Data for:A Hybrid Deep Learning-Geostatistical Framework for PM2.5 Spatial Estimation with SHAP Interpretation

Description This study improves the spatial prediction of ground-level PM₂.₅ concentrations by integrating meteorological, land-cover, remote-sensing, socioeconomic, and geographic predictors within a hybrid deep learning–geostatistical framework. The proposed DeepResKriging model first uses a deep neural network (DNN) to estimate the nonlinear mean component and then applies ordinary kriging (OK) to the remaining spatially correlated residuals. The accompanying dataset and source code support the analyses reported in the study, including nested random cross-validation, nested 500-km spatial-block cross-validation, repeated cross-validation robustness analysis, residual spatial autocorrelation analysis, spatial prediction and uncertainty assessment, and SHAP-based interpretation. Data Description The dataset contains monthly PM₂.₅ concentrations and associated predictors for January–December 2023 across 901 valid monitoring stations in the contiguous United States (CONUS). Each monthly CSV file (01.csv–12.csv) includes: Ground-level PM₂.₅ measurements (µg m⁻³) from the U.S. Environmental Protection Agency Air Quality System (EPA AQS). Meteorological variables from ERA5, including 2-m dew-point temperature (2d), 2-m air temperature (2t), 10-m zonal and meridional wind components (10u and 10v), 10-m wind speed (s10), convective precipitation (cp), large-scale precipitation (lsp), surface pressure (sp), and boundary-layer height (blh). Land-cover and vegetation variables, including high- and low-vegetation cover (cvh and cvl) from ERA5 and NDVI and EVI from MOD13A3.061. Remote-sensing and topographic variables, including aerosol optical depth (AOD) from MERRA-2 and elevation (DEM) from the Copernicus GLO-90 digital elevation model. Socioeconomic and fire-activity variables, including population density from the U.S. Census Bureau, real GDP from the U.S. Bureau of Economic Analysis, and monthly fire counts from NASA FIRMS. Geographic coordinates (longitude and latitude) for each monitoring station. During preprocessing, records with missing or negative PM₂.₅ values or missing coordinates were removed, and invalid predictor fill values were converted to missing values. For model evaluation, all preprocessing operations that required estimation from the data were performed using only the corresponding training data within the cross-validation procedure, thereby preventing information leakage from validation or test folds. The response variable was modeled using the transformation log(PM₂.₅ + 1), and predictions were back-transformed to the original concentration scale for evaluation.