Jonathan Frame , Grey Stephen Nearing, Frederik Kratzert, Mashrekur Rahman 

We build three Long Short-Term Memory (LSTM) daily streamflow prediction models (deep learning networks) for 531 basins across the contiguous United States (CONUS), and compare their performance: (1) a LSTM post-processor trained on the U.S. National Water Model (NWM) outputs (LSTM_PP) as a target variable, (2) a LSTM post-processor trained on the NWM outputs and using atmospheric forcings (LSTM_PPA), and (3) a LSTM model trained on USGS average daily streamflow data and using atmospheric forcing (LSTM_A). We trained the LSTMs for the period 2004-2014 and evaluated on 1994-2002, and compared several performance metrics to the NWM reanalysis. Overall performance of the three LSTMs is similar, with median NSE scores of 0.73 (LSTM_PP), 0.75 (LSTM_PPA), and 0.74 (LSTM_A), and all three LSTMs outperform the NWM validation scores of 0.62. Additionally, LSTM_A outperforms LSTM_PP and LSTM_PPA in ungauged basins. While LSTM as a post-processor improves NWM predictions substantially, we achieved comparable performance with the LSTM trained without the NWM outputs (LSTM_A). Finally, we performed a sensitivity analysis to diagnose the land surface component of the NWM as the source of mass bias error and the channel router as a source of simulation timing error. This indicates that the NWM routing scheme should be considered a priority for NWM improvement.