A Transformer-based approach to air quality prediction in Milan through satellite imagery combined with meteorological and morphological data

Air pollution is one of the most pressing problems of our time, causing serious issues for human health and the environment. It is a key factor in the development of respiratory ailments, ranging from mild conditions like asthma to more severe complications such as decreased lung function, cardiovascular diseases, and premature mortality. Additionally, it has serious environmental repercussions, such as ecosystem degradation, biodiversity loss, and adverse effects on plant growth and agricultural productivity. To better understand and mitigate this problem, remote sensing technologies have been adopted to monitor large areas, facilitating comprehensive air quality assessments. For example, satellite data from the Copernicus Sentinel-5p mission provide valuable information and are useful for forecasting pollutants. In this study, Copernicus Sentinel-3 and Sentinel-5p data combined with other meteorological and morphological data were used to predict the air quality in the city of Milan, focusing on five pollutants: ozone (O3), nitrogen dioxide (NO2), sulfur dioxide (SO2), particulate matter with a diameter of less than 10 micrometers (PM10), and less than 2.5 micrometers (PM2.5). Deep learning models, such as CNN, RNN, and LSTM, have demonstrated considerable success in analyzing such data due to their ability to uncover complex relationships. Expanding on this success, this study explores the adaptation of TimeSformer, a transformer-based model initially designed for video classification, to work on satellite data for the forecasting task. The primary modification is the initial encoding, which is designed to accommodate diverse sources, reflecting the heterogeneous nature of the dataset. The developed model is capable of forecasting all pollutants simultaneously, creating a forecast mechanism that could be of significant value to policymakers, as it enables them to implement timely protection and prevention measures. It achieved a Mean Absolute Percentage Error of 30.6%, improving the prediction accuracy for three out of the five pollutants compared to a baseline model.