Comparative performance of random forest and support vector regression for reference evapotranspiration estimation: a critical review

Abstract

Precision irrigation has emerged as a critical adaptation strategy for alleviating global water scarcity under climate change, relying heavily on the accurate estimation of reference evapotranspiration (ET 0 ). Although physically based models, such as the FAO-56 Penman-Monteith equation, serve as the global standard, their reliance on extensive and continuous meteorological data significantly limits their application in data-scarce and heterogeneous regions. Machine Learning (ML) offers a promising alternative to overcome these data limitations. However, the trade-offs between model complexity and adaptability remain to be fully elucidated. This systematic review traces the evolution of precision irrigation from reactive scheduling to proactive predictive control. It focuses on the comparative performance of support vector regression (SVR) and random forest (RF) models for ET 0 and evapotranspiration-related prediction tasks. Specifically, it benchmarks the predictive capabilities of both algorithms across diverse climatic conditions and varying data availability. Model performance was interpreted using R 2 together with RMSE and study-design information, including target-variable definition, predictor availability, and validation strategy. Comparative evaluations suggest that apparent model superiority is conditional on data dimensionality, predictor availability, target construction, and validation strategy. Current literature suggests that SVR may be competitive in limited-input, low-dimensional, or relatively stable settings, whereas RF/RFR may be advantageous for heterogeneous, multi-source, or spatially distributed predictor sets. These tendencies should be interpreted as conditional patterns rather than universal model-selection rules, because reported performance is strongly shaped by target construction, predictor availability, sample size, climate setting, and validation design.