Location generalizability of image-based air quality models

Eleanor Byler, Christian Svinth, Kirsten Chojnicki; Proceedings of the Winter Conference on Applications of Computer Vision (WACV) Workshops, 2025, pp. 1181-1190

The ability to rapidly quantify atmospheric pollutants is important both for global emissions monitoring and for mitigating the adverse effects that follow a hazardous chemical release. In the aftermath of a chemical release imagery is often the only available resource to assess local conditions. Recent work has demonstrated initial success in predicting particulate matter pollution from imagery; however these results are tied to a specific site and do not generalize to new geographic locations. In this work we seek to understand how easily deep learning models generalize to new locations in the context of image-based air quality assessments targeting two distinct tasks: (1) broad measures of particulate matter pollution and (2) the mass of a given chemical released in hazardous plumes. For the latter we focus on sulfur dioxide a toxic aerosol and a major component of particulate matter pollution caused by industrial fossil fuel consumption. To develop a model that operates in the widest possible range of environments we test different training strategies including the use of new geolocation foundation models. The best performing models achieve >80% accuracy when evaluating unseen imagery at previously seen sites but we find significant drops in performance when evaluating imagery from unseen sites at best 65%. Additionally we present the public release of the National Parks Air Quality Index Dataset a new medium-sized dataset that pairs imagery with sensor-based air quality measurements at 15 different national parks.