Integrating machine learning with analytical surface energy balance model improved terrestrial evaporation through biophysical regulation

Abstract

Global evaporation modeling faces challenges in understanding the combined biophysical controls imposed by aerodynamic and canopy-surface conductance, particularly in water-scarce environments. We addressed this by integrating a machine learning (ML) model estimating surface relative humidity (RH0) into an analytical model (Surface Temperature Initiated Closure - STIC), creating a hybrid model called HSTIC. This approach significantly enhanced the accuracy of modeling water stress and conductance regulation. Our results, based on the FLUXNET2015 dataset, showed that ML-RH0 markedly improved the precision of surface water stress variations. HSTIC performed well in reproducing latent and sensible heat fluxes on both half-hourly/hourly and daily scales. Notably, HSTIC surpassed the analytical STIC model, particularly in dry conditions, owing to its more precise simulation of canopy-surface conductance (gSurf) response to water stress. Our findings suggest that HSTIC gSurf can effectively capture physiological trait variations across ecosystems, reflecting the eco-evolutionary optimality of plants. This provides a fresh perspective for process-based models in simulating terrestrial evaporation.