Explainable Machine Learning for High-Resolution Modeling of Long-Term Atmospheric Nitrogen and Sulfur Wet Deposition across the United States

Wet deposition of nitrogen and sulfur compounds is a major pathway through which atmospheric pollutants influence ecosystem nutrient cycles and regional air quality. Existing methods for estimating atmospheric nitrogen and sulfur wet deposition, such as sparse in situ observations and computationally intensive chemical transport models (CTMs), often suffer from limited spatial coverage, coarse resolution, or high uncertainty in complex terrain and emission environments. In this study, we developed a high-resolution machine learning framework integrated with SHapley Additive exPlanations (SHAP) to investigate the spatiotemporal variability of the wet deposition of nitrate (NO₃^-), ammonium (NH₄⁺), and sulfate (SO₄^2-) across the contiguous United States (CONUS) from 2002 to 2017. The model achieved strong predictive performance (<i>R</i>² > 0.60), outperformed the high-resolution CTMs, and accurately reproduced both seasonal and spatial deposition patterns across nine dominant land-use categories. Over the study period, SO₄^2- and NO₃^- wet deposition decreased by 31.0% and 21.0%, respectively, reflecting the long-term effectiveness of national controls on SO₂ and NO_<i>x</i>. The most pronounced decreases occurred over croplands and urban/forested ecosystems. In contrast, NH₄⁺ deposition increased by 13.0% on average, with the strongest rise (+16.6%) over agricultural landscapes, indicating growing contributions from ammonia emissions that remain largely unregulated. By integrating high predictive accuracy with transparent interpretation of dominant environmental and anthropogenic drivers, this framework provides a robust basis for assessing deposition trends and supporting targeted mitigation strategies for reactive nitrogen and sulfur.