Abstract
Accurate estimation of reference evapotranspiration (ETo) is crucial for irrigation scheduling, hydrological modeling, and water resource management, particularly in arid coastal regions where microclimatic variability hinders accurate predictions. This study introduces two hybrid frameworks—Variational Mode Decomposition–Long Short-Term Memory–eXtreme Gradient Boosting–Categorical Boosting (VLXCat) and Ensemble Empirical Mode Decomposition–Long Short-Term Memory–eXtreme Gradient Boosting–Categorical Boosting (ELXCat) to improve daily ETo prediction in Muscat, Oman. Meteorological data from 2018 to 2025 were used for model training and testing across three input scenarios: (i) temperature only, (ii) temperature with wind speed (U2), and (iii) temperature with vapor pressure deficit (es–ea). The dual boosting mechanism integrates XGBoost and CatBoost, where the former enhances bias–variance optimization through gradient-based regularization, and the latter refines residual errors and categorical interactions, resulting in complementary learning that strengthens generalization. Model performance was evaluated using R2, RSR, and RMSE. with interpretability analyses based on saliency maps. VLXCat consistently outperformed ELXCat and the benchmark LSTM across all scenarios, achieving optimal accuracy in Scenario 3 (R2 = 0.999, RSR= 0.038, RMSE = 0.139 mm/day). The results demonstrate that integrating signal decomposition with deep learning and dual boosting yields a robust and interpretable approach for ETo prediction. The proposed framework is a scalable solution for climate-responsive irrigation management and drought risk mitigation in arid regions with limited data availability.
