MACHINE LEARNING-BASED REGRESSION MODELING FOR PREDICTING SURFACE PM2.5 IN KANTO REGION, JAPAN

Acquiring detailed information on the distribution of fine particulate matter (PM2.5) is important for comprehensive reporting of air pollution levels in urban areas. However, for large-scale projects, it is difficult to rely only on ground measurements and limited number of sensors. This study presents a machine learning (ML)-based modeling approach to estimate surface PM2.5 levels in Kanto region based on input variables from an atmospheric chemistry model. Anthropogenic heat emissions (AHE), anthropogenic and biogenic emission sources were considered in the numerical simulations using Weather Research and Forecasting coupled with Chemistry (WRF-Chem). Meteorological parameters (i.e., aerosol optical depth, temperature, relative humidity, wind speed, surface albedo, emissivity, rainfall, pressure, shortwave radiation) and topographical variables (i.e., AHE, leaf area index, vegetation fraction, terrain height, land use and distance from coastline) were set as predictors in training tree-based regression ML models. Statistical results show that in 50 prediction trials, XGBoost achieves better accuracy than LightGBM and Random Forest in terms of MAE (0.094 μg/m³), RMSE (0.147 μg/m³) and R2 (0.990). The processes presented here serves as a preliminary workflow in developing a robust model for operational use in reporting surface PM2.5 levels in urban environments.