Solar Radiation–Driven Machine Learning for Modelling Reference Evapotranspiration Using <scp>ERA5</scp> ‐Land in a Semi‐Arid Microclimatic Basin in Türkiye

ABSTRACT Accurate estimation of reference evapotranspiration (ETo) is critical for effective water resource management, irrigation planning, and climate adaptation, particularly in semi‐arid regions where water scarcity and climatic variability are pronounced. This study assessed the performance and temporal stability of four machine learning (ML) algorithms [Multilayer Perceptron (MLP), Random SubSpace (RSS), M5P model tree (M5P), and Random Forest (RF)] in estimating daily ET in the Iğdır Basin, one of eastern Türkiye's most productive yet hydrologically vulnerable agricultural zones. Using ERA5‐Land reanalysis data, the models were trained and validated across two distinct periods (1950–1999 and 2000–2024) to evaluate both historical accuracy and robustness under changing climatic conditions. Solar radiation (SR) was identified as the primary driver of ETo, highlighting the dominant role of radiative energy in modulating evapotranspiration variability, even under the predominantly water‐limited conditions of the basin's semi‐arid environment. A near‐perfect correlation between ETo and SR confirmed the dominant influence of solar energy on evaporative demand. Among the tested models, M5P and MLP achieved the highest predictive accuracy and maintained consistent performance across both periods, indicating strong generalizability and resistance to climatic non‐stationarity. In contrast, RS and RF exhibited higher error rates and reduced sensitivity to nonlinear interactions, limiting their applicability in dynamic environmental conditions. These results emphasise the ecological and hydrological relevance of energy‐based modelling approaches for drought‐sensitive regions. The integration of ERA5‐Land data with machine learning provides a scalable and data‐efficient alternative to conventional ETo estimation methods, particularly where in situ measurements are limited. The proven accuracy and stability of the M5P and MLP models offer a reliable foundation for long‐term hydrological forecasting, adaptive irrigation planning, and climate resilience efforts in eastern Türkiye's agricultural landscapes.