A Regression Framework for Spatial Downscaling of NASA GEOS-CF PM2.5 Predictions in Africa Fusing Satellite Observations, Spatial Covariates, and Ground-Based Measurements

This archive contains the code and analysis-ready data needed to reproduce the results of "A Regression Framework for Spatial Downscaling of NASA GEOS-CF PM2.5 Predictions in Africa Fusing Satellite Observations, Spatial Covariates, and Ground-Based Measurements" (Earth and Space Science). The study downscales the NASA GEOS-CF PM2.5 field from 0.25 degrees to 0.01 degrees (about 1 km) across Africa for 2019 through 2024. The coarse field is first corrected to satellite-observed aerosol optical depth, which defines the downscaling target. The target is then downscaled with random forest and ordinary least squares regressions and calibrated against United States embassy and consulate monitors. The final model is a single continental Ground-Calibrated random forest. The archive holds the seven Jupyter notebooks used for the analysis, in run order, together with the data they require. Notebook 01 is the Google Earth Engine predictor export. Notebook 02 is the gap-fill method evaluation (donor-mask reconstruction, four candidate methods, IDW optimization). Notebook 03 is the random forest and OLS predictor selection. Notebook 04 fits the Baseline and Ground-Calibrated models, evaluates them under all four criteria (in-sample, random 10-fold cross-validation, leave-one-year-out, and leave-one-site-out), reports per-station skill, and isolates the effect of the AOD correction. Notebook 05 fits the deployment model, predicts the 1 km surfaces, and computes the mass balance. Notebook 06 computes the variable importances, both mean decrease in impurity and the Saabas path attribution, at 0.25 and 0.01 degree resolution. Notebook 07 draws the figures. All model fitting is deterministic, so a rerun regenerates the same results by the same mechanism. The data folder provides the analysis-ready 0.25 degree training matrices for both target configurations (AOD-corrected and uncorrected), the published predictor selections, the published result tables, attribution summaries, the per-monitor monthly series behind the SI time-series figures, reference renders of the published figures, the AOD-corrected GEOS-CF coarse fields, and the five-region definition shapefile. The deposit also includes the published 1 km product itself: surfaces_1km.zip holds the 78 monthly and annual Ground-Calibrated random forest GeoTIFFs at 0.01 degrees. Unzipped into the data folder, it enables the mass-balance and Figure 5 steps directly. Every analysis step is present as code. Two raster inputs remain too large to archive, the native monthly composites and the aligned 1 km predictor composites; the steps that need them skip with a note unless those inputs are attached via path variables, and their archived outputs are included so every reported number can still be inspected. Our original work is released under CC-BY 4.0. Predictors are derived from public datasets (NASA MODIS, VIIRS, and GEOS-CF, Copernicus Sentinel-5P and Land Cover, USGS GMTED2010, NASA SEDAC GPWv4, and WorldPop) and from U.S. Department of State monitors distributed through AirNow. Attribution for each source is given in PROVENANCE.md. Please cite the journal article as the primary reference and this archive for the code and data.