A physically constrained and interpretable deep learning framework for PM_2.5 inversion under sparse monitoring...

Monitoring fine particulate matter (PM_2.5) in arid regions remains difficult due to sparse ground-based networks and frequent extreme pollution episodes. Conventional inversion models often fail to capture these sudden concentration spikes. To address this gap, we developed ResNet-DFL, a physically constrained and interpretable deep learning framework aimed at satellite-based PM_2.5 estimation in data-scarce areas. We applied a distribution-aware learning strategy to reduce long-tail bias and integrated Shapley Additive Explanations (SHAP) with two-dimensional Partial Dependence Plots (2D-PDPs) to quantify key drivers and their nonlinear interactions. Validation using independent sparse-station scenarios showed that ResNet-DFL achieves a spatial R² of 0.79 and a temporal R² of 0.68. It consistently outperforms benchmark baseline models like Random Forest and Geographically and Temporally Weighted Regression, particularly during heavy pollution events. Our interpretability analysis shows that relative humidity, temperature, and NDVI are the dominant factors governing local pollution. Furthermore, 2D-PDPs confirm that low temperature combined with high humidity primarily triggers PM_2.5 accumulation in this region. We also found that a vertically normalized aerosol optical depth (R_AOD) is critical to link column-integrated satellite observations with near-surface PM_2.5 by adjusting for boundary layer dynamics. Ultimately, ResNet-DFL provides a reliable tool for PM_2.5 estimation in data-scarce arid regions, providing practical data support for local air quality management and targeted pollution control.