Automating the Classification of Hysteresis in Event Concentration-Discharge Relationships

The response of in-stream sediment concentration and discharge during rainfall-runoff events provides information about dominant watershed processes as it represents the amalgamation of the connectivity, erodibility, and the spatial location of sediment sources. A common way to collapse the sediment and streamflow response into a readily interpretable visualization is to utilize an event concentration-discharge (C-Q) plots which frequently exhibit patterns of hysteresis. However, challenges exist given the subjective nature of visual classifications and when scaling to large data sets. Hysteresis indices have been used to facilitate an automated and objective analysis method. In this study, we present an alternative method for automating hysteresis classification utilizing all the information present in the event C-Q plots. Thus,avoiding the loss of information that may occur when collapsing data into metrics and enabling the local sediment dynamics to be interpreted to a greater extent.

We developed an automated machine learning tool using images of event C-Q plots to classify storm events into pre-defined hysteresis pattern types. The classifier utilizes a convolutional neural network, a machine learning method that has achieved excellent predictive accuracy in image classification tasks. We then applied this tool using surrogate suspended sediment data from turbidity monitoring in eight watersheds within the Lake Champlain Basin in Vermont encompassing 760 individual storm events. The tool accurately and efficiently classifies events and represents an advancement over manual visual classification