Potential of machine learning for prediction of traffic related air pollution

Abstract

Land use regression (LUR) has been extensively used to capture the spatial distribution of air pollution. However, regional background and non-linear relationships can be challenging to capture using linear approaches. Machine learning approaches have recently been used in air quality prediction. Using data from a mobile campaign of fine particulate matter and black carbon in Toronto, Canada, this study investigates the boundaries of LUR approaches and the potential of two different machine learning models: Artificial Neural Networks (ANN) and gradient boost. In addition, a moving camera was used to collect real-time traffic. Models developed for fine particulate matter performed better than those for black carbon. For the same pollutants, machine learning exhibited superior performance over LUR, demonstrating that LUR performance could benefit from understanding how explanatory variables were expressed in machine learning models. This study unveils the black-box nature of machine learning algorithms by investigating the performance of different models in the context of how they capture the relationship between air quality and various predictors.