Improving prediction of PM2.5 in Metro Manila using XGBoost with Optuna hyperparameter optimization

Reporting air pollution levels in Metro Manila, Philippines remains dependent on records from few ground monitoring stations. For impact studies on human health, grid-based pollution levels datasets will enhance the assessment of the exposure and risk of the local population with a finer spatial resolution. This paper presents an enhanced model for fine particulate matter (PM2.5) in Metro Manila using extreme gradient boosting (XGBoost) with predictor variables from daily satellite observations and climate reanalysis datasets. This study employs XGBoost with Optuna workflow, a Bayesian optimization algorithm, to determine optimal learning parameters achieving least mean absolute error (MAE), root mean squared error (RMSE) and highest coefficient of determination R2 for both training and test datasets to obtain more accurate predictions of PM2.5. Predictor variables assessed are Aerosol Optical Depth (AOD), emissivity, direct solar radiation (DSR), albedo, land surface temperature during daytime (LSTDT) and nighttime (LSTNT), Normalized Difference Vegetation Index (NDVI), wind speed, air temperature, pressure and total precipitation. The results indicate that using a Kriging-based interpolated PM2.5 as target variable achieves the highest R2 = 0.73 and lowest RMSE (1.08 μg/ m3) in the test data using K-fold cross-validation approach (k = 5). The predictors in the best XGBoost model with relatively high feature importance in both gain scores and SHAP values are DSR, precipitation, LSTDT, air temperature and AOD in the 0.47 μm band. These outputs can be utilized in generating daily gridded PM2.5 maps for long-term health impact studies and assessment of seasonal variations of PM2.5 at regional level.