S2WFS - Worldwide Sentinel-2 satellite imagery for wildfire detection

This dataset has been produced as part of FIRE-SAT collaborative project funded by the Austrian Research Promotion Agency (FFG). It has been used as a key contribution to the following publication: Xu, C., Rodriguez, N.D., Laux, D., Dornauer, V., Rashkovetsky, D., Langer, M., Ratschbacher, L. Manuscript submitted, Global Data Set for Early Active Wildfire Detection with RGB Satellite Images. Description: S2WFS is a dataset produced by Silicon Austria Labs GmbH, from Copernicus Sentinel-2 satellite images collected and labelled by OroraTech GmbH and their WildFire Service. The images contain Earth observations of worldwide landscapes with wildfires, collected from July 2020 to Oct 2021, as well as non-fire scenes of the same locations collected from 30 to 90 days before and after the wildfire occurred. Dataset Contents: The dataset presents the following files: "S2WFS_png.zip": ZIP file that containes the images and supplementary masks as PNG files, subdivided into classes. Img_png/: folder that contains 448x448 True Color Images (TCI) of both classes, collected from Sentinel-2 L2A products. Fire/: class folder with worldwide smoking wildfire scenes acquired from July 2020 to Oct 2021, named after their tile number (from UTM-UPS/WGS84 projection) and acquisition time. Non-fire/: class folder with scenes extracted from the same locations as "Fire" class images, 30-90 days before and after the wildfire occurred. "Non-fire" pictures are named after the corresponding "Fire" image, with an extra identifier to indicate whether the image represents a "before" or "after" wildfire scene. Msk_png/: class folder that contains supplementary masks compressed as PNG images, with one mask per channel. Mask images in class subfolders follow the same naming convention as their TCI couterparts. Fire/: class folder with Wildfire and cloud masks separated in channels as follows: "red" channel: binary fire mask provided by Ororatech WildFire Service. "green" channel: corrected fire mask (binary) using original "red" channel and manual annotations. "blue" channel: cloud mask (0-100 probability) obtained from Sentinel-2 data. Non-fire/: folder that presents mask for non-fire scenes, with only cloud mask stored as "blue" channel. "S2WFS_data.json": dictionary with metadata for each image in the dataset, where the keys are the names (without extension) of each sample. The metadata for each sample is given as another dictionary with the following keys: "nc_file": original netCDF4 file that contained Sentinel-2 product, from which the sample has been extracted "slice_num": number to identify the extracted slice from original TCI in Sentinel-2 product. "class": class identifier that can be "Fire" for wildfire class, "Non_fire_B" and "Non_fire_A" for "Non-fire" class ("before" and "after" the wildfire respectively). "linked_TP": only for "Non_fire" class, name used to identify the associated wildfire (True Positive) image. "coordinates": dictionary with geographic coordinates of the image's center point. "lat": latitude as a floating point value within the range [-90, 90]. "lon": longitude as a floating point value within the range [-180, 180]. "mission": Sentinel-2 identifier for the mission ID (S2A/S2B). "descriptor": Sentinel-2 identifier for the product level (MSIL1C for L1C, or MSIL2A for L2A). "time": Sentinel-2 identifier for datatake sensing start time (in the format YYYYMMDDHHMMSS). "baseline": Sentinel-2 identifier for the Processing Baseline number (in the format Nxxyy). "orbit": Sentinel-2 identifier for the Relative Orbit number (R001 - R143). "tile": Sentinel-2 identifier for the Tile Number field (in the format of Txxyyy with "xx" for longitude and "yyy" for latitude). "product": Sentinel-2 identifier for the Product Discriminator, used to distinguish between different end user products from the same datatake. To support the continent-specific splits of S2WFS dataset, given that spatially adjacent tiles may still exhibit considerable similarity, we used the location of the Sentile-2 product tile, denoted as ``Txxyyy'' in the product name (Naming Convention) to group tile images acquired from different continents with distinctive scenes. In this tile naming identifier, the two numeric digits (“xx’’) specify the longitude, whereas the three alphabetic characters (“yyy’’) correspond to the latitude. Table 1 presents continental grouping based on the tile naming identifier, using only the first alphabetical character for latitude (``y - -''). Although the resulting tile intervals form rectangular regions may extend beyond the actual continental boundaries, this approach substantially simplifies the implementation of the continent-wise image selection algorithm.  Potential Applications: Wildfire classification: the dataset can be used to train Machine Learning models to perform on-board satellite classification for early alert systems. Fire segmentation: with the fire mask annotation on pixel level, the dataset can be used to train Machine Learning models to perform fire segmentation for fire coverage estimation. Target Audience: The S2WFS dataset is supporting research studies, the development and benchmarking of wildfire detection algorithms using satellite imagery. Usage: S2WFS dataset is in PNG format (separated in subfolders for classes) and a JSON file for metadata, making it ready to use in any coding language that supports such formats.