Data for:Hybrid Deep Learning-Geostatistical Algorithm for Spatial Estimation of PM2.5 Concentrations and Interpretability Analysis of Influencing Factors

This study improves the accuracy of ground-level PM₂.₅ concentration prediction by integrating meteorological, land cover, remote sensing, and socioeconomic factors into a hybrid modeling framework that combines deep neural networks (DNN) and geostatistical techniques (Ordinary Kriging, OK).
Data Description
The dataset contains monthly PM2.5 concentrations and related variables for January to December 2023 across 901 valid monitoring stations in the United States. Each monthly CSV file (01.csv to 12.csv) includes:
1.Ground PM2.5 measurements (µg/m³) from the U.S. EPA Air Quality System (AQS).
2.Meteorological variables from ERA5, including:
2m dew point temperature (2D)
2m air temperature (2T)
10m U- and V-components of wind (10U, 10V)
10m wind speed (S10)
Convective and large-scale precipitation (CP, LSP)
Surface pressure (SP)
Boundary layer height (BLH)
3.Land cover variables: high and low vegetation cover (CVH, CVL) from ERA5; NDVI and EVI from MOD13A3.061.
4.Remote sensing variables: aerosol optical depth (AOD) from MERRA-2, and digital elevation model (DEM) from GLO-90.
5.Socioeconomic variables: population density from USCB, real GDP from BEA, and fire counts from NASA FIRMS.
6.Geospatial coordinates (latitude and longitude) for each station.
All data are processed to remove missing or negative values, ensuring consistency across variables and months.