Which Machine Learning Algorithm Is Best Suited for Estimating Reference Evapotranspiration in Humid Subtropical Climate?

Abstract

Timely and reliable estimates of reference evapotranspiration (ET₀) are imperative for robust water resources planning and management. Applying machine learning (ML) algorithms for estimating ET₀ has been evolving, and their applicability in different sectors is still a compelling field of research. In this study, four Gaussian process regression (GPR) algorithms—polynomial kernel (PK), polynomial universal function kernel (PUK), normalized poly kernel (NPK), and radial basis function (RBF)—were compared against widely used random forest (RF) and a simpler locally weighted linear regression (LWLR) algorithm at a humid subtropical region in India. The sensitivity analysis of the input variables was followed by application of the best combination of variables in algorithm testing and training for generating ET₀. The results were then compared against the Penman–Monteith method at both daily and monthly time steps. The results indicated that ET₀ is least sensitive to wind speed at 2 m height. Additionally, at a daily time step, RF, followed by PUK, generated the best results during both training and testing phases. In contrast, at a monthly time step, using multiple model evaluation matrices, PUK followed by RF performed best. These results demonstrate the application of the ML algorithms is subjected to user-required time steps. Although this study focused on Northwest India, the findings are relevant to all humid subtropical regions across the world.