Improving terrestrial evapotranspiration estimation of the Tibetan Plateau by coupling SEBS with machine learning-derived aerodynamic resistance

Abstract

Land evapotranspiration (ET) on the Qinghai–Tibet Plateau (TP) is crucial for regulating worldwide atmospheric circulation. This research explores the integration of machine learning (ML) with a physical framework to improve ET estimation. We developed a coupled model that combines the surface energy balance system (SEBS) model with ML algorithms (SEBS-ML) to estimate ET effectively. Specifically, we employed the random forest (RF) algorithm to estimate aerodynamic resistance (ra), which significantly influences the turbulent transport between the surface and air. We evaluated the instantaneous and daily ET estimates using data from 17 eddy covariance flux tower sites and compared these estimates with those derived from three other ML strategies: linear regression (LR), ridge regression (RR), and support vector machine regression (SVM). We also compared the hybrid model with the SEBS model and pure machine learning strategy (PML). The results indicate that RF provides the most accurate estimates of daily ET among the 4 hybrid models, with R² and Kling–Gupta efficiency (KGE) values equal to 0.70 and 0.82. SVM performed less effectively than RF, while LR and RR were the least effective. The ensemble learning approach in the RF model appears to reduce the bias of the ensemble results by compensating for individual tree biases. Under extreme conditions, the hybrid model demonstrates superior generalization capability, with a relatively low energy irrationality rate and better extrapolation performance compared to the PML model.. The enhanced strategy enhances the quality of ET estimates but also adheres to physical constraints, thereby preventing the generation of implausible results. This research introduces an innovative approach for estimating ET that enhances the physical mechanisms and performance of the SEBS model, thereby improving its precision and scalability. This model provides vital insights into the hydrological and climatic changes occurring on the TP.