Advancing Regional NO2 Estimation Using Machine Learning, TEMPO Satellite Data, and Geospatial Embeddings

Ground-level air quality affects human health, well-being, and social outcomes. Gold-standard monitoring relies on calibrated, ground-based sensors that can be costly to install and maintain, limiting spatial coverage especially in resource-constrained regions that face higher pollution burdens. Recent advances in satellite remote sensing and machine learning offer a pathway to address these limitations by extending the spatial and temporal coverage of monitoring networks. However, models trained at national or global scales often exhibit substantial variability in regional performance, particularly in environments that are data-sparse or with heterogeneous emission sources. In this work, we investigate the benefits of developing a regional model for the American Gulf South. We develop an ensemble machine learning model that integrates satellite-derived NO2 retrievals with auxiliary data sets, including pretrained AlphaEarth location embeddings, to incorporate globally learned spatial representations into a regional framework. We show that these embeddings improve model performance and enable the most detailed estimates to date: ∼1 km × 1 km estimates of daily maximum NO2 across the Gulf South. These results demonstrate how combining regional modeling approaches with global foundation model representations can improve air quality estimation in complex environments, providing new information to support exposure assessment and decision-making in under-monitored communities.