SegTHRawS: Thermal Hotspots segmentation in raw Sentinel-2 imagery

Overview This is the first publicly available segmentation dataset for thermal hotspot detection using raw, 10-band Sentinel-2 imagery. It is designed to support research in automated environmental monitoring and early-warning systems. Data Structure and Content The dataset is organized into 9 zip archives containing the images, segmentation masks, and training data. Image Categories: event: Confirmed thermal hotspot. non-event: Confirmed absence of a hotspot. potential event: Unconfirmed but potential hotspot. Note: The non-event images are split across multiple archives by band combination to manage file sizes. Band Organization: The 10 Sentinel-2 bands are grouped into the following combinations, each in its own folder: RGB: Bands 2, 3, 4 VNIR: Bands 5, 6, 7 NIR1: Band 8 NIR-SWIR: Bands 8A, 11, 12 train_geo_split_weakly_B12_B11_B8A_dataset.zip corresponds to the reduced dataset used for training the segmentation models, divided in three main groups: training, validation, and testing. The dataset is geographically splitted, geographical areas contained in training and validation are not included in the testing dataset. Technical Specifications All images and masks are stored in binary format (.bin). Image Shape: [1,256,256,3] Mask Shape: [1,256,256,1] Code Repository The Python code for generating and processing this dataset is available on GitHub: https://github.com/Ubotica/SegTHRawS_ext