The Shelf-Bench Dataset: A benchmark dataset for Antarctic ice shelf front and coastline delineation from multi-sensor radar satellite data

Shelf-Bench is a comprehensive benchmark dataset for the delineation of Antarctic ice shelf fronts and coastlines using Synthetic Aperture Radar (SAR) imagery. The dataset comprises 161 manually annotated scenes spanning three decades (1992–2021) from Sentinel-1, ERS and Envisat satellites. Thirty-eight scenes have been designated as a standardised test set to enable model comparison. Processed patches (265 x 265 pixels) of the satellite scenes are also provided in PNG format for training, validation, and testing purposes, along with trained weights for five baselines (U-Net, DeepLabv3, FPN, DINOv3 7B, and ViT), which were trained on the PNG patches.  Additionally, the Geopackage contains the extents of all the training and test scenes, along with their metadata, to provide a quick geospatial overview in GIS software such as QGIS.

This repository contains the following files:

• Shelf-Bench.zip: Satellite scene-based benchmark dataset, georeferenced GeoTiffs 
• Shelf-Bench-extents.gpkg: Geospatial extents of train and test scenes
• Shelf-Bench-tiles.zip: Tiled PNG patches for train, validation and test extracted from Shelf-Bench.zip
• trained_models.zip: Weights trained based on the five baseline models

The scene-based dataset is structured as follows:

{Satellite}/
  ├── masks/
  ├── scenes/
  └── test_{satellite}/
        ├── masks/
        └── scenes/

File naming follows the structure:

[SATELLITE-ID]_[YYYYMMDD]_[POLARISATION]_[SCENE_ID].tif

• SATELLITE-ID: `S1A`, `S1B`, `ERS`, or `ENV`. S1A and SLB refers to Sentinel-1a and Sentinel-1b respectively.
• YYYYMMDD: Acquisition date.
• POLARISATION: e.g., `HH`, `VV.`
• SCENE_ID: Product-specific identifiers (e.g., scene code, crop number, or location prefix).