DeepAQ: Unsupervised Domain Adaptation for Air-Quality Mapping Using High-Resolution Satellite Imagery

High-resolution urban air quality (AQ) estimation is critical for addressing problems such as quantifying the health impacts of air pollution and devising mitigation strategies, among others. Traditional methods rely on AQ monitoring stations on the ground. However, many cities throughout the world lack adequate coverage of such stations. Advances in machine learning (ML) combined with the availability of high-resolution satellite imagery at a global scale offer an alternative solution. Yet, generalization to unseen, often data-poor, domains (regions) remains a challenge. In this work, we propose a novel ML-based modeling approach using unsupervised domain adaptation (DeepAQ) for AQ mapping over cities lacking sufficient training data. We show that AQ (mean annual NO2 levels) models trained on data-rich cities such as Los Angeles and New York can be transferred even to cities in low-income countries such as Accra in Ghana, Africa, with a sigma-normalized RMSE of 0.66 and r2 score of 0.5. This work demonstrates the utility of ML methods in deriving predictive information from satellite imagery over regions with limited ground data, suggesting many potential applications across scientific domains.