A Regression Framework for Spatial Downscaling of NASA GEOS-CF PM2.5 Predictions Using Satellite Observations, Spatial covariates, and Ground-Based Measurements

This repository contains the datasets and source code required to reproduce the 1-km resolution PM2.5 estimates for Africa as described in the accompanying manuscript: "A Regression Framework for Spatial Downscaling of NASA GEOS-CF PM2.5 Predictions Using Satellite Observations, Spatial Covariates, and Ground-Based Measurements."

The project utilizes a Random Forest-based downscaling framework to bridge the gap between global chemical transport models (CTMs) and sparse ground-level monitoring. This framework produces high-resolution surface estimates that resolve intra-urban pollution gradients and align with measured patterns  at U.S. embassy locations across the African continent.

Contents:

Code and Notebooks:

• 00_GEE_Data_Extraction.ipynb: This notebook executes the multi-source data extraction from the Google Earth Engine Python API. It performs the aggregation of GEOS-CF and environmental predictors into monthly and annual mean rasters at 0.25° (training) and 0.01° (downscaling) resolutions.
• 01_Baseline_Model.ipynb: This notebook executes the primary downscaling pipeline. It implements the stepwise, importance-driven forward predictor selection for each of the five African regions, followed immediately by the training and 10-fold cross-validation of the uncalibrated (CTM-only) Random Forest and OLS baseline models.
• 02_Calibrated_Model.ipynb: This notebook implements the final ground-calibration phase. It integrates the spatial scale indicator (Si) and U.S. Embassy ground monitoring data to bias-correct the baseline estimates and produce the final 1-km pollution surfaces.
• 03_Downscaled_Estimates_ByModel.ipynb: Generates final 1-km continental PM2.5 estimates. It applies the trained baseline and calibrated models to the high-resolution (1km) predictor grids to produce the final GeoTIFF outputs at annual and monthly timesteps.
• 04_Performance_Evaluation.ipynb: Performs statistical validation against independent ground measurements. It generates the continental and regional performance metrics (Tables 2 and 3) and produces the monthly concentration time series (Figure 3).
• 05_Spatial_Analysis_&_Mass_Conservation.ipynb: Evaluates the mass conservation of the downscaling framework. It compares aggregated 1-km estimates against the 25-km GEOS-CF truth (Table 4) and generates continental maps of annual mean PM2.5 and error distribution (Figure 4).
• 06_Predictor_Influence.ipynb: Analyzes variable importance across both model types and all regions. It generates the normalized importance heatmaps (Figure 5) used to interpret the primary environmental drivers of pollution in each region

Data Files:

• Ground_Measurements.zip: Contains monthly aggregated PM2.5 concentrations (2019–2024) from U.S. Embassy monitors across the African continent used for model calibration and validation.
• Annual_Maps.zip: Final downscaled 0.01° (~1 km) annual mean PM2.5 mass (RH35) GeoTIFFs for all African regions.
• Monthly_Maps.zip: Final downscaled 0.01° (~1 km) monthly mean PM2.5 mass (RH35) GeoTIFFs for all African regions.
• All_Monthly_Datasets_25km.zip & All_Annual_Datasets_25km.zip: The regional 0.25° predictor and GEOS-CF target rasters required for model training and mass-balance evaluation

Metadata:

• variable_selection_summary.json: The output generated by the baseline notebook containing the parsimonious list of predictors and performance progression (R², RMSE) for each region. Created by running 01_Baseline_Model.ipynb.
• performance_evaluation_results.csv: The compiled prediction-observation pairs used to generate the manuscript figures and tables. Created by running 04_Performance_Evaluation.ipynb.