A machine learning-based method for predicting the spatio-temporal distribution of heavy metal polluted water bodies

To enhance the prediction accuracy of heavy metal pollution in water bodies across space and time, this study constructs a composite framework integrating a graph convolutional network (GCN) and bidirectional gated recurrent unit (Bi-GRU). By combining high-frequency surface water monitoring data with multi-source environmental factors, a refined model of pollutant concentration and trend evolution is developed. Focusing on a typical composite watershed in East China, a “fixed + mobile”sampling strategy is used to build a multidimensional feature tensor, incorporating pollutant levels, water quality, meteorological drivers, and spatial structure variables. The model demonstrates strong spatial structure representation and temporal trend response, effectively identifying pollution diffusion paths and high-risk areas. Results show the method is adaptable and robust in multi-scale accuracy evaluation, spatial consistency, and pollution trend early warning, with promising potential for regional-level engineering applications.