Predicting Atmospheric Air Pollution: A Convolutional- Transformer Approach for Spatial and Temporal Analysis of PM2.5

6.7 million premature deaths occur annually due to household and ambient air pollution. Air pollution affects individuals globally and can be derived from a variety of factors including household cooking fuel, motor vehicles, industrial practices, and natural fires. To tackle this global crisis, our research focuses on understanding the spatial and temporal patterns between air pollutants to predict future levels of air pollutants. Our approach uses a novel deep learning methodology that involves a spatiotemporal Convolutional-Transformer archi-tecture (ConvTransformer) We harnessed multiple data sources as inputs to our ConvTransformer, including several remote sensing instruments such as NASA's Terra+Aqua satellites with the Moderate Resolution Imaging Spectroradiometer (MODIS), and the European Space Agency's Sentinel-5P satellite with the Tropospheric Monitoring Instrument (TROPOMI), gridMET ground-based meteorological remote-sensing data from North American Land Data (NLDAS-2), US Census TIGER Roadways, and ground-level PM2.5 sensing data sourced from EPA AirNow. Our approach shows a 38.8% decrease in 5-frame average SSIM error compared to existing modern deep learning architectures utilizing remote-sensing data, satellite imagery, and ground-level data for PM2.5 prediction.