiration dynamics across the region, which are crucial for improving irrigation practices under varying climate conditions. The BPNN-AMM model, deploying eleven EF-DI-SI features, delivered superior performance (R²=0.914, RMSE=6.115) compared to the standalone BPNN model (R²=0.850, RMSE=7.289). It also outperformed the XGB-AMM model with seven hybrid traits (R²=0.869, RMSE=9.285), in contrast to the separate XGB model (R²=0.684, RMSE=10.584). By refining the precision of AET predictions, the model clarifies water balance processes in arid regions. These insights have the potential to guide regional water resource management and enable real-time AET monitoring. The developed software is available for access at (https://drive.google.com/file/d/1dPMiDzngtzDIY65xIyo8MAIVbJMudeBc).