KMaras-MSBD

KMaras-MSBD Dataset KMaras-MSBD is a post-earthquake building-damage segmentation dataset derived from multi-source WorldView-2 satellite imagery collected after the 2023 Kahramanmaras earthquake in southern Turkey. It was created for research on collapsed-building mapping using very-high-resolution RGB imagery and 8-band multispectral data. This dataset supports the study reported in the MS-SwinNet paper and is intended for semantic segmentation of three classes: background, collapsed/severely damaged buildings, and intact buildings. Data Source The imagery was collected from the Maxar Open Data Program using post-event WorldView-2 acquisitions from February 8 and 11, 2023, corresponding to approximately 2-5 days after the mainshocks. The original data include: 3-band RGB imagery at about 0.30 m spatial resolution Panchromatic imagery 8-band multispectral imagery at about 2.56 m spatial resolution The multispectral bands extend beyond visible light into the Red Edge and Near-Infrared (NIR) ranges, which are helpful for distinguishing damaged building materials from intact roofs, soil, and surrounding background. Study Area and Annotation The dataset focuses on earthquake-affected urban regions in the Kahramanmaras event. Ground-truth masks were manually delineated at the pixel level with support from pre- and post-earthquake high-resolution imagery and official disaster assessment references. Each pixel is labeled as: 0: Background 1: Collapsed or severely damaged building 2: Intact building Dataset Construction To reduce spatial leakage, the original large scenes were first divided into non-overlapping Geographic Units (GUs) of 2048 x 2048 pixels. GUs without collapsed-building pixels were removed, and only damage-relevant regions were retained. Patch generation was then performed using a sliding-window strategy: Patch size: 512 x 512 Overlap: 25% Stride: 384 After removing boundary-adjacent patches to avoid overlap between neighboring GUs, the final dataset contains 826 valid sample pairs. Official Split The dataset is split at the GU level, not at the patch level, to prevent geographic leakage: Train: 571 patches Validation: 134 patches Test: 121 patches This design ensures that samples from different subsets are geographically disjoint and non-adjacent. Preprocessing Variants This release provides several preprocessing variants of the same core dataset. They differ in how the multispectral data were sharpened or resampled before being paired with RGB imagery. Included Variants KMaras_v1_gs_256 Gram-Schmidt pan-sharpened multispectral patches. KMaras_v2_gsa_256 GSA-based pan-sharpened multispectral patches. KMaras_v3_bicubic_256 Multispectral patches generated with bicubic resampling. KMaras_v4_bilinear_256 Multispectral patches generated with bilinear resampling. KMaras_v5_pannet_256 Multispectral patches generated using the PanNet pan-sharpening method. Notes All variants correspond to the same earthquake case and annotation system. The main difference between folders is the preprocessing strategy applied to the multispectral modality. Users should keep the train/validation/test split unchanged for fair comparison with the paper. Because pan-sharpening and resampling may affect spectral fidelity and spatial alignment, results from different variants may not be directly interchangeable. Recommended Use This dataset is intended for research in: Post-disaster damage assessment Building-damage semantic segmentation Multi-source or multispectral remote sensing Image fusion and pan-sharpening evaluation Transformer- or CNN-based dense prediction Citation If you use this dataset, please cite the corresponding MS-SwinNet paper and acknowledge the KMaras-MSBD dataset release. License Please refer to the Zenodo record for the official license and distribution terms. Acknowledgment We gratefully acknowledge Maxar Open Data for providing the post-earthquake satellite imagery used to construct this dataset.