KMaras-MSBD

Multi-Source Building Damage Dataset: 2023 Turkey Earthquake 1. Overview This dataset provides high-resolution paired optical and multispectral imagery for building damage extraction. It focuses on the catastrophic earthquake that struck southern Turkey in February 2023, offering a benchmark for disaster situation generation and rapid damage assessment. 2. Dataset Specifications Sensor: WorldView-2 (WV-2) Source: Maxar Open Data Program (ODP; now under Vantor) Total Size: 2.3 GB Total Channels: 11 Channels (3-channel High-Res RGB + 8-channel Multispectral) Spatial Configuration: RGB Patches: $512 \times 512$ pixels (High-resolution texture) MS Patches: $256 \times 256$ pixels (Spectral-rich information) Note: MS patches are resampled to exactly half the resolution of RGB patches to support multi-scale feature fusion. 3. Event & Geographic Coverage Disaster Event: Kahramanmaraş Earthquake (Feb 06, 2023) Imagery Dates: Feb 08, 2023 (Post-event) and Feb 11, 2023 (Post-event) Study Areas: Representative urban and suburban scenes from: Kahramanmaraş (Epicenter area) Nurdağı (Severely impacted) Antakya (Significant structural collapse) 4. Preprocessing & Technical Details To ensure high data quality for the Transformer architecture, the following steps were performed: Pan-sharpening: Gram-Schmidt (GS) pan-sharpening was applied to the Multispectral (MS) bands using the Pan-chromatic band. Resampling: The MS imagery was resampled to a spatial resolution exactly 50% of the RGB product to maintain a consistent $2:1$ scale ratio. Patch Extraction: * RGB imagery was cropped into $512 \times 512$ patches. Corresponding MS imagery was cropped into $256 \times 256$ patches to ensure perfect spatial alignment. Data Stacking: The resulting tensors contain 11 bands: Coastal, Blue, Green, Yellow, Red, Red Edge, NIR1, NIR2 (from MS) and the high-res R-G-B. 5. Annotation & Classification The dataset follows a 3-class semantic segmentation protocol, manually annotated by experts: Class 0 (Background): Roads, vegetation, water, and open land. Class 1 (Damaged Building): Buildings showing structural collapse, roof displacement, or significant debris. Class 2 (Intact Building): Structurally sound buildings with no visible damage. 6. Folder Structure /KMaras-MSBD├── train_images_RGB/ # 512x512├── train_images_MS/ # 256x256├── train_labels/ # 512x512├── valid_images_RGB/├── valid_images_MS/├── valid_labels/├── test_images_RGB/├── test_images_MS/└── test_labels/ 7. Contact & Support Principal Investigator: Longkun Zhang (张龙坤) Email: [e-mail] License & Attribution Data Source The raw satellite imagery is provided by the Maxar Open Data Program (Vantor). We gratefully acknowledge Maxar/Vantor for making this data available to the disaster response and research community. Licensing Imagery: Licensed under CC BY-NC 4.0. Terms of Use By downloading this dataset, you agree to: Provide appropriate credit to Maxar/Vantor and the authors of this dataset. Use the data for non-commercial purposes only.