Research Project

GeCO: Generating high-resolution CO2 maps by Machine Learning-based geodata fusion and atmospheric transport modelling

The spatiotemporal distribution of greenhouse gases and their sources on Earth has so far been considered mainly at relatively coarse resolutions. There is a lack of sound information on local emissions and their spatiotemporal distribution. However, these data are urgently needed to design local options for action in climate change mitigation and to validate mitigation efforts.

In the GeCO project, significant progress will be achieved by very high-resolution detection and modeling of the sources of selected greenhouse gases, as well as their turbulent dispersion in the atmosphere. Both the modeling of the distribution of the sources, as well as the dispersion of the gases will be done by means of computer-aided data-intensive methods through collaboration of expertise from environmental physics and geoinformatics. Current methods such as machine learning (ML) and geospatial data fusion are used. The computer computation time is kept low despite the high resolution by using a so-called catalog approach, which enables the dispersion calculation of greenhouse gases over longer time periods despite the high resolution.

Partners from geoinformatics and environmental physics are working in a team on data-driven methods to analyze the sources of greenhouse gases on the Earth's surface and their turbulent dispersion in the atmosphere. The heterogeneous database for the sources of greenhouse gases on the Earth's surface consists of diverse sensor readings, citizen science observations, the social web, and official and other statistical data. The discovered patterns are validated by measured values. Machine Learning (ML) methods are used and further developed for the analysis and fusion of the heterogeneous input data and generation of spatially (and temporally) highly resolved emission inventories of relevant greenhouse gases.

The GIScience Research Group develops the methods for data preparation for the emission inventories using methods of Spatial Data Science and ML (esp. deep learning). For this purpose, different geodata sets, in particular on land use and other relevant sources of greenhouse gases (industry, transport, housing, waste, agriculture, etc.) will be analyzed. An important data source is OpenStreetMap (OSM), which contains high-resolution spatial information on buildings, industrial facilities, transport infrastructure and also land use information, which provides the basis for the localization of emissions. Challenges include the spatially heterogeneous data quality of OSM on the one hand, and the spatially heterogeneous emission factors on the other hand. The question of data quality is investigated and used for the evaluation of the OSM objects.

The team at the Institute of Environmental Physics (IWR, HCE) uses the resulting high-resolution emission maps to set up turbulent dispersion calculations for the dispersion of greenhouse gases in the atmosphere on top of them, and to compare them at the end with in-situ concentration measurements for validation purposes.