Explainable deep learning for dual goals: Predicting total phosphorus concentrations and identifying pollution sources in the Taihu Lake Basin, a tidal river network

Predicting the concentration and identifying the source of phosphorus in aquatic systems are essential for ecosystem health. This study tackled two primary challenges: the intricate biogeochemical cycle of phosphorus, which hinders the accuracy of process-based models, and the time-intensive, resource-demanding nature of experimental and model-based phosphorus tracing methods. We adopted a novel attention physics-guided spatiotemporal graph convolutional neural network, which employs convective diffusion equations to constrain deep learning training for more accurate spatiotemporal multi-node total phosphorus (TP) predictions, and is coupled with an attention-based interpretability method to trace pollution sources. In application to the Taihu Lake Basin (China), this model enhanced TP concentration prediction accuracy by 7.1%–12.3% compared with baseline models. It also effectively identified and quantified the primary pollution source in Gehu Lake under varying seasonal and hydraulic engineering conditions. Examination of the microscale TP migration process revealed an equilibrium mode between TP concentration dilution and sediment disturbance–release under specific river velocity, with an equilibrium velocity of 0.19 m/s. This study underscores the critical role of hydrodynamics, shaped by hydraulic engineering and hydrological variability, in influencing pollutant migration and transformation within tidal river networks, thereby offering new insights into phosphorus prediction and source tracing in complex habitats.