Extended Large-Scale Semantic Change Detection (LsSCD-Ex) dataset

LsSCD-Ex comprises remote sensing images captured in September 2013 and August 2015, with a spatial resolution of 0.6 m and a total size of 48,000 × 32,500 pixels (approximately 471 km²) in Nanjing city, China. Its semantic classification system aligns with OpenEarthMap, encompassing eight categories: bare land, rangeland, developed space, road, tree, water, agriculture, and building. The number of annotated image pairs (2048 × 2048 pixels) is 100, including 70 for training and 30 for testing, with no spatial overlap between sets to ensure reliable evaluation. All images were labeled by professional remote sensing experts using QGIS geospatial software. The dataset captures complex urban-rural land-cover transitions and severe atmospheric style differences, providing a challenging benchmark for change detection. 256.zip, 512.zip, and 2048.zip represent datasets of different patch sizes, obtained by cropping the original images using non-overlapping sliding windows Citation: @article{tang2026dreamcd, title = {DreamCD: A Change-Label-Free Framework for Change Detection via a Weakly Conditional Semantic Diffusion Model in Optical VHR Imagery}, author = {Tang, Kai and Zheng, Zhuo and Chen, Hongruixuan and Chen, Xuehong and Chen, Jin}, journal = {International Journal of Applied Earth Observation and Geoinformation}, volume = {146}, pages = {105125}, year = {2026}, issn = {1569-8432}, doi = {10.1016/j.jag.2026.105125},} 1. Dataset Splits train — Training samples test — Testing samples 2. Folder Structure Patch Scale Folders: 256, 512, 2048 These folders represent different patch sizes. 512 patches are cropped from the 2048 patches using non-overlapping sliding windows. 256 patches are generated using the same method. Inside each scale folder: 2013 — Pre-event imagery 2015 — Post-event imagery bcd — Binary change masks (0 = no change, 255 = change) images — Image files semantic_masks — Semantic labels (integer encoded) semantic_masks_rgb — Semantic labels in RGB format 3. Geolocation Information The shpfile folder contains geolocation.shp, which provides the geographic locations of the image patches. The shapefile can be opened with standard GIS software. 4. Semantic Classes Integer-encoded semantic labels (semantic_masks): 0: Bare land 1: Rangeland 2: Developed space 3: Road 4: Tree 5: Water 6: Agriculture land 7: Building RGB-format semantic labels (semantic_masks_rgb): Bare land: [128, 0, 0] Rangeland: [0, 255, 36] Developed space: [148, 148, 148] Road: [255, 255, 255] Tree: [34, 97, 38] Water: [0, 69, 255] Agriculture land: [75, 181, 73] Building: [222, 31, 7] 5. License: Use of the dataset must respect the "Google Earth" terms of use. All images and their associated annotations in LsSCD-Ex can be used for academic purposes only, but any commercial use is prohibited. (CC BY-NC-SA 4.0)