Comment on amt-2023-241

Abstract. The largest anthropogenic emissions of carbon dioxide (CO2) come from local sources such as cities and power plants. The upcoming Copernicus CO2 Monitoring Mission (CO2M) will provide satellite images of the CO2 and NO2 plumes associated with these sources at a resolution of 2 km × 2 km and with a swath of 250 km. These images could be exploited with atmospheric plume inversion methods to estimate local CO2 emissions at the time of the satellite overpass and the corresponding uncertainties. To support the development of the operational processing of satellite column-average XCO2 and NO2 imagery, this study evaluates “data-driven inversion methods”, i.e., computationally light inversion methods that directly process information from satellite images, local winds and meteorological data, without resorting to computationally expensive dynamical atmospheric transport models. We have designed an objective benchmarking exercise to analyse and compare the performance of five different data-driven inversion methods: two implementations with different complexity for the cross-sectional flux approach (CSF and LCSF) and one implementation for the Integrated Mass Enhancement (IME), the Divergence (Div) and the Gaussian Plume model inversion (GP) approaches. This exercise is based on pseudo-data experiments with simulations of synthetic “true” emissions, meteorological and concentration fields, and CO2M observations in a domain of 750 km × 650 km centred on Eastern Germany over 1-year. The performance of the methods is quantified in terms of accuracy in the single-image (from individual images) or annual average (from the full series of images) emission estimates and in terms of number of instant estimates for the city of Berlin and 15 power plants in this domain. Several ensembles of estimations are conducted, using different scenarios for the available synthetic datasets. These ensembles are used to analyse the sensitivity of the performance to the loss of data due to cloud cover, to the uncertainty in the wind or to the added value of simultaneous NO2 images. The GP and the LCSF methods generate the most accurate estimates from individual images with similar Interquartile Ranges (IQR) in the deviations between the emission estimates and the true emissions between ~20 % and ~60 % for all scenarios. When taking the cloud cover into account, these methods produce respectively 274 and 318 instant estimates from the ~500 daily images that cover significant portions of the plumes from the sources. Filtering the results based on the associated uncertainty estimates can improve the statistics of the IME and CSF methods, but at the cost of a large decrease in the number of estimates. Due to a reliable estimation of uncertainty and thus a suitable selection of estimates, the CSF method achieves similar if not better statistics of accuracy for instant estimates compared to the GP and LCSF methods after filtering. In general, the performances for retrieving single-image estimates are improved when, in addition to XCO2 data, collocated NO2 data are used to characterise the structure of plumes. With respect to the estimates of annual emissions, the root mean square errors (RMSE) are for the most realistic benchmarking scenario 20 % (GP), 27 % (CSF), 31 % (LCSF), 55 % (IME) and 79 % (Div). This study suggests that the Gaussian plume and/or the cross-sectional approaches are currently the most efficient tools to provide estimates of CO2 emissions from satellite images and their relatively light computational cost will enable analysis of the massive amount of data provided by future missions of satellite XCO2 imagery.