A comparative evaluation of empirical equations and machine learning algorithms for daily reference evapotranspiration estimation in the Eastern Ghats region of India

Accurate estimation of reference evapotranspiration (ET0) is essential for sustainable water resource management in the climatically varied Eastern Ghats region of India. In the present study, eleven empirical methods and four models for machine learning, namely artificial neural network (ANN), k-nearest neighbour (KNN), decision tree (DT) and multiple linear regression (MLR) were assessed with 23 years of meteorological data (2002–2024) of Koraput, Odisha. The Penman–Monteith (PM) method was used as the reference standard. Empirical methods were compared with PM. Artificial intelligence and machine learning (AI/ML) models were developed using input-variable selection approaches such as regression-based selection, principal component analysis, path analysis and cluster analysis. The mean annual ET0 was 1186.49 mm and daily ET0 values ranged from 1.18 to 8.05 mm day−1. Empirical approaches were dominated by the modified PM and Priestley–Taylor, with the Oudin method being characterised by underestimation. The AI/ML models improved ET0 prediction by learning the nonlinear relationships between the meteorological variables. The ANN had the highest accuracy (R2 = 0.957, root mean square error = 0.264, mean absolute percentage error = 5.879%, E = 0.936 and D = 0.983), followed by KNN, MLR and DT. Coupling feature selection with AI/ML, especially ANN, provides a strong base for reliable ET0 estimation for sustainable agricultural water management under conditions of limited meteorological data availability.