From Local Knowledge to Decision Support: A Causal Top-Down Bayesian Network for Drinking-Water Intake Risk Assessment

Source water protection is one of the most critical barriers in the multi-barrier approach to ensure safe drinking water. However, identifying and prioritizing upstream hazards are still significant challenges for utilities. Several methods, including machine learning, deep learning, and process-based models, have been applied to risk assessment. These approaches are typically developed using numerical scientific measurements. Despite their high analytical precision, traditional monitoring programs are often expensive and difficult to implement in remote regions. They also frequently miss short-term pollution events such as Combined Sewer Overflows (CSOs). Given the uncertainty this discrepancy creates in risk assessment, independent sources of evidence are required to verify assessment results. In such cases, observations from residents and local users of a water body could represent a valuable data source for water quality monitoring and offer essential reference data to validate models where scientific records are limited. To make these qualitative observations comparable with quantitative scientific data, a structured modeling framework is required. Bayesian Networks can address this challenge by quantifying uncertainty and by integrating non-scientific inputs, such as local knowledge, into a structured risk assessment framework.

Using scientific datasets, including municipal CSO records, meteorological observations, and water quality measurements, together with local knowledge from surveys of watercourse users, this study develops a causal top-down Bayesian Network. In this approach, the network structure is constructed a priori based on theoretical causal mechanisms and expert knowledge rather than being learned computationally from data, ensuring physical interpretability. A fuzzy algorithm was used to quantify subjective expert knowledge into the numerical probabilities required for conditional probability tables. The proposed framework compares the capabilities of these distinct data sources in assessing microbial risk levels at selected drinking-water intakes in southern Quebec, Canada. This research investigates the assessment capacity of non-scientific data sources for microbial risk level estimation at drinking-water intakes, comparing their reliability relative to available scientific monitoring records. Compared with findings from previous studies and reports in the same area, this study shows that information reported by water body users can produce realistic and rational estimates of microbial risk levels. The proposed approach offers a lower-cost data source suitable for remote areas and capturing event-based pollution episodes.