Cross-reservoir Sentinel-1/Sentinel-2 shoreline delineation benchmark: manual labels, model predictions, and evaluation outputs

This dataset supports a leakage-controlled cross-reservoir benchmark of shoreline delineation using Sentinel-1, Sentinel-2, and optical–SAR fusion data. It contains manual shoreline reference masks for 15 reservoir states across five Mexican reservoirs—Cointzio, Queréndaro, Soledad, Esperanza, and Mata—together with scene-pairing registries, derived feature metadata, model predictions, evaluation metrics, multi-seed results, common-state definitions, global and reservoir-adaptive rankings, canonical method-selection records, and quality-control outputs. The benchmark compares 43 configurations spanning spectral indices, Sentinel-2 Scene Classification Layer baselines, SAR thresholding, Random Forest, XGBoost, U-Net, SegFormer, Mask2Former, and SAM 2. Fourteen states were eligible for optical evaluation, while 10 temporally matched states—two per reservoir—formed the strict common subset used for controlled optical, SAR, and fusion comparisons. The full benchmark comprises 505 scene–method evaluations. The dataset also includes screening-level transfer outputs for Little Rock Reservoir, Randy Poynter Lake, and Croton Falls Reservoir in the United States. These external sites lack temporally matched manual shoreline labels and therefore must not be interpreted as supervised external validation. The materials are intended to support reproducibility, independent inspection of the benchmark protocol, comparison of shoreline-delineation approaches, and development of leakage-controlled cross-reservoir evaluation strategies. Bathymetric reconstruction, reservoir-storage estimation, hydrometric validation, and downstream uncertainty analysis are outside the scope of this dataset. Raw Sentinel satellite imagery is not redistributed; scene identifiers and processing registries are provided to support data retrieval from the original public services.