Development of geospatial AI and deep learning models to estimate high-resolution multipollutant ground air quality using satellite remote sensing at hyperlocal level

Abstract

Tracking air quality in urban areas is particularly important for the continued urbanization of civilization. Urban air quality is decreasing due to natural and anthropogenic events, causing living quality degradation and giving rise to various population health issues. Air quality monitoring is heavily dependent on the sparse and expensive network of ground-based sensors. Due to the scarcity of ground sensors and the resultant lack of spatiotemporal coverage, it cannot capture the spatial and temporal variations and fails to delineate localized zones of heightened pollution within urban settings. Satellite remote sensing offers a promising alternative by providing continuous spatial coverage. Recent advancements in computing have enabled widespread use of deep learning in remote sensing tasks, including air quality estimation. This remote sensing-based approach of air quality modeling, with or without applications of deep learning, has predominantly been directed at national or regional assessments, primarily due to the reliance on images of lower resolution. Utilizing dense ground sensor networks data from the Microsoft Eclipse project, along with high-resolution Sentinel-2 and low-resolution tropospheric measurements from Sentinel-5P, this research contributes to urban air quality estimation in three steps. First, it explores spatiotemporal dynamics at a finer level, which provides the necessary background for developing models for air quality estimation. Second, it develops deep learning models using transfer learning and convolutional neural network architectures to estimate ground-level air quality at the hyperlocal level. Finally, it enhances state-of-the-art transformer-based deep learning models by incorporating spatiotemporal dependencies directly into the modeling frameworks. The results demonstrate that integrating multi-source satellite data with geospatial deep learning significantly improves the accuracy and spatial granularity of air quality estimation. This research highlights the potential of remote sensing and deep learning as scalable and cost-effective alternatives for urban air quality monitoring, particularly in regions where ground-based measurements are limited or unavailable.