Stacking Regressor Model for PM2.5 Concentration Prediction Based on Spatiotemporal Data

Abstract

This study presents the development of a predictive model for PM2.5 concentrations resulting from forest and peatland fires in Riau Province, utilizing the stacking regressor technique within an ensemble learning framework. The model integrates spatiotemporal data from remote sensing and ground-based sensors at a resolution of 1 km x 1 km, demonstrating its effectiveness in capturing the intricate patterns of PM2.5 concentrations. By combining Random Forest, Gradient Boosting Machine (GBM), and XGBoost, with RidgeCV as a meta-learner, the model attained optimal performance, achieving R² = 0.851, MAE = 0.045 µg/m³, and MSE = 0.003 µg/m³. The incorporation of temporal feature engineering techniques, including lag and rolling window methods, significantly enhanced prediction accuracy, enabling the model to effectively capture seasonal variations and temporal dynamics. Key variables, such as air temperature, evapotranspiration, and Aerosol Optical Depth (AOD), were found to exhibit strong correlations with PM2.5 concentrations. The findings from this research contribute to the formulation of data-driven policies for air quality management and pollution mitigation, with the potential for broader application in regions encountering similar environmental challenges.