Improving OMI-NO2 spatial resolution using a stochastic convolutional neural network over central southern Chile

Abstract

Both precision and spatial resolution of atmospheric data are fundamental aspects when evaluating air quality and the effects of different factors on the distribution of air pollutants. Satellite-borne instruments have been very useful to collect this type of data, providing valuable information for process understanding and when deciding public policies of both regional and national relevance. However, as sensors and satellites retrievals have improved, one finds discontinuities in quality and resolution that require homogenisation for establishing long-term trends. This work seeks to assess the feasibility of improving the spatial resolution of satellite measurements while, simultaneously estimating the expected error. A stochastic approach based on Convolutional Neural Networks is presented, which succeeds in increasing the spatial resolution of nitrogen dioxide (NO2) columns collected with spectrometers onboard satellites. Furthermore, the approach allows estimating the aleatoric uncertainty generated by errors inherent in spectrometer measurements. Also, the methodology allows achieving the estimation precision of conventional deep learning models. The results show that the reconstructed fields are robust to added noise on the data, presenting slight decreases in the evaluated metrics above 5% noise. An advantage of the presented methodology is that the models can be trained with small-scale images, and then applied without domain restriction, if the resolution used during training is maintained. The results indicate that the methodology is appropriate for the stated objectives. Also, it is subject to further improvement by considering state-of-the-art models (ResNet, GAN). An application to reconstruct NO2 column data from the Ozone Monitoring Instrument (OMI) onboard the Aura satellite is shown, illustrating the potential of the methods. Thus, this work contributes to an improvement in the monitoring of air quality for the country, and it is expected that it can be applied to obtain better prediction results (both precision and error estimation) and cover a larger area of application.