Investigating the ability of deep learning on actual evapotranspiration estimation in the scarcely observed region

Abstract

The estimation of actual evapotranspiration is a difficult issue in hydrological research, particularly in the scarcely observed region. Deep learning (DL) has been increasingly used in the field of hydrology in recent years. In this study, we investigated the ability of DL on actual evapotranspiration estimation using three sets of controlled experiments at a typical region with scarce observations, the Qinghai-Tibetan Plateau. The results suggest that the DL model can utilize a few key types of observation data to simulate actual evapotranspiration, and more in-situ observation data types did not significantly improve the accuracy of DL simulations. A multi-source DL model established by integrating data from distantly distributed stations showed a better performance than the model built separately using data at individual sites. Moreover, further analysis of climate pattern and input data correlation was conducted for the similarity of stations for multi-source learning of the DL model. Using inputs in the lead-time period can improve the simulation of daily ET by DL. Compared with traditional process-based physical methods, the DL model is more flexible to simulate actual evapotranspiration in areas with insufficient observed data such as the Qinghai-Tibetan Plateau. The results of this study highlight the potential power of DL model to improve the actual evapotranspiration estimation in the scarcely observed region.