U-net for learning and inference of dense representation of multiple air pollutants from satellite imagery

ABSTRACT

Air pollution is an important topic on countless fronts and is an active area of research. The goal of this work is to provide a machine learning model for learning and inference of pollution concentrations and air quality measures, namely Particulate Matter 2.5, NO3, Nitrate Pollution, and NH4, Atmospheric Ammonium, with high granularity by using easily obtainable satellite imagery data. In order to achieve this, we propose the fully convolutional network U-net that, unlike previous work, can predict these pollutant values at a pixel-level high-resolution instead of being able only to predict a single value for an entire geographical region. We demonstrate that this approach can reconstruct the considered pollutant concentrations on ground-truth data and can predict the concentrations and their spatial structure reasonably well, even for data that the network has temporally not yet seen. Finally, we illustrate that the model's pollutant predictions can offer valuable insights into the current COVID-19 pandemic.