Sk-21/CryoMesh-T4-1pct-sample

CryoMesh T=4 temporal NPY - 1% public sample

This repository contains a public 1% sample of the CryoMesh four-timestamp temporal NPY corpus from all_temporal_npy. It is hosted as a Hugging Face model repository per release request, but the contents are data files for inspection, prototyping, and reproducibility checks.

The sample is selected at the grid level so every included grid has all 4 timestamps. The source directory has 2,371 complete four-timestamp grid samples (9,484 files); this release includes 24 grids (96 files), or 1.012% of complete samples. Total payload size is about 92.74 GB (86.37 GiB).

Contents

• data/*.npy - temporal raster arrays, one file per grid timestamp.
• index.csv - one row per .npy file with grid id, timestamp index, date, region, exact NPY shape, CRS, affine transform, bounds, centroid, and file size.
• samples.csv - one row per grid sample with the four timestamp dates and per-timestamp shape summary.
• crs_metadata.csv - per-file CRS, affine transform, bounds, width, and height metadata recovered from the corresponding source GeoTIFFs.
• region_composition.csv - source and selected sample counts by region.
• selection_summary.csv - source directory, selected fraction, and byte counts used for this release.

Array Format

Each .npy file is a float32 NumPy array in channels-first order with shape (10, H, W). Grid dimensions vary by location, and a few temporal stacks have different H/W values across timestamps. Use index.csv for the exact per-file shape and geospatial transform.

Index	Band	Source
0	Blue	Sentinel-2
1	Green	Sentinel-2
2	Red	Sentinel-2
3	NIR	Sentinel-2
4	SWIR1	Sentinel-2
5	SWIR2	Sentinel-2
6	Slope	Copernicus DEM derived layer
7	Elevation	Copernicus DEM
8	VV	Sentinel-1 SAR
9	VH	Sentinel-1 SAR

Note on SAR polarization: over some polar regions Sentinel-1 acquisitions are single-polarization. In those cases the available polarization may be duplicated into the VV/VH slots to preserve a stable 10-channel schema.

Region Composition

Region	Complete Source Samples	Selected Samples
antarctic	821	8
arctic	774	8
hma	276	3
north_america	240	2
south_america_andes	117	1
other	98	1
europe_alps	45	1

Usage

import numpy as np
import pandas as pd

idx = pd.read_csv("index.csv")
sample = idx[idx.grid_id == "grid_12153"].sort_values("timestamp_index")
sequence = [np.load(path) for path in sample.filename]
print(sample.date.astype(str).tolist())  # ['20210109', '20210204', '20211006', '20211125']
print(sequence[0].shape)                 # (10, H, W)

For geospatial placement, use the per-file crs, transform_*, and bounds_* columns in index.csv or crs_metadata.csv.

Selection Method

Only grid IDs with exactly four .npy timestamps were eligible. The selected count is ceil(0.01 * 2,371) = 24 complete grid samples. Allocation is proportional by region using largest-remainder apportionment, then deterministic even spacing over numeric grid IDs within each region.

Provenance and Attribution

Derived from public Earth-observation sources, 2021:

• Contains modified Copernicus Sentinel-2 and Sentinel-1 data.
• Includes Copernicus DEM derived and elevation layers. Copernicus DEM copyright: DLR e.V. / Airbus Defence and Space GmbH.

Curation, co-registration, multi-sensor harmonization, cloud screening, tiling, and NPY packaging are original CryoMesh processing work.

License

Released under CC-BY-4.0. Please retain the Copernicus attributions above and cite CryoMesh when using this sample.

Citation

@misc{cryomesh_t4_1pct_sample_2026,
  title  = {CryoMesh T=4 temporal NPY - 1% public sample},
  author = {Kaushik, Saurabh and collaborators},
  year   = {2026},
  note   = {One percent public sample of the CryoMesh four-timestamp temporal NPY corpus}
}