Comparison of physical-based, data-driven and hybrid modeling approaches for evapotranspiration estimation

Abstract

Obtaining accurate estimates of evapotranspiration is a key issue in many fields like hydrology, ecology, and agriculture. There exist different ways, including physical-based and data-driven approaches, to estimate evapotranspiration. Here, the physical-based model (the surface energy balance system, SEBS), data-driven models (using three machine learning techniques, deep neural network (DNN), random forest (RF), and symbolic regression (SR)), and hybrid model were compared using the FLUXNET2015 dataset. More importantly, different search strategies were investigated to optimize the network architecture and to explore the maximum accuracy of data-driven models. The Shapley additive explanations (SHAP) was introduced to quantify the contributions of input features to evapotranspiration estimation. The results show that there is a large error in the SEBS model with RMSE of 121 W m−2, while the data-driven models acquire the best evapotranspiration estimation with RMSE of 32–53 W m−2. The hybrid approach, whose RMSE is 60 W m−2, improves the SEBS model but still underperforms data-driven models. The complexity of data-driven models influences the accuracy of evapotranspiration estimations. By optimizing the network architecture, DNN and RF obtain results with similar precision. But the performance of SR, which generates a simple algebraic formula for evapotranspiration estimation, is significantly degraded. SHAP reveals that all models regard net radiation as the most crucial feature. The SEBS and hybrid models attach much importance to temperature and humidity. Three data-driven models learn the different relationships between input features and evapotranspiration. The DNN-based model is likely to learn a relatively correct relationship among the data-driven models because it has a similar feature importances pattern with physical-based and hybrid models.