Machine Learning and Citizen Science for Catchment-Scale Water-Quality Assessment

Abstract Assessing and managing water quality in data-scarce tropical basins is increasingly challenging under accelerating land-use change, intensifying human pressures, changing livelihood practices, and climate driven shifts in precipitation regimes. In the transboundary Mara River Basin (MRB), these pressures strongly influence sediment and nutrient dynamics, yet conventional monitoring networks operated by national authorities lack the temporal and spatial resolution required to characterize pollution sources and event-driven responses. To address these gaps, this study integrates citizen-science observations, with known and quantified uncertainty characteristics, with satellite derived hydro-climatic and land-use variables within an interpretable machine learning framework to investigate the drivers of turbidity (NTU), nitrate (NO₃), and phosphate (PO₄) dynamics. Citizen scientists were recruited, trained, and equipped by local ministry staff following standard methods and measurements were quality controlled before analysis. Random Forests (RF) and Artificial Neural Networks (ANN) were developed to predict water-quality variability and quantify the relative influence of environmental and anthropogenic drivers. RF provided consistently better predictive performance than ANN, indicating that tree-based ensemble methods are well suited to heterogeneous citizen-science datasets while providing interpretable estimates of predictor importance. Model interpretation using SHAP values revealed that precipitation and river flow velocity dominate short-term, event-based fluctuations of turbidity, while population density was an important predictor of NO₃, consistent with diffuse sources. Site-level analyses further demonstrated pronounced spatial heterogeneity in precipitation sensitivity, identifying mid- to lower-basin reaches as particularly responsive to rainfall-driven sediment mobilization. These findings demonstrate that integrating citizen science with interpretable machine learning can strengthen pollution assessment, identify catchment-scale drivers and hotspots, and support adaptive water-quality management in data-scarce river basins undergoing rapid land-use and climate change.