Applying Deep Learning to Event Concentration-Discharge Hysteresis Patterns to Reveal Differences in Sediment Dynamics across Contrasting Watersheds

Analysis of suspended sediment responses in rivers during storm events and characterization of hysteresis in the concentration-discharge relationship has emerged as a method for identifying sediment sources, controls on transport of sediment, and improving load estimates in watersheds. The availability of high-frequency monitoring data also has continued to increase providing an opportunity for the development and application of automated methods for analyzing concentration-discharge relationships during storm events. In this study, we used high-frequency turbidity-based monitoring of suspended sediment (SS) from twelve watersheds in the Lake Champlain Basin, in northeastern U.S., to study spatial variability in event sediment dynamics and transport across watersheds with contrasting land use, geology, drainage area, and topography. TSS-discharge relationships from 1,000+ storm events allowed for testing the use of deep-learning methods to automate the categorization of hysteresis patterns in the SS-discharge relationship.

The SS-discharge event relationships were transformed into grayscale images, and the resulting pixelated data were used as model inputs. We developed a deep 2-D convolutional neural network (CNN) based on the existing ResNet-50 model architecture to classify storm events. The CNN model achieved an overall classification accuracy of 69% when using previously identified categories of hysteresis patterns. In addition, we developed a denoising auto-encoder network to refine the categories of hysteresis patterns through unsupervised learning. Encoded feature representations of the storm events were extracted from the autoencoder and clustered to confirm hysteresis categories and identify candidate categories of hysteresis types. Using the 1,000+ categorized storm events, we assessed variability in the distribution of observed hysteresis types from different watersheds. This allowed differences in the primary erosional and sediment transport processes to be characterized across the varying watershed types.