Abhinav Gupta, Sean A. McKenna

This study examines streamflow simulations using deep learning (DL) to: (1) Understand why global DL models trained on multiple watersheds outperform local DL models trained on single watersheds, given the watershed uniqueness hypothesis and (2) Improve recession flow simulation accuracy. It introduces a novel global-local (GL) modeling strategy, where global model outputs are fed as input to a locally trained model, with the hypothesis that this strategy can leverage both global and watershed-specific information. GL models demonstrate enhanced accuracy in recession flow prediction for 30% of the watersheds compared to global and local models. However, considering the entire hydrograph, GL models often perform worse than the global model. Our results suggest that watershed uniqueness play a significant role in the performance of global models, suggesting that even global LSTM models should be tailored to individual watersheds.