This dataset contains 3,330 satellite imagery samples (1,352 landslide, 1,978 non-landslide) for landslide segmentation in Pakistan's Gilgit-Baltistan region, developed for the PKLandSeg study (IEEE Geoscience and Remote Sensing Letters, DOI: 10.1109/LGRS.2026.3673080). Each sample includes RGB imagery (PlanetScope, 3m resolution), Copernicus DEM elevation data, NDVI, slope, and a binary landslide annotation mask, all at 512x512 pixel resolution. The dataset is split into fixed train (2,177), validation (660), and test (493) sets for reproducibility. Annotations were produced through a two-phase AI-assisted workflow: an initial set of 1,072 sparse expert annotations was used to train a model that generated probability maps, which annotators then used to refine and expand the final labels. Dataset quality was independently validated using DeepLabV3+, achieving a TTA F1-score of 0.578. RGB imagery is released as JPEG in compliance with Planet's Education and Research Program data sharing guidelines; DEM, NDVI, and slope layers are unrestricted derivative/external data. See the included README.md for full folder structure, methodology, licensing, and known limitations. Recommended CitationPlease cite the following publication when using this dataset: S. B. Hammad, E. Naseer, M. A. Siddique and R. Ahmed, "PKLandSeg: Annotation-Guided Hybrid CNN-Vision Transformer for Landslide Segmentation," in IEEE Geoscience and Remote Sensing Letters, vol. 23, pp. 1-5, 2026, Art no. 5001505. For methodology details (e.g., annotation workflow, model architecture, performance metrics), please refer to the above paper.