UNRAVELING SEDIMENT DYNAMICS WITHIN WATERSHEDS FROM PATTERNS IN SUSPENDED SEDIMENT-DISCHARGE RELATIONSHIPS

Studying the variation in in relationships between suspended sediment concentration (SSC) and discharge during individual storm events is an efficient way to gain insight into the drivers and sources of riverine sediment. Classification of hysteresis patterns (e.g. linear, clockwise, counter-clockwise, and figure-eight patterns) in SSC-discharge relationships have commonly been used to characterize events. However, existing gaps in the research remain given challenges associated with visual classification by experts or the collapse of these patterns (visual images) to a hysteretic index. Advances in pattern recognition, such as machine learning methods used for handwritten character recognition, provide an automated and more robust way to recognize and classify these patterns.

In this study, we leverage three-years of high-frequency suspended sediment, discharge, and meteorological monitoring from within the Mad River watershed to demonstrate a new machine-learning approach for classifying storm events based on the type of hysteresis observed. The main stem of the Mad River and five of its tributaries were monitored between 2013 and 2015, providing slightly more than 600 unique events for analysis. Fourteen types of hysteresis were identified within the 600-event data set. Classification was automated by training a restricted Boltzmann machine (RBM), a type of probabilistic artificial neural network, on images of the suspended sediment-discharge plots. The network predicted the correct or next most similar class 71% of the time.

The expanded classification system allowed for new insight into drivers of types of hysteresis including spatial scale, antecedent conditions, hydrology and rainfall. Additionally, differences in the type and frequency of hysteresis were observed between sites and between seasons. This provided insight into types of erosion occurring within the watershed and the proximity and connectivity of sources of riverine sediment. With increased availability of high-frequency suspended sediment data, the approach presented here can be used to characterize watershed erosion and inform watershed management efforts to identify sediment sources and reduce fine sediment export.