Machine Learning Techniques for Estimation of Suspended Sediment Loading in Urban Watersheds

Abstract:

Urbanization leads to water quality degradation in virtually all urban streams. Stormwater runoff contains suspended solids, nutrients, trace organic compounds, heavy metals, and pathogens that are discharged into natural water bodies, impairing ecosystems, and human health. Physically-based models are frequently used for predicting pollutant loading from urban areas. We used two different machine learning techniques for predicting total suspended solids (TSS) concentration from storm events. The National Stormwater Quality Database (NSQD) was used to generate data for the analysis. Six different factors that are relevant to TSS loading were included in the analysis. The factors that were considered are precipitation depth, the volume of flow, land use, percent imperviousness, antecedent days, and drainage area. Two well knowns machine learning techniques were used in the study; Support Vector Machine (SVM) and Uniform k-Nearest Neighbor (KNN). We used 66% of the data set for training and 34 % for validation. We compared the goodness of fit of each method using the coefficient of determination (R²). Both methods were able to predict TSS loading based on contributing factors. The R² values were 0.65 and 0.7 respectively for SVM and kNN for the training dataset, and 0.58 and 0.36 for SVM and kNN for prediction dataset. The SVM algorithm performed relatively well during both the training and prediction phase. kNN algorithm had an overfitting problem during the training phase and was unable to capture TSS loading during the prediction phase. Machine learning techniques have been primarily used for estimation of runoff from rainfall. This study demonstrated that machine learning techniques can also be applied to predict water quality parameters.