U-Net geographic-holdout predictions for Arctic lake and pond water surfaces from Sentinel-2 imagery (2023)

This dataset provides U-Net predictions of lake and pond water surfaces for 1,629 geographic-holdout test tiles in Arctic permafrost lowlands, using summer 2023 Sentinel-2 imagery. It includes 1,305 tiles from Yamal and European Russia and 324 tiles from the buffered Mackenzie Delta experiment. Each tile contains a water-probability raster and a binary prediction raster, giving 3,258 georeferenced GeoTIFFs on a 10 m grid. Binary predictions use a probability threshold of 0.5. The release also includes model checkpoints, training configurations, split records, evaluation results, machine-readable metadata, processing code and SHA256 checksums. The dataset supports reproducible evaluation of geographic model transfer and reuse of the released predictions. The associated training resource uses a hybrid label-construction workflow combining spectral constraints, object screening and model-assisted refinement. Evaluation scores describe agreement with the supervision targets; independent shoreline accuracy has not been established. Coverage consists of selected test tiles rather than a continuous Arctic lake inventory. Associated training data are archived separately at the National Cryosphere Desert Data Center:https://doi.org/10.12072/ncdc.permafrost.db7645.2026