Enhancing streamflow estimation by integrating a data-driven evapotranspiration submodel into process-based hydrological models

Abstract

Hydrological models are powerful tools for estimating streamflow. However, they typically simplify the representation of the evapotranspiration (ET) process, which increases the uncertainty in ET estimation and further introduces extra error in streamflow simulation. This study constrains the uncertainty of ET and improves the accuracy of streamflow estimation by replacing the classical ET module with a data-driven submodel in process-based hydrological models. The data-driven ET submodel was first built based on the FLUXNET2015 dataset and the random forest (RF) algorithm and then integrated into two classic hydrological models with different runoff generation mechanisms, the Xinanjiang (XAJ) model, which is driven by the saturation excess runoff mechanism, and the Soil and Water Assessment Tool (SWAT) model, which is based on the infiltration excess runoff mechanism. The sensitivity of streamflow to ET under the XAJ and SWAT models, as well as the performance of the hybrid hydrological models named the ETRF-XAJ model and the ETRF-SWAT model, was evaluated using the ET and streamflow observations in nine U.S. watersheds. The results showed that streamflow was more sensitive to ET in the XAJ model. The ETRF-XAJ model obviously outperformed the XAJ model, while the ETRF-SWAT and SWAT models had similar accuracies. This similar result was also founded in different climatic zones and seasons. This phenomenon can be attributed to the stronger link between ET and streamflow under saturation excess runoff mechanisms. The ETRF-XAJ model effectively improved the original XAJ model, which underestimated peak flow. This study highlights that the hybrid model has great potential for improved streamflow simulation while building the well-performed hybrid model highly relies on the physical framework of the hydrological model.