Regional scale simulations of daily suspended sediment concentration at gauged and ungauged rivers using deep learning

• LSTM models developed to simulate daily suspended sediment concentration (SSC) in the Mississippi River Basin (MRB). • About 78% and 50% of the studied watersheds had positive NSEs in the gauged and ungauged scenarios, respectively. • VIC-simulated streamflow helped improve SSC simulation, but MMF-simulated sediment yield did not improve the simulations. • LSTM models performed better in the wetter regions of the MRB. • The volume of training data impacted the LSTM performance, underscoring the importance of long-term monitoring data. This study examined the simulations of riverine suspended sediment concentration (SSC) using a deep learning model called Long Short-Term Memory Network (LSTM) and areal averages of daily scale meteorological data and watershed attributes as predictor variables. Observed streamflow, although strongly correlated with SSC, was not used as a predictor since it is not measured for most rivers. The SSC data between 1983 and 2017 from 167 USGS stations in the Mississippi River Basin were used. The LSTM models were trained using two objective functions: normalized mean square errors (NMSE) and normalized mean absolute error (NMAE). The models were tested in gauged and hypothetical ungauged scenarios. In the gauged scenario, the LSTM was trained on the data from all the watersheds; in the ungauged scenario, the LSTM was trained using a leave-location-and-time-out scheme. About 78% and 50% of the watersheds had positive Nash-Sutcliffe Efficiencies (NSEs) in the gauged and ungauged scenarios, respectively. A slight improvement in model performance was achieved when streamflow simulated by the variable infiltration capacity model was used as one of the predictors. The model trained on the NMAE instead of the NMSE objective function improved the simulations of low SSCs in several watersheds. The model performance was poor in some of the watersheds due to the small amount of training data available in these watersheds, suggesting that it is crucial to collect riverine SSC data to develop reliable models. Overall, these results are encouraging as the SSC could be simulated accurately in many of the watersheds.