High-Resolution Disaggregation of Pollution Loads through Deep Learning Fusion of Multi-source Spatial Proxies
Precision environmental management demands high-resolution pollution source attribution, yet traditional emission inventory methods are fundamentally limited by their reliance on static, linear assumptions that fail to capture spatial heterogeneity.To bridge this gap, we propose a physics-constrained deep learning framework that transforms the emission inventory paradigm from deterministic estimation to verifiable intelligent disaggregation.A core methodological innovation is the embedding of a mass conservation constraint directly into the Convolutional Neural Network (CNN) architecture via a Softmax layer, ensuring rigorous physical consistency between grid-level predictions and aggregate-level supervision (wastewater treatment plant influent loads).By fusing multi-source spatial proxies—specifically building volume (VC) and cellular base station density (BS)—the model effectively captures the non-linear emission fingerprints of dynamic human activities.Validated in the Erhai Lake Basin, our model generates high-fidelity 500 m maps for wastewater (WW), chemical oxygen demand (COD), total nitrogen (TN), and total phosphorus (TP), achieving high predictive accuracy (R² = 0.82–0.98, RMSE < 0.045). The results reveal a fundamental characteristic of urban pollution: extreme spatial concentration, with the top 5% of grid cells contributing 60%–74% of total loads, predominantly in southern and western urban-tourism zones. Strong inter-pollutant synergies were uncovered, with WW and TN exhibiting near-perfect spatial correlation (r = 0.96). Model interpretability via Shapley Additive Explanations (SHAP) identifies VC and BS as the dominant drivers, accounting for over 80% of explanatory power, as they robustly proxy static population density and dynamic human activity, respectively. This study presents a transferable methodology for high-resolution pollution attribution, particularly valuable in data-scarce regions, and offers a transformative tool for advancing from coarse-scale mitigation to precision water resource management.