enscg/calvdb

calvingdb — Calving Front Benchmark

First benchmark dataset for data-driven annual calving front forecasting, covering 123 marine-terminating glaciers in Svalbard from 2013 to 2023.

Each benchmark sample asks: given five past calving front observations (spanning roughly four years), predict where the calving front will be 365 days in the future.

---

Quick start

Requires zarr >= 3.0 and Python >= 3.11 — see Requirements. Read the archives directly; do not unzip them.

pip install "zarr>=3.0" huggingface_hub

import zarr
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="enscg/calvdb",
    filename="zarr_zipped/RGI60-07.00552.zarr.zip",   # one glacier, 30 MB
    repo_type="dataset",
)

group = zarr.open_group(zarr.storage.ZipStore(path, mode="r"), mode="r")

print(dict(group.attrs))    # glacier_id, T, H, W, CRS, channel names
print(group["sdt"].shape)   # (T, H, W)  signed distance transform, metres
print(group["dates"][:5])   # observation dates, YYYY-MM-DD

Or with xarray, for named dimensions and lazy loading:

import xarray as xr

ds = xr.open_zarr(zarr.storage.ZipStore(path, mode="r"))
print(ds.sdt.dims)          # ('time', 'y', 'x')

A complete runnable example — download, inspect, plot, and load a benchmark sample — is in example_loading.ipynb.

---

Requirements

Component	Requirement	Why
zarr	>= 3.0 (tested 3.1.5)	stores are Zarr format 3; zarr 2.x cannot read them
Python	>= 3.11	hard floor of zarr-python 3.x
numcodecs	>= 0.14	installed with zarr 3; provides the zstd codec
numpy	>= 1.26	zarr 3 floor
xarray	>= 2024.10	optional, for the xarray route
huggingface_hub	any recent	download only

Troubleshooting

GroupNotFoundError: group not found at path '' — you are on zarr 2.x. A v2 reader looks for .zgroup / .zarray, which Zarr v3 stores do not contain. Check with import zarr; print(zarr.__version__); it must be 3.x. Do not upgrade zarr inside an existing 2.x environment unless you are ready for a breaking API change — create a fresh environment instead.

TypeError: cannot cast dtype StringDType() — under numpy >= 2 the dates array (vlen-utf8) arrives as StringDType. Use dates.tolist(), not dates.astype(str).

Opening the .zip with zipfile, or unzipping it first — unnecessary. zarr.storage.ZipStore reads chunks directly from the archive.

---

Dataset at a glance

Property	Value
Glaciers	123 (RGI 6.0, region 07 — Svalbard)
Total observations	17,358 calving front scenes
Benchmark samples	1,234 (866 train / 109 val / 259 test)
Temporal coverage	2013–2023
Spatial resolution	30 m
CRS	EPSG:3995 (Arctic Polar Stereographic)
Prediction horizon	365 days
Input sequence length	5 snapshots
Download size	5.4 GB (all glaciers); 0.4–300 MB per glacier
Version	1.1.0
License	CC BY 4.0

---

Repository layout

calvdb/
├── zarr_zipped/                 # one Zarr v3 store per glacier, zipped
│   └── RGI60-07.XXXXX.zarr.zip  #   read directly with zarr.storage.ZipStore
├── zarr_sample/                 # one store, unzipped, for browsing only
│   └── RGI60-07.00025.zarr/     #   click through it in the file browser above
├── splits/
│   ├── train.json               # 866 benchmark samples
│   ├── val.json                 # 109 benchmark samples
│   ├── test.json                # 259 benchmark samples
│   ├── normalisation_stats.json # channel statistics (train split only)
│   └── sampling_params.json     # reproducibility config
├── croissant.json               # MLCommons Croissant metadata
└── README.md

Browsing the layout without downloading

zarr_sample/RGI60-07.00025.zarr/ is one glacier (Braasvellbreen, 55 timesteps) left unzipped so you can explore the store in the file browser above and read the zarr.json metadata documents in place — for example zarr_sample/RGI60-07.00025.zarr/sdt/zarr.json shows the shape, dtype, chunking and dimension names of the SDT array.

Its arrays are identical to zarr_zipped/RGI60-07.00025.zarr.zip; it is provided for inspection only. For actual use, download from zarr_zipped/ — one file of 275 MB rather than 617 loose files, and readable without unzipping.

---

Store contents

Each archive holds one glacier with T observations on an H × W grid at 30 m.

Array	Shape	Dtype	Description
sdt	(T, H, W)	float32	Signed distance transform to the front, metres, clipped ±2000; the front is the zero level set
trace	(T, H, W)	bool	Rasterised front line. Evaluation ground truth only — never a model input
imagery	(T, 5, H, W)	float32	Optical bands: blue, green, red, nir, swir1
geophys	(T, 6, H, W)	float32	bed_elevation, surface_elevation, thickness, velocity, strain_rate, ice_fjord_mask
climate	(T, 2)	float32	mar_runoff, ocean_temp
dates	(T,)	string	Acquisition date, YYYY-MM-DD
sensors	(T,)	uint8	0=S2, 1=LE07, 2=LC08, 3=LC09, 255=unknown
valid_sdt, valid_img, valid_clm, valid_trace	(T,)	bool	Per-timestep reliability flags — check these before using a timestep

Root attributes carry glacier_id, T, H, W, resolution_m, crs, sdt_clip_m, and the channel-name lists.

---

Benchmark splits

Each entry in splits/{train,val,test}.json is one forecasting sample:

{
  "glacier_id": "RGI60-07.00030_D",
  "zarr_path": "zarr_zipped/RGI60-07.00030_D.zarr.zip",
  "reference_date": "2017-05-20",
  "horizon_days": 365,
  "input_zarr_indices": [0, 5, 7, 14, 22],
  "target_zarr_index": 53,
  "anchor_offsets": [-1460, -1095, -730, -365, -30]
}

Index into the arrays with input_zarr_indices (the five inputs) and target_zarr_index (the front 365 days after the last input):

group = zarr.open_group(zarr.storage.ZipStore(path, mode="r"), mode="r")
inputs = group["sdt"][sample["input_zarr_indices"]]   # (5, H, W)
target = group["sdt"][sample["target_zarr_index"]]    # (H, W)

Entries also carry zarr_path (e.g. zarr_zipped/RGI60-07.00030_D.zarr.zip), the repository-relative location of that glacier's archive, plus input_days, target_day and day_gaps giving the actual observation offsets in days — sampling is irregular, so these differ from the nominal anchor_offsets.

Suggested task

Predict displacement rather than absolute position: sdt(target) − sdt(last input). The natural baseline is persistence — predict zero displacement, i.e. the front does not move — which any useful model must beat.

Normalisation

splits/normalisation_stats.json contains per-channel mean and standard deviation computed from the training split only. Apply z-score normalisation before training.

---

Changelog

1.1.0 — Fixed archive packaging. Stores were previously nested under a zarr/<GLACIER>.zarr/ prefix inside each zip, which made zarr.open_group() and xarray.open_zarr() fail with GroupNotFoundError. Stores are now at the archive root. Added dimension_names to all arrays (xarray previously could not open the stores at all) and consolidated metadata for faster opens. Added loading documentation and a runnable example notebook. Regenerated the unzipped zarr_sample/ with the same corrected metadata, so it matches the archives. Array data is unchanged — only packaging and metadata.

1.0.0 — Initial release.

License

Released under CC BY 4.0.