Data for: Benchmarked and leakage-validated deep-learning glacier segmentation: a 42-year, five-sensor snow-fraction record for western North America

Le résumé fourni par la source

Dataset and trained models accompanying: Tarka, Maxime and Baraër, Michel and Aubry-Wake, Caroline, Benchmarked and leakage-validated deep-learning glacier segmentation: a 42-year, five-sensor snow-fraction record for western North America (September 21, 2026). Available at SSRN: https://ssrn.com/abstract=7418104 or http://dx.doi.org/10.2139/ssrn.7418104 A four-class U-Net (Snow / Ice / Cloud / Other) was trained on 540 annotated scenes from 206 glaciers across five optical sensors (Landsat 5/7/8/9, Sentinel-2) and 42 years (1984-2025). Under strict glacier-level spatial splitting, the model reaches test mIoU = 0.705, outperforming four independently tuned baselines (NDSI/NDVI threshold, Random Forest, DeepLabv3, SegFormer). Applied continentally to 22,155 glaciers across Alaska and Western Canada/USA, the pipeline produces 786,014 glacier-year observations of end-of-season snow fraction (Fsnow), a remotely measurable complement to the accumulation-area ratio. This deposit also documents the spatial-autocorrelation leakage diagnostic, the benchmark suite, and every robustness analysis (attention-gate ablation, hyperparameter-search confirmation, predictive-uncertainty, throughput measurement) that supports the paper's conclusions. Contents: - UNet_Models/: trained model checkpoints, per-run evaluation outputs, and confusion matrices for the production model plus three ablation splits (glacier-level, scene-level, boundary-excluded).- UNet_Train_Data/: the 540-scene, 206-glacier annotated corpus (TOA imagery + pixel-level Snow/Ice/Cloud/Other masks), released as a standalone, independently citable benchmark dataset for glacier surface-state segmentation.- Baselines/: four independently tuned, Optuna-searched, five-seed-confirmed baselines (NDSI/NDVI threshold, Random Forest, DeepLabv3, SegFormer) benchmarked against the published U-Net on the identical split.- Attention_Gate_Ablation/: the U-Net's attention gates tested present vs. absent on two independent glacier-level splits.- HPO_Robustness/: a substantially wider hyperparameter search than the one used to select the published configuration, confirmed on ten seeds.- Leakage_Significance_Test/: formal statistical tests (paired and Welch t-tests, bootstrap confidence intervals) quantifying the scene-level vs. glacier-level split leakage central to the paper.- Predictive_Uncertainty/: predictive entropy evaluated as a pixel-level error-detection signal (AUROC = 0.872).- Glacier_Attribute_Regression/: a regression of per-glacier segmentation performance against elevation, area, and aspect.- Throughput_Measurement/: directly measured inference throughput on a single consumer GPU.- Visible_Glacier_Surface/ and Glacier_Evolution/: the fixed reference footprint (VGS) used to compute snow fraction, plus annual per-class (Snow/Ice/Other/Cloud) footprint polygons and centroid trajectories for every glacier-year.- Analysis_Inputs/: per-glacier and per-glacier-year snow-fraction statistics, plus the terrain and climate covariates compiled alongside the record. See the top-level and per-folder README.md files in this deposit for full file-level documentation. Code: https://github.com/Research-Tarka/glacier-fsnow-unet This archive can be browsed per glacier, without downloading it in full, through GlacierScope, an open-source, standalone viewer (world map, per-glacier snow/ice state, centroid drift, area change): https://github.com/Research-Tarka/GlacierScope. GlacierScope only needs two parts of this deposit, picked in its file picker on first run: Glacier_Evolution/glacier_evolution_full.gpkg and Glacier_Evolution/DEM_tiles/ (required), plus optionally any parquet/csv table under Analysis_Inputs/ (Static/glacier_ref.parquet, Temporal/glacier_year_state.parquet, Temporal/stats_normalized.parquet, and anything under Covariates/) to browse alongside the map. The pixel-level annotation interface used to build the training corpus has since been rebuilt and generalised beyond glacier imagery, and is released separately as MaskForge: https://github.com/Research-Tarka/maskforge (MaskForge itself was not used to produce the annotations in this deposit).

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.