Predicting atmospheric particle number concentration from roadway surveillance video

Abstract:

Traffic-related air pollution is a major environmental cause of death and disease in urban areas. Pollution measurement networks exist to monitor concentrations, but accurate sensors are expensive and the number of sensors deployed by the US federal government has been decreasing over recent years. Additionally, within-city spatial variation in pollution concentration is often poorly captured by routine monitoring networks. We explore the use of roadway surveillance data—which is widely collected—as an input into a machine-learned system for estimating near-roadway fine particle number concentrations.

We collect 15 hours of roadside particle number concentration and record corresponding road traffic video at a location 3–6 m from an arterial road in Champaign, IL. We collect particle number data with a TSI Optical Particle Sizer with 16 size bins ranging from 0.3 - 10 μm, and video data with a commercially available webcam. We additionally collect auxiliary data such as temperature, wind direction and speed, and distance from the roadway. We use the video and auxiliary data to train and test a machine learning model to predict the measured particle number concentration. The model consists of two deep neural network components. The first component extracts features from the frames of the input video. The second component combines those features and the auxiliary data to form the regressor that gives the final prediction. We use 80% of the data to train the model, 10% for validation set, and 10% to produce the results shown in Figure 1.

The model was able to predict concentration measurements with R² = 0.85 and MAE = 4.0 #/cm3. Our initial results show the potential for predicting particulate matter counts from readily available data. Additional analysis is required to explore model performance over a wider range of locations and conditions.