CUGMine: A Fine-Grained Remote Sensing Dataset for Unravelling Open-Pit Mine Ecological Pattern Evolution

It is designed to train and evaluate semantic segmentation models for land cover classification, particularly in regions affected by mining activities.<br><b>Geographic coverage:</b> Yunnan Province, China<br><b>Image source:</b> Google Earth<br><b>Spatial resolution:</b> approximately 1.1-1.2 m (pixel scale 1.07e-5 deg, WGS 84)<br><b>Tile size:</b> 512 x 512 pixels; the training subset additionally contains 966 tiles of 256 x 256 pixels at the same resolution<br><b>Time periods:</b> 2016 and 2022<br><b>Total number of samples:</b> 2,656 image-label pairs (train 2,010 / val 322 / test 324)<br><b>File format:</b> GeoTIFF in the WGS 84 geographic coordinate system (EPSG: 4326), with a per-tile longitude-latitude tie point; images are 8-bit RGB, label maps are single-band integer masks; each image-label pair shares the same numeric file name (1.tif, 2.tif, ...)<br><b>Label encoding (pixel values 0-5, fixed order):</b><br>0 - Grassland<br>1 - Forest<br>2 - Buildings<br>3 - Road<br>4 - Bareland<br>5 - Water<br>Pixel value 0 is the class Grassland; there is no background or no-data value.<br><b>Typical applications:</b> land degradation monitoring; mining environmental impact assessment; model benchmarking for semantic segmentation.<br>All annotations have been manually reviewed.

The dataset is intended for research purposes under the CC-BY 4.0 license.</p>