Environmental factors and sources influencing Cryptosporidium river concentrations: A data-rich spatio-temporal...

Cryptosporidium is a waterborne parasite which poses significant risks to drinking water safety. Despite its importance, considerable uncertainty remains regarding the effect of catchment characteristics and environmental conditions on the scale and timing of Cryptosporidium concentrations in rivers. This paper performs a spatio-temporal analysis of the environmental factors influencing Cryptosporidium by applying eXplainable AI (XAI) tools to a large-scale, long-term dataset from across the British mainland. The study brings together extensive land use, topographical, meteorological, hydrological and human/livestock density data to analyse the seasonality and magnitude of Cryptosporidium in rivers and advance understanding of how these factors influence Cryptosporidium concentrations and dynamics across heterogeneous river systems. A machine learning (ML) model - XGBoost - was trained with environmental inputs at different spatial scales. Interpretation of feature effects using SHapley Additive exPlanations (SHAP) found that humidity and rainfall had a strong positive effect on modelled Cryptosporidium concentrations. Human population density was shown to be more highly predictive of Cryptosporidium concentrations than either grassland area or livestock density, suggesting that human sources of Cryptosporidium may be more important than animal sources in many rivers. Densities of combined sewer overflows (CSOs) and sewage treatment works (STWs) were most predictive within 5 km and 25 km upstream, respectively, indicating the possible significance of local point source effects. The study revealed strong seasonality in Cryptosporidium at many sites, with the timing of peak concentrations varying significantly with catchment characteristics, such as human population, STW density, urban area and prevailing meteorological conditions. The work demonstrates how an XAI-based modelling framework can combine static, pseudo-static and dynamic environmental datasets across multiple catchments to identify modelled drivers of pathogen occurrence and seasonality in river systems, yielding more generalisable insights than those provided by previous single-catchment or short-term studies.