GeoCH4PlumeNet - Global dataset for Methane point source detection and quantification

Overview Purpose: Binary segmentation (methane plume binary segmentation), Regression (Estimation of emission rate)Source of data: Satellite provider – Sentinel 2Images: (512, 512, bands) window cropped out of a L1C tile, with all 13 Sentinel 2 bands saved in a GeoTiff, compressed using ZSTD.Labels: 0=no plume, 1=plume, saved in GeoTiff, compressed using ZSTD.Geographic coverage: Worldwide (oil and gas regions) defined as gographic points.Date of acquisition: Around April 2024 (depending on cloud coverage < 30%)Emissionrates: Provided in the Name of the files given the identifier “XX_Rzzzz…” where zzzz is the emission rate in kg/h.Quality control: Pumes are not overlapping, plumes are aligned with the main wind direction of the scene (disturbed sometimes), no plume over water bodies. What data is included Within this dataset are modified Sentinel 2 Images (all bands) with corresponding methane plume masks and respective emission rate resulting from simulations. The dataset contains 1000 samples. The data has been split into a training set (train) containing 70% of data, a validation set (val) containing 10% of data, and a test set (test) containing 20% of the data. For the split of training data is has been considered that: The split is 70 % train, 10 % val, 20 % test. One whole location (geographical point id) and at least one plume is only in the test split. The other plumes may be sorted in different splits. This ensures a good test for generalizability of the models. The validation set contains an own location (not in test or train set). All Sentinel 2 bands (also modified Band 11 – SWIR 1 and Band 12 – SWIR 2) are integrated into the “XX_nonorm_bands20m.tiff”. As the name suggests all bands are interpolated to 20m resolution using bilinear resampling. The values of the bands are unchanged – therefore roughly between 0 and 10000 as the data include L1C data. The binary labels (plume / no plume) are marked as “XX_bin_labels.tiff” and in a separate label folder within the train, val, test folders. The image data and label can be aligned by sorting or matching the identifiers (“XX…”), satellite and time. Overall, this dataset contains 1000 samples from 10 locations, 147 different plumes and mostly two timestamps per location. Further information Label created by using quantile of transmissivity only from the methane plumes (upper 70% are marked as plume). The transmissivity are derived by a simulated 3D concentration map of methane as input for a radiative transfer model using the most accurate atmospheric parameters of that time in that scene. The images and labels are georeferenced, and it is recommended to use GDAL or Rasterio for reading the data. The locations and exact filenames of the original satellite tiles are described in “Locations_and_tiles.csv” The different datasplits are packed in a .tar file to be uploaded here. Planned improvements More data samples, more locations, more different emission rates, include different Satellites, provide data loader, real detected plumes for comparison of models, further information on data generation in upcoming publication. Disclaimer This dataset is still under construction and will be adapted and enhanced in the future. In case you have any questions or requests: [e-mail] The dataset contains modified Copernicus Sentinel data [2024]. Funding This work has been funded by OHB Digital Connect GmbH and ESA via the Open Space Innovation Platform