Description

This repository contains the fully labelled multimodal images cropped from the Sentinel-5P TROPOMI level-2 Carbon Monoxide data. This dataset was used in Wąsala et al. 2025 for carbon monoxide (CO) plume detection with AutoMergeNet. The training data is saved per image in .npy format.

This dataset was generated using the TROPOMI CO data. The Copernicus Datahub provides access to this data. 

The dataset consists of 5207 images of the African content with 10 data layers. 24% of the images show plumes detected by Leguijt et al. 2024. The remaining images were selected randomly over the African content, using heuristics to discard potential plumes. Please see Wąsala et al. 2025 for a detailed description of how the dataset was compiled.

The images have 10 data layers: the CO data layer and 9 data layers that support the classification, making this dataset suitable for multimodal image classification or data fusion. These data layers are extracted from the operational (offline) TROPOMI CO product and are, therefore, already co-located.

The data layers are the following (see TROPOMI CO product for more detailed descriptions, variable name in the dataset indicated in brackets):

• Carbon monoxide total column (co): [ppb] the carbon monoxide concentration [mol/m2] converted to ppb and destriped to remove striping caused by the satellite sensor.
• Carbon monoxide standard error (co_precision): [mol m-2] the uncertainty in the retrieval of the carbon monoxide concentration.
• Data quality assurance (qa): [-] a flag indicating the quality of each pixel, based on observation conditions and output of the carbon monoxide retrieval algorithm.
• Geolocation flag (geolocation_flag): [-] flags describing the influence of the solar angle.
• Surface pressure (surface_pressure): [hPa] surface pressure in hPa.
• Degrees of freedom (dof): [-] degrees of freedom for the CO signal.
• Height scattering layer (height_scattering_layer): [m]  Scattering layer height above the topographic surface.
• SWIR scattering optical thickness (scattering_optical_thickness): [-]  Scattering optical depth in the SWIR channel at 2330 nm.
• Surface Albedo in the SWIR channel (albedo): [-] surface albedo at the wavelength of 2335 m.
• Ground pixel (ground_pixel): ground_pixel is the dimension perpendicular to the flight direction, comparable to the x-axis but in the satellite’s frame of reference. This measure indicates how close to the edge of the satellite’s swath a particular pixel is.

The annotations file contains the latitude and longitude of the top right corner of each image. Furthermore, Wąsala et al. defined a location-based test split. The images belonging to this test split are marked in the “test” column of the annotations file and “test” attribute in the NetCDFs (where 0=False and 1=True).