Deep Learning-Based Method for an Assessment of Road Traffic Pollutant Estimation from Predicted Driving Behaviors

Nowadays air quality is a major public health issue. Transport, and especially road traffic is one of the main contributor of greenhouse gases and pollutant emissions. Improving air quality in urban areas implies reducing road traffic pollutant emissions. Conventionally, public authorities and air quality experts monitor road traffic emissions using macroscopic models based on emission factors that depend solely on average speed. Unfortunately, different road sections with the same average speed can exhibit significantly variable actual pollutant emissions. To address this bias, we propose a mesoscopic model based on deep learning to compute more precise emissions from infrastructure characteristics. This model utilizes neural networks that rely on raw scalar features from geographic information systems. The first neural network outputs a driving behavior distribution, while the second provides the mean and standard deviation values of five pollutant emissions per road link and vehicle. Our global method demonstrates real-time emissions estimation with an inference computation time of 2 minutes per one-hour step for a city the size of Lyon, France (40000 road links) and 30 representative vehicles. Our methods demonstrate the ability to accurately learn driving behavior, achieving an 81% accuracy on the test set. Moreover, by accounting for driving behavior, pollutant emissions are accurately estimated, resulting in a 35% reduction in the mean absolute error for the CO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> estimate compared to a macroscopic model.