Terrace dataset of the mountainous and hilly region of western Liaoning, China: GF-6 imagery, DEM and terrace labels for VEP-MDTNet

This dataset was used to train and evaluate VEP-MDTNet, a dual-modal deep learning model for terrace extraction described in "VEP-MDTNet: a deep learning model for terrace extraction in the mountainous and hilly region of western Liaoning, China" (Huang et al., under review). Study area: 22 small watersheds in Jianping County, Chaoyang City, Liaoning Province, China, including the Zhalanyingzi River headwater, Hongshan River, Dasanjia River and Bayan River headwater watersheds (about 550 km², mean elevation 690 m), a typical low-mountain and hilly region with complex topography and a dense network of gullies. Data sources: 2 m Gaofen-6 (GF-6) imagery acquired in spring 2024 and a 5 m digital elevation model (DEM) obtained from the Geographic Remote Sensing Ecological Network platform (www.gisrs.cn), resampled to the image grid. Slope and contour layers were derived from the DEM. Labels: 1,720 terraces were visually interpreted in ArcGIS 10.7. The dataset provides terrace masks and terrace boundaries. Tiles and splits: The study area was divided into 10 non-overlapping spatial subregions, which were assigned to the training, validation and test sets at a ratio of 7:2:1. Images, DEM and labels were cropped synchronously into 256 × 256 pixel GeoTIFF tiles with a sliding window and 20% overlap, and tiles with labelling errors or very low terrace coverage were removed. Each location is represented by the original tile and its horizontally and vertically flipped copies, giving 5,946 training, 1,701 validation and 873 test tiles; 153 of the test tiles contain no terraces. Folder structure:82lb/train/, 82lb/val/ and 82lb/test/ each contain image/, dem/, label/, boundary/, contour/ and slope/; files are matched by name across folders. 82lb/source/ contains full-area rasters (boundary.tif, contour.tif, dem_aligned.tif, slope.tif). Layers:- image: GF-6 imagery, 2 m, 4 bands, uint8 (VEP-MDTNet uses the first three bands, RGB)- dem: 5 m DEM resampled to the image grid, float32, elevation in m- label: terrace mask, uint8 (0 = background, 255 = terrace)- boundary: terrace boundary mask, uint8 (0/255), used only for boundary-band evaluation- contour: 5 m contour lines, float32 (contour elevation; -9999 elsewhere)- slope: slope in degrees, float32 Files: 82lb.zip (4.59 GB) and SHA256SUMS.txt, which lists the SHA-256 checksums of the archive and of the three split parts distributed on GitHub. Code: https://github.com/hongyuanhuangaa/VEP-MDTNet Please cite this dataset and the associated paper when using the data.