Kurdistan Surface Water Dynamics (KSWD): A multi-temporal Sentinel-2 and SRTM tensor dataset for machine learning-based reservoir change detection in Iraq

Kurdistan Surface Water Dynamics (KSWD) v1.0 — A multi-temporal Sentinel-2 and SRTM tensor dataset for machine learning-based reservoir surface water change detection in the Kurdistan Region of Iraq KSWD is an analysis-ready, deep-learning-oriented geospatial dataset documenting surface water dynamics of the three principal dam reservoirs of the Kurdistan Region of Iraq — Dukan (8 ROIs), Darbandikhan (5 ROIs), and Duhok (1 ROI) — over seven consecutive hydrological years (2019–2025). This window captures the exceptionally wet 2019 season, when regional reservoirs approached full storage, and the severe multi-year drought drawdown from 2021 to 2025 driven by reduced precipitation and declining transboundary inflows. Dataset contents The dataset comprises 98 sample units (dam × ROI × year), totaling 686 GeoTIFF files. Each sample unit contains: Four seasonal 9-band tensors (winter, spring, summer, autumn): 512 × 512 pixels at 10 m resolution, uniformly cast to 16-bit signed integer (Int16). Bands: (1–3) display-scaled RGB from Sentinel-2 B4/B3/B2; (4–6) MNDWI × 100, NDWI × 100, NDVI × 100; (7) binary water mask (MNDWI > 0.1; 0 = land, 1 = water); (8–9) SRTM elevation (m) and slope (degrees) for topographic context to suppress mountain-shadow false positives. Three cross-seasonal binary change maps (winter→spring, spring→summer, summer→autumn): 1 = water transition (gain or loss), 0 = stable. All imagery was generated in Google Earth Engine from COPERNICUS/S2_SR_HARMONIZED (SCL-based cloud/cirrus/shadow masking, seasonal median compositing; winter spans Dec of the previous year through Feb) and USGS/SRTMGL1_003. Winter of target year Y covers December (Y−1) to February (Y). Files in this record kswd_dukan.zip, kswd_darbandikhan.zip, kswd_duhok.zip — per-dam GeoTIFF archives organized as data/{dam}/r{NN}/{year}/{dam}_r{NN}_{year}_{product}.tif kswd_docs_and_metadata.zip — manifest.csv (one row per sample unit with paths, CRS, WGS84 bounds, per-season water/valid-pixel fractions, per-map change fractions), band_descriptions.md, roi_footprints.geojson, sha256sums.txt, and splits/train.csv, val.csv, test.csv README.md, CITATION.cff, LICENSE.txt — loose copies for direct viewing Train/validation/test splits Splits are ROI-held-out to prevent spatial autocorrelation leakage: every ROI, with all seven of its years, belongs to exactly one partition. Train: Dukan r01–r06 + Darbandikhan r01–r03 (63 units); validation: Dukan r07 + Darbandikhan r04 (14 units); test: Dukan r08 + Darbandikhan r05 + Duhok r01 (21 units). The Duhok reservoir appears only in the test set, providing an unseen-dam generalization benchmark. Splits are declared in CSV tables, so users may construct alternative partitions (e.g., a temporal hold-out of 2025) without restructuring the archive. Intended uses Training and benchmarking semantic segmentation and Siamese change-detection networks on reservoir water dynamics; transfer learning for drought- and flood-driven surface water monitoring; hydrological analysis of reservoir surface-area fluctuation in semi-arid mountainous terrain. Known limitations Persistent winter cloud cover over the Zagros Mountains can reduce valid-pixel coverage in some winter composites (see valid_frac_* columns in the manifest); the water labels derive from a single global MNDWI threshold rather than manual annotation; SRTM elevation is static (February 2000 epoch) at native 30 m resolution. License and citation Released under CC BY 4.0. Please cite this record using the DOI above (machine-readable metadata in CITATION.cff). Contains modified Copernicus Sentinel data (2019–2025), processed with Google Earth Engine.