Day 4

Detailed paper information

Back to list

Paper title Reducing uncertainties in annual CO2 point source emission estimates from CO2M CO2 and NO2 images using computer vision techniques and multi-plume models
  1. Erik Koene Empa - Swiss Federal Laboratories for Materials Science and Technology, Dübendorf, Switzerland Speaker
  2. Gerrit Kuhlmann Empa - Swiss Federal Laboratories for Materials Science and Technology
  3. Dominik Brunner Empa - Swiss Federal Laboratories for Materials Science and Technology
Form of presentation Poster
  • A1. Atmosphere
    • A1.04 Greenhouse Gases
Abstract text To support the ambition of national and EU legislators to substantially lower greenhouse gas (GHG) emissions as ratified in the Paris Agreement on Climate Change, an observation-based "top-down" GHG monitoring system is needed to complement and support the legally binding "bottom-up" reporting in national inventories. For this purpose, the European Commission is establishing an operational anthropogenic GHG emissions Monitoring and Verification Support (MVS) capacity as part of its Copernicus Earth observation programme. A constellation of three CO2, NO2, and CH4 monitoring satellites (CO2M) will be at the core of this MVS system. The satellites, to be launched from 2026, will provide images of CO2, NO2, and CH4 at a resolution of about 2 km × 2 km along a 250-km wide swath. This will not only allow observing the large-scale distribution of the two most important GHGs (CO2 and CH4), but also capturing the plumes of individual large point sources and cities.
Emissions of point sources can be quantified from individual images using a plume detection algorithm followed by data-driven methods computing cross-sectional fluxes or fitting Gaussian plume models. To estimate annual emissions, a sufficiently large number of estimates is required to limit the uncertainty due to the temporal variability of emissions. However, the number of detectable plumes is limited, because the signal-to-noise ratio of individual plumes is too low or because neighboring plumes are overlapping. We present methods for increasing the number of plumes available for emission quantification using computer vision technqiues and improved data-driven methods that can estimate emissions from overlapping plumes.
Using synthetic data generated in the SMARTCARB project (Kuhlmann et al., 2020), we show that a joint denoising of coincident CO2 and NO2 images can result in signficantly improved signal-to-noise ratios for the individual images (notably, improving the peak signal-to-noise ratio of the CO2 images by +13 dB). Furthermore, by using a generative adverserial neural network approach, we show that it is possible to fill in missing data due to, e.g., cloud cover, with as an additional input wind direction information to steer the interpolation for the missing data. This ‘inpainting’ method helps the segmentation step, as it becomes possible to connect otherwise disjoint parts of a plume. Finally, we show how plume detection may be improved to be particularly receptive to plume-like features on satellite images (e.g., stretched out and narrow enhancements over the background) using a method referred to as Meijering.
A remaining challenge is to quantify the emissions from overlapping plumes, e.g., those occurring when one point source lies in the downwind direction of another plume, or when two diffusive plumes are positioned close to each other. We developed a data-driven approach using a multi-plume model that alleviates this problem. First, the approach obtains a best fitting center line for each of the individual plume sources, using effective wind data information and the multimodal distribution in the CO2 and NO2 images as inputs. Once such center lines are available, a cross-sectional flux method assuming a Gaussian cross-sectional structure can be computed for the multiple plume sources simultaneously. The upstream part of the plume (prior to overlapping) can be used to constrain the estimated fluxes. An alternative solution is to find best-fitting parameters for two or more Gaussian plume models simultaneously to estimate the emissions of each point source.
The improvements in the plume detection algorithm, and the multi-plume models for estimating emissions of overlapping plumes, increase the number of satellite images from which emission can be quantified. The larger number of emission estimates reduces the uncertainties in estimated annual emissions for point sources.