calvdb / zarr_sample /README.md
enscg's picture
Upload zarr_sample/README.md with huggingface_hub
6c4a523 verified

calvdb — Dataset Sample

This folder contains a small sample of the full calvdb dataset, provided so that reviewers can inspect the data format without downloading the complete archive (6.02 GB).

Contents

The sample includes Braasvellbreen glacier (RGI60-07.00025) taken from the train split of the benchmark.

How the sample was created

Each Zarr store in this folder is taken directly from the full dataset without modification. The stores were copied from the full zarr_zipped/ collection and are structurally identical to every other glacier store in calvdb.

Zarr store structure

Each glacier is stored as a Zarr v3 group (zstd compression). Opening any store reveals the following arrays:

Array Shape Description
sdt (T, H, W) Signed distance transform of the calving front; values clipped to ±2000 m
imagery (T, 5, H, W) Multispectral imagery: blue, green, red, NIR, SWIR-1
geophys (T, 6, H, W) Geophysical channels: bed elevation, surface elevation, ice thickness, surface velocity, strain rate, ice–fjord mask
climate (T, 2) Climate forcing: surface runoff, ocean temperature
trace (T, H, W) Rasterised calving-front line (used for evaluation only, not a model input)
dates (T,) Observation dates in YYYY-MM-DD format
valid_sdt (T,) Boolean mask — True if the SDT observation is valid
valid_img (T,) Boolean mask — True if the imagery observation is valid
valid_clm (T,) Boolean mask — True if the climate observation is valid
valid_trace (T,) Boolean mask — True if the trace observation is valid
sensors (T,) Sensor identifier for each imagery observation

All arrays share the same time axis T. Spatial dimensions H × W vary per glacier. Spatial resolution is 30 m; CRS is EPSG:3995 (Arctic Polar Stereographic).

Note on ice_fjord_mask (geophys channel 5): Although ice_fjord_mask is stored in the zarr archive as geophysical channel 5 for completeness, it is excluded from the model input tensor in the baseline experiments due to inaccuracies in the mask boundary for several glaciers. It is retained in the dataset to support future work that may use or improve upon this variable.

Further details

  • Full dataset, splits and normalization: see the main README