MSR-BACD-00

We have made both the dataset and the code publicly available on Zenodo (https://doi.org/10.5281/zenodo.15788378). To facilitate inspection of the dataset quality, we also provide burned area vector overlays on pre-fire and post-fire true-color and false-color composite imagery. Specifically, the following materials are included in our Zenodo repository: (i) Burned Area Vector Overlays:We provide vector files of burned area overlays aligned with Sentinel-2 imagery to enable intuitive visual inspection of annotation quality. These files include: 2019shp_FCI_TCI.zip, 2020shp_FCI_TCI.zip, and 2021shp_FCI_TCI.zip. Due to the large file sizes, we used multi-part compression to facilitate upload and download. For example, 2019shp_FCI_TCI.zip consists of 14 split archives: the main file 2019shp_FCI_TCI.zip and parts 2019shp_FCI_TCI.z01 through 2019shp_FCI_TCI.z13. In total, there are 42 image files across the three years, all in PNG format. These materials are available at https://doi.org/10.5281/zenodo.15767195. (ii) Model Code and Pretrained Weights:We provide the full implementation of the model along with the pretrained weights to help readers better understand the model architecture. The files are packaged in MSR-BACD model.zip, available at https://doi.org/10.5281/zenodo.15770516. The /data/after/ folder contains post-fire samples in TIFF format, each with 10 bands. The /data/before/ folder contains pre-fire samples, also in TIFF format with 10 bands. The /data/label/ folder contains the corresponding labels in PNG format. In the /logs/ folder, the files B12118A432.pth and B12118A8765432.pth are pretrained weights for 6-band and 10-band inputs, respectively. The 6 bands correspond to Sentinel-2 MSI bands B12, B11, B8A, B4, B3, and B2; the 10 bands include B12, B11, B8A, B8, B7, B6, B5, B4, B3, and B2. The code parameter selected_nums = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9] represents these 10 Sentinel-2 bands in order. To use the 6-band weight file (B12118A432.pth), set selected_nums = [0, 1, 2, 7, 8, 9]. Notably, the shared model supports 20-band input (10 pre-fire + 10 post-fire) to allow flexible experimentation with various band combinations—users can simply adjust the selected_nums parameter. (iii) Burned Area Vector Samples and Acquisition Metadata:To help users locate the corresponding Sentinel-2 tiles, we provide vector files that include acquisition dates for both pre-fire and post-fire imagery. A total of 11 files have been uploaded (available at: https://doi.org/10.5281/zenodo.15770516), including: 2019shp_BAs.zip, 2019shp_data.zip, 2019shp_Extent.zip 2020shp_BAs.z01, 2020shp_BAs.zip, 2020shp_data.zip, 2020shp_Extent.zip 2021shp_BAs.z01, 2021shp_BAs.zip, 2021shp_data.zip, 2021shp_Extent.zip The XXXXshp_data.zip files contain text files (.txt) listing the acquisition dates of pre- and post-fire imagery for each sample. For example, 10000007_0_20190126.txt indicates that the image for sample 10000007_0 was acquired on January 26, 2019.The XXXXshp_Extent.zip files contain the spatial extent of each sample in shapefile (.shp) format.The XXXXshp_BAs.zip files contain shapefiles delineating the burned area boundaries. (iv) Negative Sample Vector Data:We also provide vector files for negative samples used during dataset construction, along with the corresponding acquisition dates. A total of six files are included (available at: https://doi.org/10.5281/zenodo.15770516): Land cover background samples and acquisition times: Negative_sample1.zip, Negative_sample1_time.zip Land cover change samples and acquisition times: Negative_sample2.zip, Negative_sample2_time.zip Cloud-contaminated samples and acquisition times: Negative_sample3.zip, Negative_sample3_time.zip The Negative_sampleX.zip files contain shapefiles that define the spatial extent of the samples.The Negative_sampleX_time.zip files contain text files listing the acquisition dates. For example, 1000000_9_3_20190404.txt refers to sample 1000000_9_3, with the image acquired on April 4, 2019. Important Note:The entire dataset is approximately 2.6 TB in size. Due to Zenodo's upload limitations, we are unable to include all original Sentinel-2 imagery. Therefore, we provide the burned area vector extents and corresponding acquisition dates, which users can use to retrieve the raw imagery via Google Earth Engine for sample reconstruction. Due to the large number and size of files, we used batch compression to accelerate the upload process. Even so, the process was time-consuming, and we carefully verified the files to ensure accuracy. However, network issues may still result in incomplete uploads. If you encounter any problems while using the dataset, please do not hesitate to contact us at [e-mail]. We will respond as soon as possible. Thank you for your understanding and support.