Incorporating Light Gradient Boosting Machine to land use regression model for estimating NO2 and PM2.5 levels in Kansai region, Japan

This study incorporates Light Gradient Boosting Machine (LightGBM) to a land use regression (LUR) model for estimating NO 2 and PM 2.5 levels. The predictions were compared with LUR-based machine learnings models of Extreme Gradient Boosting (XGBoost) and Random Forests (RF). Weather Research and Forecasting (WRF) model-simulated meteorological parameters, Community Multiscale Air Quality modeling system (CMAQ)-simulated NO 2 /PM 2.5 concentrations, land use variables, and population data were used as predictor variables. The model performances were evaluated through spatial and temporal cross-validations (CV). The CV results indicated that the LightGBM model was moderately superior in NO 2 and PM 2.5 predictions compared to the RF and XGBoost models. Moreover, the LightGBM model had high performance in NO 2 and PM 2.5 predictions at high concentrations, which is essential for risk assessment. Our findings demonstrate that LightGBM can greatly improve the accuracy of NO 2 and PM 2.5 estimates. • Light Gradient Boosting Machine algorithm was added to Land use regression model. • Daily average NO 2 and PM 2.5 levels are estimated. • Light Gradient Boosting Machine model processes faster than other models. • Light Gradient Boosting Machine model shows superior prediction accuracy. • The developed model has high predictability at high concentration.