Machine learning model prediction of unmonitored PM2.5 concentration using meteorological, land use, and stationary monitoring data

Abstract

As the adverse impacts of particulate matter on human health are widely known, the interest in the current concentration and accurate prediction of particulate matter is increasing. However, the existing stationary particulate matter monitoring sensors are sparse and expensive, and the measurement frequency is also insufficient, once an hour in general. In order to solve this problem, this study performed mobile particulate matter monitoring by attaching low-cost particulate matter sensors to a vehicle and identified the characteristics of the data. Using the data obtained from mobile monitoring, this study predicted the particulate matter concentration using several machine learning techniques combined with high-resolution spatiotemporal data under various spatial conditions. The several machine learning techniques stand for classical machine learning techniques, Artificial Neural network (ANN) and Support Vector Machine (SVM), Extreme Gradient Boost (XGBoost), an emerging technique, and Bayesian Neural Network (BNN). The prediction results of the machine learning techniques were compared using model performance evaluation metrics, including Mean Squared Error, Mean Absolute Error, Root Mean Squared Error, and Nash-Sutcliffe efficiency. Furthermore, the BNN computed the uncertainty of predicted particulate matter concentrations under several spatial conditions. The results of the BNN showed spatial mapping of the uncertainty of the predicted concentration as well as the predicted values. Furthermore, the results can suggest where to be more monitored with mobile monitoring. This study presented the characteristics of particulate matter under certain conditions and revealed the need for mobile monitoring. In addition, it determined the best model for the prediction of mobile monitoring particulate matter and showed the location, that preferentially needed mobile monitoring, using the uncertainty of predicted values.