CO2 plume detection and inversion using convolutional neural networks: application to synthetic images of XCO2 fields over urban areas

Carbon dioxide emissions, accounting for more than 70% of global anthropogenic greenhouse gas releases, are the main driver of climate change. Current emissions estimates, which are needed to guide reduction policies, rely on statistical data of energy consumption including self-reporting from emitters and are subject to important uncertainties. In order to assess these emissions in an independent, timely and accurate manner, the Copernicus CoCO2 project aims to build a prototype system for a CO2 emission monitoring service exploiting atmospheric CO2 measurements. As part of this project, our goal is to build an atmospheric transport modelling inverse system to improve the quantification of CO2 sources of large magnitude at urban scale based on the spaceborne imagery of the CO2 atmospheric plumes from these sources. The reconstruction of such sources depends on the detection of the associated plumes in the satellite images of the vertically averaged CO2 column concentrations (XCO2), which represents a significant challenge. Indeed, the signal of CO2 plumes induced by point-source emissions is intrinsically difficult to detect since it rarely exceeds values of a few ppm and is perturbed by variable regional CO2 background signals and noise or error patterns in XCO2 images due to instrument and retrieval algorithms. To tackle the problem of CO2 plume detection and inversion, we investigate the potential of deep learning methods. Neural networks are trained on hourly simulated XCO2 fields in the regions of Paris, Berlin, and several power plants, consisting of the plume from the city or the power plant and of other biogenic and anthropogenic fluxes. Convolutional neural networks are trained to evaluate the presence and the contour of the CO2 plume in an image and to reconstruct the corresponding emissions. In 75% of the estimates, the relative error between predictions and actual emissions is less than 0.2.