Commercial Vehicle Traffic Detection from Satellite Imagery with Deep Learning

Road freight traffic is a major greenhouse gas emitter: commercial vehicles (CVs) contribute ∼7% to the global CO 2 emission budget, a fraction that is likely to increase in the future. The quantitative monitoring of CV traffic rates, while essential for the implementation of targeted road emission regulations, is costly and as such only available in developed regions. In this work, we investigate the feasibility of estimating hourly CV traffic rates from freely available Sentinel-2 satellite imagery. We train a modified Faster R-CNN object detection model to detect individual CVs in satellite images and feed the resulting counts into a regression model to predict hourly CV traffic rates. This architecture, when trained on ground-truth data for Switzerland, is able to estimate hourly CV traffic rates for any freeway section within 58% (MAPE) of the actual value; for freeway sections with historic information on CV traffic rates, we can predict hourly CV traffic rates up to within 4% (MAPE). We successfully apply our model to freeway sections in other countries and show-case its utility by quantifying the change in traffic patterns as a result of the first COVID-19 lockdown in Switzerland. Our results show that it is possible to estimate hourly CV traffic rates from satellite images, which can guide civil engineers and policy makers, especially in developing countries, in monitoring and reducing greenhouse gas emissions from CV traffic.