ABSTRACT Terrestrial evapotranspiration (ET) is crucial for the sustainable assessment of water resources and ecosystem functions. While machine learning (ML) models show promise in ET estimation using flux tower data (FLUXNET), existing ML application approaches inadequately address systematic errors caused by turbulence hysteresis (time‐lag) effects in eddy covariance systems. Leveraging data from 103 global flux towers, this study introduces a novel approach integrating time‐lagged meteorological variables (e.g., lag = 1–3 days) to enhance ET modelling precision. Validated at global flux tower sites, the use of ML algorithms such as random forest and gradient boosting networks reduced model application bias and significantly improved the model's coefficient of determination; the model achieved an R 2 increase of 39.7% (from 0.63 to 0.88) and reduced MAE by 45.6% (from 12.7 to 6.9 W/m 2 ) compared with benchmark models. Model interpretation (SHAP/Causal Forest Algorithm) confirmed that lagged variables contributed comparably (±15%) to concurrent variables in ET estimation. Collectively, the notable enhancements and novel methodological perspectives in this study may serve as a reference for improving the global modelling of evapotranspiration, thereby further optimizing global water and ecosystem monitoring under climate change.
Shu et al. (Thu,) studied this question.