Analyzing the Effects of Data Variability and Quantity on Predicting Particulate Matter (PM2.5) Concentrations: Insights from a Machine Learning Approach

Accurately predicting particulate matter, 2.5 microns or less in diameter (PM2.5), concentrations is imperative to the future of public health and environmental policies. Machine learning models incorporating spatial and temporal datasets to predict PM2.5 concentrations are often limited by data availability and poor-resolution satellite imagery. In this study, we present multiple predictive models designed for generalized PM2.5 predictions, the output of which has been utilized for different spatial locations. Using Random Forest (RF) and Extreme Gradient Boost (XGB) algorithms, these predictive models follow a multidisciplinary approach using Moderate Resolution Imaging Spectroradiometer Aerosol optical depth (MODIS AOD) and surface datasets (relative humidity, barometric pressure, outdoor temperature, wind speed and wind direction). Models are trained and validated based on historical data to evaluate the impact of training data variability and quantity on the predictive performance of RF and XGB models for PM2.5 concentrations. Using MODIS AOD alone yielded weak predictive performance, with average R2 values ranging from -0.06 to 0.07 across the three urban areas (Washington, D.C., Boston, and New York City), highlighting its limited capability. The integration of meteorological data (temperature, wind speed, wind direction, relative humidity, and barometric pressure) along with MODIS AOD significantly improved the model performance. RF models achieved R² values of 0.30–0.62, while XGB models had R² values of 0.25–0.63, with corresponding RMSE values reduced by 20–30% relative to AOD-only models. Feature importance analysis revealed that PM2.5 predictions were most strongly influenced by temperature (average importance of 0.21), wind speed (0.20), and wind direction (0.15). MODIS AOD exhibited moderate importance (≈0.12), indicating that although satellite-based aerosol observations contributed to the predictions, ground-based meteorological variables remained the primary drivers. These quantitative results highlighted that combining satellite observations with meteorological measurements substantially enhanced PM2.5 predictive accuracy, informing urban planning, environmental policy, and public health interventions to better protect vulnerable populations.