Scalable Bayesian Geostatistical Regression Model for Bias-Correcting Large-Scale Daily Satellite-Retrieved Aerosol...

Area of Science:

• Environmental Science
• Biostatistics
• Geostatistics

Background:

• Ambient fine particulate matter (PM2.5) exposure is associated with adverse health outcomes.
• Accurate, high-resolution PM2.5 prediction and uncertainty quantification are vital for epidemiological research.
• Existing methods may lack scalability or sufficient accuracy for large-scale health analyses.

Purpose of the Study:

• To introduce a scalable Bayesian spatial-temporal geostatistical regression model (GRM) for improved PM2.5 prediction.
• To bias-correct Chemical Transport Model (CTM) outputs or satellite-retrieved Aerosol Optical Depth (AOD) data.
• To enhance spatial and temporal resolution in PM2.5 predictions for health-related studies.

Main Methods:

• Developed a scalable Bayesian spatial-temporal geostatistical regression model (GRM).
• Employed nearest neighbor Gaussian process (NNGP) random effects for computational efficiency.
• Applied the model to CTM outputs (12-km resolution, US) and AOD data (1-km resolution, California).
• Assessed hybrid GRM-Random Forest (RF) models for enhanced predictive accuracy.

Main Results:

• The GRM demonstrated efficient spatial correlation exploitation and faster computation compared to regular Gaussian processes.
• Cross-validation confirmed the model's spatial predictive performance and accurate uncertainty quantification over the US.
• In California, the GRM-RF hybrid model achieved superior predictive accuracy (RMSE 7.325 μg/m3) over standalone RF (7.732 μg/m3) and GRM (8.557 μg/m3).
• The GRM-RF maintained accurate predictive uncertainty (95% credible interval coverage 0.957).

Conclusions:

• The proposed scalable Bayesian GRM, particularly the GRM-RF hybrid, offers a robust approach for high-resolution PM2.5 prediction.
• The model provides accurate uncertainty quantification crucial for health impact assessments.
• Open-source R software is provided to facilitate broader research implementation.