Estimation of CO2 total columns from synthetic MicroCarb spectral images using machine learning

Following the Paris Agreement, there has been an increasing need to better understand the global sources and sinks of greenhouse gases (GHG). Most previous space missions aiming to measure GHG concentrations did so using point-based observations of the solar radiation reflected by Earth’s surface (corresponding to the short-wave infrared domain). These measurements – with resolutions in the tens of kilometers – allowed to estimate column average of CO₂ and CH₄ through the inversion of the radiative transfer equation.

This discrete measurement method presents nonetheless the disadvantage of being limited in spatial coverage and resolution, which entails that the satellite can miss important emission zones such as large cities or power plants. New GHG observation missions therefore use spectro-imagers to provide surface-based measurements: this is the case of the new GOSAT-GW satellite, as well as the exploratory mode (called city mode) of the new MicroCarb satellite. Both were launched in the summer of 2025 and have similar measurement techniques as the upcoming CO2M mission. They will allow to cover large areas of tens of kilometers consistently and with a kilometer-scale precision.

However, analysing such images with the traditionnal method based on optimal estimation is doomed to be too computationally expensive, and there is a need to develop a faster retrieval technique which would allow to reach the same precision level while keeping into consideration constraints such as instrumental noise and limited spectral bands.

The use of machine learning seems to present a convenient solution to this problem: by training neural networks with known atmospheric datasets, the machine learning model can learn to extract the column average of the gases with an inversion of the radiative transfer. This presentation thus aims to present a retrieval method based on a neural network designed using the radiative transfer model 4A/OP and the spectroscopic database GEISA, as well as atmospheric and surface variables and their effect on spectroscopy. The combination with a Bayesian approach will provide an uncertainty for the retrievals through the use of a Bayesian neural network. Exploiting the spatial information contained in the spectral images could further improve the retrieval process.

The method will be applied to two synthetic spectral images which could be observed by the MicroCarb city mode, representing emission plumes of a power plant near Berlin and a factory near Reims, and the performance of the retrieval process will be assessed using both systematic and random estimation errors. A perspective on future missions will also be presented.