Modelling Cryptosporidium risks in catchments to protect drinking water supplies

Understanding source water characteristics – and the catchments which influence them – can play a vital role in protecting drinking water supplies. Many water providers collect vast quantities of water quality monitoring data, but this data is often neglected after first use, though it may contain important information for understanding contaminant sources, processes and risks. The research in this thesis examines Cryptosporidium – a major waterborne pathogen – by applying machine learning (ML) and Quantitative Microbial Risk Assessment (QMRA) approaches to long-term water quality, hydro-meteorological and geographical datasets. The overall aim is to advance understanding of how catchment conditions influence Cryptosporidium river concentrations and the consequent effect on public health risk in drinking water supply systems. The opening work describes a first-of-its-kind ML-based daily Cryptosporidium prediction tool for use at a major abstraction site on the River Thames. The principal advance is the comparison and integration of catchment-averaged and spatially distributed environmental and source data to demonstrate predictability at operational timescales and identify potential source areas, conditions and lag times associated with elevated Cryptosporidium concentrations. The second study presents a QMRA for London’s drinking water supply, integrating river, reservoir and final treated water data across a complex multi-source system. This enables comparison between loading-sensitive raw water-based and treated water-based risk estimates, improving assessment of seasonal effects and treatment robustness. A third study applies an explainable ML approach to evaluate the impact of static, pseudo-static and dynamic environmental variables across multiple catchments. Higher humidity and recent rainfall events were associated with higher Cryptosporidium river concentrations, and 33% of sites were identified as dominated by Cryptosporidium shed by human – as opposed to animal – hosts. Taken together, these findings have important implications for drinking water providers and catchment managers, and demonstrate the latent value within large-scale historical monitoring datasets, which can be readily harnessed using ML tools.