Hydrograph and recession flows simulations using deep learning: Watershed uniqueness and objective functions

Abstract

This study examines streamflow simulations using deep learning (DL) to understand the information extraction capability of global DL models trained on multiple watersheds. The study separately examined the entire streamflow time series and recession flow predictions. It introduces a global–local (GL) modeling strategy, where the global model outputs are fed as input to a locally trained model, with the hypothesis that the local model can leverage watershed-specific information that the global model may miss. The GL models demonstrate enhanced accuracy in recession flow prediction for 20-30% of the watersheds compared to the global and local models. However, considering the entire hydrograph, the GL models often perform worse than the global model. Further, the DL models were trained on two different objective functions. The performance of the global model in a watershed depended strongly upon the objective function used. These results suggest that the performance of global models is affected by watershed uniqueness, suggesting that even a global DL model should be tailored to individual watersheds for optimal performance.