EVAPOTRANSPIRATION ESTIMATION USING EMPIRICAL AND MACHINE LEARNING MODELS A COMPARATIVE ASSESSMENT OF ACCURACY AND TRANSFERABILITY

Abstract

This study compared empirical and machine learning (ML) models for estimating reference evapotranspiration (ET₀) at two agroclimatically contrasting stations in Kerala: RARS Pattambi (inland, semi-arid) and the College of Agriculture, Vellayani (coastal, humid). Using thirty years (1995–2025) of daily meteorological data, the FAO-56 Penman–Monteith equation served as the benchmark. Five ML models - CNN, LSTM, CNN-LSTM,XGBoost, and Random Forest—were developed under four input configurations (full variable, temperature-based, mass transfer-based, and radiation-based) and compared against six empirical models, with accuracy assessed via R², RMSE, and MAE, and spatial transferability quantified through a transferability score. ML models consistently outperformed empirical approaches across all input categories.With complete inputs, all ML models achieved R² > 0.95, with XGBoost RF performing best. Under reduced inputs, radiation-based models performed strongest, with LSTM and XGBoost exceeding an R² of 0.87, while the empirical Makkink model also performed comparatively well, reflecting the strong physical link between solar radiation and ET₀ in Kerala's climate. Temperature-based models performed weakest overall. Transferability analysis showed that models trained at Pattambi transferred more effectively to Vellayani than the reverse, particularly for radiation-based inputs, where external validation exceeded internal accuracy. Mass transfer-based models exhibited the weakest cross-station transferability, reflecting differing humidity dynamics between the inland and coastal environments. Validation using 2025 data confirmed these patterns. Overall, ML models - particularly XGBoost and Random Forest with complete data, and LSTM/CNN-LSTM under radiation-based inputs - offer a reliable and transferable framework for ET₀ estimation. Radiation-based inputs are recommended for data-scarce settings, and the demonstrated inland-to-coastal transferability supports extending these models to ungauged locations, providing a scientific foundation for improved irrigation planning and water resource management across Kerala.