terrafloww/rasteret-collection-major-tom-style

Rasteret Collection: Major TOM style (Sentinel-2 L2A scene index, 20k rows)

This repository publishes a Rasteret Collection artifact: a portable Parquet/GeoParquet index that points to public Sentinel-2 COGs and includes the per-band metadata Rasteret needs for fast byte-range reads.

It exists to make the HF Datasets (images-in-Parquet) vs Rasteret (COGs + index) benchmark easy to reproduce, and to provide a realistic example of an "index-first" workflow where:

• The control plane is a queryable Parquet dataset (scenes, footprints, band metadata, custom keys)
• The data plane stays in source COGs (no chip exports, no image blobs inside Parquet)

Contents

• rasteret-collection-major-tom-style/
  • A Parquet dataset directory (Hive partitions year=.../month=.../part-0.parquet)
  • 20,000 rows (one row per Sentinel-2 L2A scene)
  • Date range in this snapshot: 2023-01-20 to 2024-02-06
  • CRS diversity: 50 distinct proj:epsg values (multi-UTM)
  • Geometry footprints in CRS84 (WKB)
  • COG tile metadata for B02 and B08 only
• benchmark/
  • Repro scripts copied from the Rasteret repo (see benchmark/README.md)

What this is (and is not)

• This is a scene-level index, not a patch dataset.
• It contains no pixel payloads.
• It is "Major TOM style" because it includes a major_tom_product_id column that matches Major TOM keys (product_id), derived from s2:product_uri (with .SAFE removed).

Data sources

• Scene metadata: derived from Earth Search STAC for the Sentinel-2 L2A collection.
• Pixel data (not included here): public Sentinel-2 COGs hosted at https://sentinel-cogs.s3.us-west-2.amazonaws.com/....

This repo only stores metadata and URLs. You are responsible for complying with upstream data terms when you fetch pixels from the source COGs.

Schema (columns)

The dataset includes the Rasteret required columns plus a few extras:

• id: STAC item id
• datetime: acquisition timestamp
• geometry: footprint geometry (WKB, CRS84)
• assets: mapping of band -> asset dict (here: B02, B08) containing href and Rasteret hints
• scene_bbox: [minx, miny, maxx, maxy] in CRS84
• bbox_minx, bbox_miny, bbox_maxx, bbox_maxy: scalar bbox columns for fast predicate pushdown
• proj:epsg: scene native CRS EPSG (multi-UTM)
• s2:product_uri: STAC s2:product_uri property
• collection: STAC collection id (sentinel-2-l2a)
• B02_metadata, B08_metadata: Rasteret COG metadata structs (tile offsets, byte counts, transform, dtype, nodata, etc.)
• major_tom_product_id: normalized Major TOM-style product id (no .SAFE suffix)
• year, month: partitions derived from datetime

Quickstart: load with Rasteret

Requirements:

• Python 3.12+
• pip install rasteret (or uv pip install rasteret)

from huggingface_hub import snapshot_download
from pathlib import Path
import rasteret

root = Path(snapshot_download(
    "terrafloww/rasteret-collection-major-tom-style",
    repo_type="dataset",
))

collection = rasteret.load(root / "rasteret-collection-major-tom-style")
print(collection.dataset.count_rows())
print(collection.bands)  # ['B02', 'B08']

Run the HF vs Rasteret benchmark

The benchmark compares:

1. Hugging Face datasets streaming Parquet reads, where each band column stores a small GeoTIFF blob
2. Rasteret reading the same patches from source COGs using this index

# One-time install (includes HF deps used by the benchmark)
uv pip install "rasteret[examples]"

# HF access token is recommended for stable throughput
export HF_TOKEN=...

uv run python benchmark/03_hf_vs_rasteret_benchmark.py \
  --collection-path rasteret-collection-major-tom-style \
  --samples 1000 \
  --bands B02 B08 \
  --sample-strategy random \
  --random-seed 42

Known limitations

• Only B02 and B08 are enriched in this published index.
  • If you want more bands, rebuild a collection with the Rasteret scripts and include additional *_metadata columns.
• This is a scene index; patch geometries are supplied at query time (see the benchmark scripts).

Reproducibility

This collection can be regenerated using Rasteret's sharded STAC builder:

• examples/major_tom_benchmark/01_build_collection_sharded.py

Citation

If you use this dataset in a paper or benchmark report, please cite Rasteret and the upstream data sources.