A New Machine‐Learning Approach for Classifying Hysteresis in Suspended‐Sediment Discharge Relationships Using High‐Frequency Monitoring Data

Studying the hysteretic relationships embedded in high‐frequency suspended‐sediment concentration and river discharge data over 600+ storm events provides insight into the drivers and sources of riverine sediment during storm events. However, the literature to date remains limited to a simple visual classification system (linear, clockwise, counter‐clockwise, and figure‐eight patterns) or the collapse of hysteresis patterns to an index. This study leverages 3 years of suspended‐sediment and discharge data to show proof‐of‐concept for automating the classification and assessment of event sediment dynamics using machine learning. Across all catchment sites, 600+ storm events were captured and classified into 14 hysteresis patterns. Event classification was automated using a restricted Boltzmann machine (RBM), a type of artificial neural network, trained on 2‐D images of the suspended‐sediment discharge (hysteresis) plots. Expansion of the hysteresis patterns to 14 classes allowed for new insight into drivers of the sediment‐discharge event dynamics including spatial scale, antecedent conditions, hydrology, and rainfall. The probabilistic RBM correctly classified hysteresis patterns (to the exact class or next most similar class) 70% of the time. With increased availability of high‐frequency sensor data, this approach can be used to inform watershed management efforts to identify sediment sources and reduce fine sediment export. Plain Language Summary In this study, the river stage (water level) and amount of suspended sediment (soil particles) within a river and five of its tributaries were monitored for 3 years; more than 600 storm events were captured across all six sites. For each storm event, traces of the sediment concentration and river stage were plotted against each other; and the emerging patterns such as clockwise, counter‐clockwise, and figure‐eight (hysteresis) loops were grouped into 14 reoccurring patterns. We also developed a machine‐learning (artificial intelligence) tool to recognize the 14 patterns using only the visual sediment‐stage image, in the same way that handwritten characters are recognized by computers. This allowed classification of the individual storm events to be automated. To better understand what these patterns tell us about the physics associated with the storm events and where on the landscape sediments may originate, we analyzed the 14 storm categories using measured rainfall, soil moisture, sediment, and river level dat