Assistant Professor University of Louisville, Kentucky, United States
Abstract Submission: Dynamic export of suspended and dissolved substances in watersheds during hydrologic events is commonly analyzed using concentration-discharge (C-Q) plots. These plots reveal hysteretic C-Q relationships, offering insights into constituent export timing and relevant watershed transport processes. Traditionally, hysteresis analyses rely on visual classification or the use of an index to characterize hysteresis patterns. However, both methods have limitations: visual classification can be subjective and time-prohibitive for large datasets; meanwhile, collapsing a hysteresis loop into a single index loses valuable information (e.g., figure-eight and linear loop hysteresis indices may be indistinguishable). Recently, machine learning approaches have emerged to address these limitations. However, existing methods require predefined classes, hindering their adaptability to diverse datasets. To overcome these challenges, we propose a novel machine learning approach that combines Self-Organizing Maps with Dynamic Time Warping to characterize hysteresis patterns. Our technique automatically extracts typical hysteresis loops from a dataset, arranging them in a bidimensional representation preserving the gradual transitions between classes. We applied this approach to over 6000 hysteresis loops of discharge and turbidity from 80+ watersheds across the United States. Notably, our algorithm not only identifies previously reported classes but also uncovers previously unnoticed patterns. We demonstrate how this method can be used to characterize frequency distributions of hysteresis patterns and detect associations with hydrologic conditions—such as storm magnitude and antecedent conditions. Our approach can be used to inform watershed management strategies by shedding light on the processes governing the export of suspended and dissolved substances within watersheds.
