Open Collections

Abstract

Water quality monitoring often relies on slow, labor-intensive laboratory tests for chemical and microbial parameters. This thesis investigates a rapid optical alternative: using the intrinsic fluorescence of dissolved organic matter, processed through excitation-emission matrix (EEM) spectroscopy and parallel factor analysis (PARAFAC), combined with machine learning, to predict key water quality indicators. Two case studies are examined: (1) a cascade of urban stormwater ponds in Kelowna, BC, for predicting bulk chemical metrics total organic carbon (TOC) and total nitrogen (TN), and (2) a chlorinated drinking water distribution system in the North Okanagan (RDNO) for predicting a microbiological metric, intact cell count (ICC). In each setting, fluorescence EEMs were decomposed into distinct dissolved organic matter (DOM) components (e.g. humic-like, protein-like), which served as input features for tree-based models (Random Forest and XGBoost). The Kelowna Pond models accurately estimated TOC and TN levels, capturing spatial trends and seasonal shifts. Fluorescence-derived features were found to explain a large portion of TOC/TN variability, enabling reliable classification of high versus low organic concentrations. In the drinking water distribution system, fluorescence alone showed moderate skill in predicting ICC, but incorporating contextual factors (chlorine residual and turbidity) markedly improved performance. The ICC models could categorically distinguish microbial levels within the distribution system, though absolute accuracy declined at sites not seen during training, indicating a need for site-specific calibration. Notably, this work is the first to assess fluorescence-based metrics for determining ICC in drinking water distribution systems. The findings demonstrate that a unified fluorescence-PARAFAC approach can provide timely alternative measures for water quality: it performs strongly for bulk carbon/nutrient monitoring and shows promise for microbial risk indication when augmented with operational data. Overall, the thesis highlights the potential of fluorescence-informed machine learning to bridge conventional chemical and microbiological monitoring, offering a pathway toward more responsive water quality management.