Spatio-Temporal Prediction of Roadside PM2.5 based on Sparse Mobile Sensing and Traffic Information

While real-time management of urban mobility has become common in modern cities, it is now imperative to attempt such management subject to a sustainable emission target. To achieve this, one would require emission estimates at spatiotemporal resolutions that are significantly higher than the usual. In this paper, we consider roadside concentration of PM2.5, and make predictions at high spatio-temporal resolution based on location, time and traffic levels. Specifically, we optimized various machine learning models, including ones involving bagging and boosting, and found Extreme Gradient Boosting (XGBoost, XGB) to be superior. Moreover, the tuned and optimized XGB utilizing traffic information achieved significant gain in terms of multiple performance measures over a reference method ignoring such information, indicating the usefulness of the latter in predicting PM2.5 concentration.