A Sentinel-2 semantic segmentation dataset for thermokarst-associated lake and pond water surfaces in Arctic permafrost lowlands

Abstract. Lake and pond water surfaces in thermokarst-affected Arctic permafrost lowlands are important for hydrological, habitat, and carbon-cycle studies. Existing inland-water products rarely provide georeferenced image-mask pairs designed for supervised segmentation of small Arctic water bodies. We present a Sentinel-2 semantic segmentation dataset for thermokarst-associated lake and pond water surfaces across 23 selected permafrost-lowland areas of interest (AOIs) spanning Alaska, northwestern Canada, European Russia, Yamal, and Siberia. The supervised target is visible summer lake and pond water, while separate tile-level fields record thermokarst relevance and non-target-water risk. The dataset contains 5513 image-mask pairs derived from summer 2023 Sentinel-2 Level-2A surface-reflectance composites. Each image is a 256 by 256 pixel GeoTIFF with five bands (B2, B3, B4, B8, and B11), accompanied by a binary water mask and a quality-assurance (QA) mask for invalid observations and ambiguous margins. The train, validation, and test splits contain 3982, 739, and 792 tiles, respectively. The masks contain 29,141,595 positive pixels, representing 8.07 % of all pixels and a pixel-count-equivalent area of approximately 2914.2 km²; 88.35 % of target-water pixels lie north of or on the Arctic Circle at 66°34′ N. The package includes 16539 image, water-mask and QA GeoTIFF files, with sample metadata, source records, processing settings, split definitions and a 150-tile low-water subset index. A geometry audit identified cross-split overlap in the legacy development split and in the original Mackenzie Delta holdout. After removing overlapping and nearby development samples and retraining with QA-aware loss, the Mackenzie Delta experiment reached a target-agreement intersection-over-union (IoU) of 0.659 (spatial-cluster bootstrap interval 0.587–0.728). The Yamal and European Russia holdout had no detected test–development footprint overlap and retained its previously reported target-agreement IoU of 0.936. These values measure agreement with the released supervision targets; independent human-reference accuracy remains to be established. The release is designed for supervised segmentation and controlled model comparison within the represented 2023 Arctic permafrost-lowland settings. The dataset is available from the National Cryosphere Desert Data Center at https://doi.org/10.12072/ncdc.permafrost.db7645.2026 (Guo, 2026). The accompanying model-output dataset is available on Zenodo at https://doi.org/10.5281/zenodo.22965055 (Zhang et al., 2026).