Multitask Learning for Estimating Power Plant Greenhouse Gas Emissions from Satellite Imagery

<strong>Power Generation Data Set</strong> This data set contains imaging data acquired by ESA's Sentinel-2<br> Earth-observing satellite constellation [1] for a sample of power stations that were picked using geographic coordinates <br> provided by the European Pollutant Release and Transfer Register [2]. The images<br> contain scenes of power stations, some of which are actively<br> emitting smoke plumes. This data set was created with the goal to automatically segment plumes, predict the type of fired fuel, predict the rate of power generation and estimate the amount of CO2 emissions, directly from remote sensing images. <br> <strong>Description</strong><br> Each image is provided in the GeoTIFF file format, contains a total of 13 bands. Images have either a shape of 120x120 or 300x300 pixels (corresponding to a square area with an edge length of respectively 1.2 km and 3.0 km on the ground)<br> . This repository contains a total of 2131 images. This<br> repository contains a collection of JSON files that hold manual segmentation labels for plumes. Segmentation<br> labels were generated using label-studio [3]. Please note that polygon edge coordinates have to be scaled to fit the images. <br> <strong>Content</strong> The following files are contained in this repository: README.md - this file images.zip [2.0GB] - contains 2131 GeoTIFF images segmentation_labels.zip [1.5MB] - contains 2131 JSON files labels.csv [310KB] - contains additional labels for each image: Generation output rate [4],[5] Country Type of fired fuel Latitude and longitude of the power plant Concurrent weather information (temperature, humidity and wind vector) <strong>Acknowledgement</strong> If you use this data set, please cite our publication: Hanna, J., Mommert, M., Scheibenreif, L., Borth, D.,<br> "Multitask Learning for Estimating Power Plant Greenhouse Gas Emissions from Satellite Imagery",<br> Tackling Climate Change with Machine Learning workshop at NeurIPS 2021. Please refer to this publication for additional information on the data set. The code used for this publication is available at https://github.com/HSG-AIML/RemoteSensingCO2Estimation. <br> <strong>Author</strong> Joëlle Hanna University of St. Gallen, AIML Lab, School of Computer Science [e-mail] <br> <strong>References</strong><br> [1]: https://earth.esa.int/web/sentinel/missions/sentinel-2<br> [2]: https://www.eea.europa.eu/data-and-maps/data/industrial-reporting-under-the-industrial<br> [3]: https://labelstud.io/<br> [4]: https://transparency.entsoe.eu/generation/r2/actualGenerationPerGenerationUnit/show<br> [5]: https://doi.org/10.5281/zenodo.3574566