High temporal resolution prediction of street-level PM_2.5 and NO_x concentrations using machine learning approach

Accurate and high temporal resolution predictions of fine particulate matter (PM_2.5) and nitrogen oxides (NO_x) concentrations are crucial for pollution control, air pollutant exposure and epidemiological studies. This study aimed to develop machine learning-based models for predicting hourly street-level PM_2.5 and NO_x concentrations at three roadside stations in Hong Kong. We comprehensively evaluated and compared the performance of six common machine learning algorithms (MLAs) including Random Forest (RF), Boosted Regression Trees (BRT), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), Generalized Additive Model (GAM), and Cubist and hence applied the most suitable MLAs to apportion the contributions from emission and non-emission factors to hourly street-level PM_2.5 and NO_x concentrations. The results show that RF outperformed other MLAs with ten-fold cross validation (CV) R² values higher than 0.81 and 0.62 for PM_2.5 and NO_x predictions, respectively. BRT, XGBoost and Cubist presented comparable predictive performances, with CV R² of 0.79–0.83 (for PM_2.5 predictions) and 0.59–0.71 (for NO_x predictions). SVM and GAM had worse predictions than other MLAs. The external validation R² values for RF and BRT models were more than 0.62 and 0.51 for PM_2.5 and NO_x concentration predictions, respectively. Non-emission factors contributed 84% and 65% to the predictions of street-level PM_2.5 and NO_x concentrations, respectively. Non-local pollution and temperature were the major non-emission factors, whereas private cars were the major emission contributor. This study highlights the capability of MLAs to produce high temporal resolution air pollution predictions, which can supplement traditional methods (e.g., land use regression) in generating accurate and high-temporal-resolution estimations of air pollution concentration.