Applying traffic camera and deep learning-based image analysis to predict PM_2.5 concentrations

Background: Air pollution has caused a significant burden in terms of mortality and mobility worldwide. However, the current coverage of air quality monitoring networks is still limited. Objective: This study aims to apply a novel approach to convert the existing traffic cameras into sensors measuring particulate matter with a diameter of 2.5 μm or less (PM_2.5) so that the coverage of PM_2.5 monitoring could be expanded without extra cost. Methods: In our study, the traffic camera images were collected at a rate of 4 images/h and the corresponding hourly PM_2.5 concentration was collected from the reference grade PM_2.5 station 3 km away. A customized neural network model was trained to obtain the PM_2.5 concentration from images followed by a random forest model to predict the hourly PM_2.5 concentration. The saliency maps and the feature importance were utilized to interpret the neural network. Results: Proposed novel approach has a high prediction performance to predict hourly PM_2.5 from traffic camera images, with a root mean square error (RMSE) of 0.76 μg/m³ and a coefficient of determination (R²) of 0.98. The saliency map shows neural network focuses on unobstructed far-end road surfaces while the random forest feature importance highlights the first quarter image's significance. The model performance is robust whether weather conditions are controlled or not. Conclusion: Our study provided a practical approach to converting the existing traffic cameras into PM_2.5 sensors. The deep learning method based on the Resnet architecture in our study can broaden the coverage of PM_2.5 monitoring with no additional infrastructure needed.