Major-TOM/COP-GEN-Benchmark

---
license: cc-by-4.0
task_categories:
- image-to-image
tags:
- earth-observation
- sentinel-2
- major-tom
- generative-benchmark
- remote-sensing
pretty_name: COP-GEN Benchmark
size_categories:
- 10K<n<100K
configs:
- config_name: real
data_files:
- split: train
path: real/data/*.parquet
- config_name: copgen
data_files:
- split: train
path: copgen/data/*.parquet
- config_name: terramind
data_files:
- split: train
path: terramind/data/*.parquet
dataset_info:
features:
- name: grid_cell
dtype: string
- name: thumbnail
dtype: image
- name: sample_idx
dtype: int32
- name: sample_id
dtype: string
- name: date
dtype: string
- name: crs
dtype: string
- name: ul_x
dtype: float64
- name: ul_y
dtype: float64
- name: B01
dtype: binary
- name: B02
dtype: binary
- name: B03
dtype: binary
- name: B04
dtype: binary
- name: B05
dtype: binary
- name: B06
dtype: binary
- name: B07
dtype: binary
- name: B08
dtype: binary
- name: B8A
dtype: binary
- name: B09
dtype: binary
- name: B11
dtype: binary
- name: B12
dtype: binary
---

# COP-GEN Benchmark

Evaluation dataset for the COP-GEN paper
([arXiv:2603.03239](https://arxiv.org/abs/2603.03239)).
Enables stochastic-benchmark evaluation of generative EO models by
comparing generated sample sets against real multi-temporal Sentinel-2
observations at 495 geographically diverse locations.

## Subsets

Three parallel subsets with **identical schema** and **shared
georeferencing** (same Major TOM v2 1056x1056 grid):

| Subset | Source | Samples/cell |
|-------------|--------------------------------|--------------|
| `real` | Sentinel-2 L2A (GEE, cloud-free) | 16 |
| `copgen` | COP-GEN outputs (subsampled from 33 seeds) | 16 |
| `terramind` | TerraMind outputs | 16 |

Each sample is a 1056x1056 tile with 12 Sentinel-2 bands
(B01, B02, B03, B04, B05, B06, B07, B08, B8A, B09, B11, B12) at their native resolution (10 / 20 / 60 m), stored as
per-band `uint16` GeoTIFF byte blobs inside each parquet row.

## Quick start

```python
from datasets import load_dataset

# Load one of the three configs
real = load_dataset("Major-TOM/COP-GEN-Benchmark", "real", split="train")
copgen = load_dataset("Major-TOM/COP-GEN-Benchmark", "copgen", split="train")
terramind = load_dataset("Major-TOM/COP-GEN-Benchmark", "terramind", split="train")

# Decode a single band from the first row
import rasterio, io
row = real[0]
with rasterio.open(io.BytesIO(row["B02"])) as src:
b02 = src.read(1) # (1056, 1056) uint16
print(row["grid_cell"], row["date"], b02.shape)
```

## Schema

All three subsets share the following columns:

| Column | Type | Notes |
|---------------|---------|---------------------------------------------|
| `grid_cell` | str | Major TOM cell ID (e.g. `106D_246R`) |
| `sample_idx` | int | 0..15 within cell |
| `sample_id` | str | STAC product ID (real) or `seed_N` (models) |
| `date` | str | ISO date for real; empty for synthetic |
| `crs` | str | UTM CRS (e.g. `EPSG:32734`) |
| `ul_x, ul_y` | float | v2 grid upper-left in CRS metres |
| `B01..B12` | bytes | per-band GeoTIFF blob, uint16, native res |
| `thumbnail` | bytes | 256x256 JPEG RGB composite |

## Reproducing the benchmark evaluation

COP-GEN and TerraMind outputs natively cover the centre 192x192 pixels
(1.92 km) of each 1056 tile. To extract this evaluation footprint
identically across all three subsets, use the provided
`benchmark_footprint.py` utility:

```python
import rasterio, io, json
from metadata.benchmark_footprint import crop_benchmark_footprint, load_grid

grid = load_grid("metadata/benchmark_grid.json")

row = real[0]
with rasterio.open(io.BytesIO(row["B02"])) as src:
window = crop_benchmark_footprint(src, row["grid_cell"], grid)
# (1, 192, 192) uint16 — same geographic footprint for all three subsets
```

See `metadata/benchmark_footprint.py` for full documentation of the
cropping convention. The function handles CRS mismatches and is the
exact method used to evaluate the results reported in the paper.

## Metadata

- `metadata/benchmark_grid.json` — cell origins for the 192x192
evaluation footprint (used by `crop_benchmark_footprint`)
- `metadata/cells.parquet` — per-cell summary (grid_cell, ul_x, ul_y,
crs, mgrs_tile)
- `metadata/benchmark_footprint.py` — the crop utility

## Citation

```bibtex
@article{copgen2026,
title={COP-GEN: Latent Diffusion Transformer for Copernicus Earth Observation Data -- Generation Stochastic by Design},
author={Espinosa, Miguel and Gmelich Meijling, Eva and Marsocci, Valerio and Crowley, Elliot J. and Czerkawski, Mikolaj},
year={2026},
journal={arXiv preprint arXiv:2603.03239},
url={https://arxiv.org/abs/2603.03239},
}
```

## Licensing

- Sentinel-2 data: CC-BY 4.0 (Copernicus).
- COP-GEN outputs: released under CC-BY 4.0 by the authors.
- TerraMind outputs: please check the TerraMind licensing terms before
redistribution.