Improving time upscaling of instantaneous evapotranspiration based on machine learning models

Abstract

ABSTRACT Evapotranspiration (ET) plays a crucial role in the global water and energy cycle. Upscaling instantaneous ET ( E T i ) to daily ET ( E T d ) is vital for thermal-based ET estimation. Conventional methods – such as the constant evaporative fraction method (ConEF), radiation-based method, and evaporative ratio method – often overlook environmental factors, leading to biased estimates of E T d from E T i . To resolve this issue, this study aimed to assess four machine learning (ML) algorithms—XGBoost, LightGBM, AdaBoost, and Random Forest—to integrate meteorological and remote sensing data for upscaling E T i across 88 global flux sites. Each ML model was tested with eight different variable combinations. Results indicated that XGBoost exhibited the best performance, with a root mean square error (RMSE) generally below 13 W m − 2 in estimating E T d from E T i . The best variable combination simultaneously considers evaporative fraction, available energy, meteorology factors, remote sensing albedo, normalized vegetation index, and leaf area index. Using this combination, the XGBoost model achieved an R 2 = 0.88 and an RMSE = 12.33 W m − 2 , outperforming the ConEF method ( R 2 = 0.71 and RMSE = 18.86 W m − 2 ) and its expansions. These findings support the application of ML models in ET upscaling, enabling ET estimation across large spatiotemporal scales.