Elucidating Reference Evapotranspiration Drivers in Contrasting Climates using Machine Learning: From Advection-Driven to Energy-Limited Processes

Accurate estimation of reference evapotranspiration (ET₀) is fundamental to sustainable water resource management, particularly in regions characterized by high climatic heterogeneity. This study evaluates the performance of machine learning models across three distinct climatic regimes in Sistan and Baluchestan Province, Iran: hyper-arid (Zabol), semi-arid mountainous (Zahedan), and coastal-humid (Chabahar). Using daily meteorological data, the predictive capabilities of Random Forest (RF) and Support Vector Machine (SVM) models were benchmarked against the classical Multiple Linear Regression (MLR) under three input scenarios: temperature-based, temperature-humidity, and full-input. The results demonstrate that no single model is universally superior across all climatic conditions. The Random Forest (RF) model exhibited the highest precision in the extreme climates of hyper-arid Zabol (R²=0.99, RMSE = 0.46 mm/day) and coastal Chabahar (R²=0.88), whereas the Support Vector Machine (SVM) was superior in the stable, semi-arid conditions of Zahedan (R²=0.98, RMSE = 0.31 mm/day). Feature importance analysis revealed fundamental divergences in the governing physical processes: the ET₀ process in Zabol is primarily “advection-dominated,” driven by wind-speed dynamics; in contrast, the Chabahar regime is “energy-limited,” with maximum temperature as the primary controlling factor. These findings underscore the necessity of tailoring ET₀ modeling strategies to regional climatic drivers. This research provides a robust framework for enhancing the precision of hydrological modeling and water-resource allocation in territories with sharp climatic gradients.