Spatial attribution analysis of river water quality by integrating remote sensing inversion and multi-source land cover data: An interpretable topological residual Knowledge-Driven Graph Attention Network (ITRK-GAT)

Accelerating urbanization places urban river networks under intense, spatially heterogeneous human pressure; yet conventional remote-sensing inversion, though effective for large-scale monitoring, reveals little about pollution sources, transport pathways, or their terrestrial mechanisms. To help bridge this “inversion-over-attribution” gap, this study proposes an interpretable Topological Residual Knowledge-Driven Graph Attention Network (ITRK-GAT), for the Lianqi River, Shanghai. The framework couples multi-source terrestrial features—land use, discharge outfalls, and pollution-related enterprises—with river-network topology through an undirected graph over river grid units, in which topological relationships and along-stream distance priors act as structural constraints for physically consistent propagation. Chemical oxygen demand (COD) is predicted with accuracy outperforming standard graph and tabular baselines, and ablation confirms the contribution of each module. Integrated Gradients attribution and attention mapping jointly indicate that residential and industrial land use, together with discharge outfalls, are most strongly associated with elevated COD, with high-attribution regions co-located with tributary confluences and discharge-dense corridors. Controlled perturbation further shows that these associations propagate along network connectivity and attenuate with topological distance, while the attribution ranking stays stable under inversion-induced label noise. Together, these analyses establish a robust, physically coherent attribution framework offering an interpretable and transferable basis for pollution-source tracing and spatially differentiated management of urban rivers.