Integrating machine learning with thermal-driven analytical energy balance model improved terrestrial evapotranspiration estimation through enhanced surface conductance
Authors
Bai Y., Mallick K., Hu T., Zhang S., Yang S., Ahmadi A.
Reference
Remote Sensing of Environment, vol. 311, art. no. 114308, 2024
Description
Global evapotranspiration modeling faces significant challenges in understanding the complex interplay between aerodynamic and canopy-surface conductance, especially in water-scarce environments. To address this issue, we developed a hybrid model called HSTIC by integrating a machine learning (ML) model for estimating surface relative humidity (RH0) into the analytically-driven Surface Temperature Initiated Closure (STIC) model, which is based on thermal remote sensing. The ML-RH0 model utilizes the extreme gradient boosting (XGBoost) algorithm and was trained with meteorological and multispectral and thermal infrared remote sensing data. This hybrid approach combines the strengths of ML in estimating RH0 with the analytical framework of STIC. Evaluations were conducted based on observed meteorological factors and heat fluxes from 200 flux sites in the FLUXNET2015 dataset, alongside MODIS remote sensing data. The results demonstrated that including the ML-RH0 significantly improved the accuracy of estimating surface water stress variations. HSTIC performed well in reproducing latent and sensible heat fluxes (λE and H). Under experimental conditions, utilizing tower-retrieved land surface temperature (TR), HSTIC demonstrated a coefficient of determination (R2) ranging from 0.76 to 0.90 and root mean square deviation (RMSD) between 35 and 48 Wm−2 for estimating (half-)hourly λE and H. In practical scenarios using MODIS TR, HSTIC exhibited an R2 of 0.60 to 0.68 and an RMSD of 21 to 26 Wm−2 for daily λE and H when upscaled from instantaneous retrievals at satellite overpass time. Notably, HSTIC surpassed the analytical STIC model, particularly under dry conditions. This improvement is attributed to the more accurate simulation of canopy-surface conductance (gSurf) and its responses to water stress. The enhanced representation of gSurf in HSTIC effectively captures variations in physiological traits across global ecosystems, reflecting how plants adapt to seasonal temperature changes. Our findings suggest that HSTIC provides a fresh perspective for process-based models in simulating terrestrial evapotranspiration. By accurately capturing the dynamic interactions between surface and atmospheric conditions, HSTIC offers a robust tool for understanding and predicting ET under varying environmental conditions.
