A Stacked Ensemble Learning Approach for Hyper-Local PM2.5 Prediction Using Satellite AOD and Meteorological Data
Accurate local air quality estimation is challenging in areas with limited ground-based monitoring stations. This study investigates a stacked ensemble approach for predicting PM2.5 concentrations by combining satellite-derived Aerosol Optical Depth (AOD), meteorological variables, and geographic information, with Google Earth Engine (GEE) supporting the geospatial workflow.
Random Forest and XGBoost serve as complementary base learners, while a linear regression meta-model combines their predictions. The study evaluates the ensemble against individual models and examines which environmental features influence its estimates. A Flask-based API connects the trained model with live weather inputs to demonstrate PM2.5 prediction and AQI categorization, linking environmental research with an interactive application.
Explore Research