Estimation of agricultural vegetation water stress using feature-optimized remote sensing and ensemble machine learning

Abstract

Agricultural production in semi-arid irrigated regions is increasingly vulnerable to episodic water stress, creating a need for reliable satellite-based assessment of vegetation condition. This study develops a feature-optimized remote-sensing and artificial-intelligence framework to estimate vegetation water stress, represented by the MODIS Enhanced Vegetation Index (EVI), in Dakahliyah Governorate, Egypt’s Nile Delta. Eight years of MODIS products (2018–2025) were compiled, including Leaf Area Index (LAI), Fraction of Photosynthetically Active Radiation (Fpar), Evaporative Stress Index (ESI), Normalized Difference Vegetation Index (NDVI), and Land Surface Temperature (LST) as candidate predictors. Data were partitioned into training (2018–2023) and independent testing (2024–2025) sets, with 5-fold cross-validation for robustness. Best Subset Regression evaluated all predictor combinations using adjusted R², Akaike and Bayesian information criteria, Mallows’ Cp, and variance inflation factor screening. The selected subset was used to compare multilayer perceptron, random subspace, regression by discretization, and random forest, with additional benchmarking against XGBoost, CatBoost, and two linear-regression baselines. NDVI showed a strong association with EVI (r = 0.934) and dominated the multivariate model (standardized coefficient = 0.935), while LAI provided a smaller but significant contribution (0.017). After feature optimization, random forest achieved the best generalization, with CC = 0.9943, MAE = 0.0063, and RMSE = 0.0161 during the 2024–2025 test period. These results demonstrate that physically interpretable feature selection combined with tree-based ensembles can provide accurate EVI estimates that generalize to independent years and support vegetation-condition monitoring in semi-arid irrigated agroecosystems.