Identifying relationships between urban stormwater signatures and watershed characteristics using interpretable machine learning

Abstract:

Urban stormwater threatens the health of natural water bodies across the United States. This threat grows with continuing urbanization and intensifying storms under climate change. We explored how urbanization intensity, land use, seasonality, climate, and weather variability of watersheds relate to stormwater pollution using machine learning. We analyzed data from the National Stormwater Quality Database, specifically, pollutant concentrations measured in 1,100 stormwater samples from 182 watersheds in 25 US cities between 1992 and 2003. First, we identified stormwater signatures, defined as combinations of distinct concentrations of 9 common pollutants (TSS, TDS, Pb, Zn, Cu, TP, TKN, NO3+NO2, and BOD). To identify them, we employed a machine learning approach to ensemble three clustering algorithms— hierarchical, k-means, and k-medoids— achieving a reduction in variance, avoidance of overfitting, improvement of cluster distinction, and domain-based interpretability. The signatures show that stormwater can be categorized as very dirty (overall high concentrations), moderately dirty with particulates (high metals, TP, and TSS), moderately clean with high organic waste (high TDS and BOD), moderately clean with organic waste (medium TDS and BOD), and moderately clean (overall low concentrations). Next, we used the 5 stormwater signatures as a categorical variable to compare watershed characteristics using multinomial logistic regression. The results showed unique associations with climate, weather, and/or landscape patterns for each signature. For example, the very dirty signature is associated with warmer days, larger storms, and higher imperviousness. We found that watersheds often produce more than one signature, yet which set of signatures they typically produce is associated with their land use and urbanization intensity. Results from this study expose fundamental relationships between place and pollution relevant to stormwater and watershed management.