Machine learning approaches for enhanced estimation of reference evapotranspiration (ETo): a comparative evaluation

Abstract

Abstract Accurate estimation of reference evapotranspiration (ETo) is critical for effective water resource management, particularly in regions with limited meteorological data. However, existing empirical and deep learning models often require extensive data or complex modeling, limiting their practical application in data-scarce environments. This study innovatively applies static (non-sequential) machine learning models K-Nearest Neighbors (KNN), Decision Tree (DT), and Random Forest (RF) justified by temporal dependency analysis, to estimate daily ETo using varying input scenarios, from full-feature datasets to minimal single-variable inputs. Results show that RF outperforms other models, achieving a root mean square error (RMSE) of 0.52 mm/day and a coefficient of determination (R²) of 0.96, with temperature and solar radiation identified as key predictors. These findings highlight the practicality of RF for robust and efficient ETo estimation, offering a reliable tool for water management and agricultural planning in resource-constrained settings.