Deep learning insights into suspended sediment concentrations across the conterminous United States: Strengths and limitations

Suspended sediment concentration (SSC) is a crucial indicator for aquatic ecosystems and reservoir management but is challenging to predict at large scales. This study seeks to test the feasibility of deep-network-based models to predict SSC at basin outlets given basin-averaged forcings and basin-physiographic attributes as inputs and extract insights by interpreting the spatially-varying model performances. We trained long short-term memory (LSTM) deep networks either separately for each of the 371 sites across the conterminous United States (local models), or on all the sites collectively (Whole-CONUS). The local and Whole-CONUS models presented median Nash-Sutcliffe Efficiency (NSE) values of 0.72 and 0.57, respectively, which are state-of-the-art results. However, this comparison disagrees with our previous “data synergy” conclusion for LSTM models and suggests there are still important yet unavailable sediment-related attributes. Both local and Whole-CONUS models tended to be more successful where SSC-streamflow correlations (R_s-q) were high - typically in the humid Eastern US - and with lower SSC. Low R_s-q basins were often found in the arid Southwest with higher SSC. The highly-nonlinear SSC-streamflow relationship is arguably due to heterogeneity in land cover and rainfall or limitations in sediment supply, suggesting these basins need to be simulated at higher spatial resolution. The local models mostly outperformed the Whole-CONUS one due to the latter lacking critical attributes, but the latter can be competitive in high-SSC regions with enough flow events. Moreover, the Whole-CONUS model also performed well for basins not included in the training dataset (median NSE=0.55), supporting large-scale modeling.