LoessScape: A Multi-Scale Semantic Segmentation Dataset for Loess Plateau Land Use

LoessScape is a pixel-level annotated semantic segmentation dataset built from high-resolution (2.0 m/pixel) remote sensing imagery of a representative region of the Loess Plateau (~1,853 km²). It serves as a standardized benchmark for training and evaluating land use / land cover (LULC) segmentation models incomplex terrain environments. The dataset contains 9,888 tiles of 512×512 pixels, organized into three anchor-aligned perception scales — scale_1_0.5 (1.0 m GSD), scale_1_1 (2.0 m, native), and scale_1_2 (4.0 m) — and split into train/val/test (7/1.5/1.5). Pixel-level masks follow a two-level hierarchy based on the China national standard GB/T 21010-2017: 14 Level-2 classes grouped into 7 Level-1 categories (Cropland, Orchard, Forest, Grassland, Built-up, Transportation, Water). NoData (255) marks areas outside the study-area boundary. For open-release compliance, all geo-transform values are normalized to relative meter offsets — no absolute geographic coordinates exist in any file. Files:- images.zip — multi-scale RGB GeoTIFF tiles- masks.zip — pixel-level semantic mask tiles- metadata.zip — class taxonomy, dataset info, and per-split tile metadata- README.md — full dataset documentation- LICENSE — CC BY-NC-SA 4.0- MD5SUMS.txt — file checksums for integrity verification