Machine learning models for reference evapotranspiration estimation in new locations under data-limited conditions in Northeastern Italy

Reference evapotranspiration (ET0) is a key variable in agro-hydrological applications, yet its estimation is challenged by the high cost of direct measurements and the extensive data requirements of indirect methods. The FAO-56 Penman-Monteith (FAO-56 PM) reference equation's high input requirements often limit its applicability, favoring simpler alternatives such as the Hargreaves equation under data scarcity. Machine learning (ML) models provide a promising pathway to reduce data input requirements while enhancing accuracy and potentially offering spatio-temporal transferability. This study evaluates linear (multiple linear regression, Lasso, Ridge, Elastic Net) and non-linear (regression tree, random forest, XGBoost, support vector regression-SVR, and artificial neural networks-ANN) ML models for ET0 estimation across 16 meteorological stations in the climatically heterogeneous Veneto region (Italy). Daily meteorological data (1994-2022) included air temperature (T), solar radiation (Rs), precipitation, and relative humidity. Model performances were benchmarked against FAO-56 PM equation and a regionally calibrated Hargreaves equation under full and reduced meteorological input configurations and assessed for spatial generalization to unseen locations and temporal transferability to unseen years. SVR, XGBoost, and ANN consistently outperformed all other models across most of the metrics used (i.e. R2 and NSE values near 0.96) in spatial generalization tests across heterogeneous climates. Although temporal transferability showed a modest performance decline, correlation and agreement metrics remained high for all three models. Input variable reduction analyses identified Rs and T as the dominant predictors, with relative humidity and precipitation playing secondary roles. SVR, XGBoost, and ANN matched or outperformed the calibrated Hargreaves equation across all configurations, including the minimal inputs one, based solely on T and extraterrestrial radiation. Overall, the results indicate that, in the study area, SVR, XGBoost, and ANN models provide accurate spatially and temporally transferable ET0 estimates, also under datalimited conditions, representing a reliable alternative for ET0 estimation.