Partitioning Ecosystem Water Fluxes Into Transpiration, Surface Evaporation, and Canopy-Intercepted Evaporation Using Knowledge-Guided Machine Learning at NEON Sites

This study introduces KG-DecompNet, a knowledge-guided machine learning framework developed to partition total evapotranspiration (ET) into its primary components: transpiration (T), surface evaporation (E_s), and canopy-intercepted evaporation (E_i). Traditional approaches have faced challenges to separate ET components, especially the dynamic, threshold-based behavior of Ei, leading to likely overestimation of T following rainfall or dew events. KG-DecompNet addresses this by integrating physical constraints into site-level machine learning models trained on multi-year, high-frequency turbulence and meteorological data from 35 National Ecological Observatory Network sites. The models achieve over 90% agreement with conditional eddy accumulation-derived T and Es during periods when Ei is likely trivial, and remain robust when compared with flux variance similarity (FVS)-derived estimates. By isolating Ei, KG-DecompNet offers new insights into surface-atmosphere water exchanges and helps set a benchmark for physically grounded ecohydrological modeling.