Datasets:
The dataset viewer is not available for this split.
Parquet error: Scan size limit exceeded: attempted to read 309515466 bytes, limit is 300000000 bytes
Make sure that
1. the Parquet files contain a page index to enable random access without loading entire row groups2. otherwise use smaller row-group sizes when serializing the Parquet files
Error code: TooBigContentError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
FLAIR multi-domain pack (aerial 20 cm + Sentinel-2) for Colab
24,475 tiles from 9 French domains, repacked from IGNF/FLAIR-HUB into Parquet so a Colab runtime can stream it.
| domain | tiles | sectors | MB |
|---|---|---|---|
D004-2021 |
3,725 | - | 310 |
D006-2020 |
3,850 | - | 321 |
D011-2021 |
2,175 | - | 162 |
D017-2021 |
1,225 | - | 91 |
D022-2021 |
2,775 | - | 222 |
D033-2021 |
4,800 | - | 372 |
D037-2021 |
1,425 | - | 113 |
D040-2021 |
1,700 | - | 132 |
D044-2022 |
2,800 | - | 212 |
Total 1.94 GB (raw equivalent ~5.9 GB uncompressed).
Columns
| column | type | meaning |
|---|---|---|
rgb |
JPEG bytes | aerial RGB, 256px, quality 95 |
nir |
JPEG bytes | aerial near-infrared, same grid |
lab |
PNG bytes | raw COSIA values 0-18, bit-exact, nearest-resized |
s2 |
zstd bytes | Sentinel-2, decompress to uint16 of shape s2_shape = (12, 10, 10, 10) |
s2_dates_raw |
int | acquisitions before compositing (38-73) |
s2_cloudy |
int | dates dropped at >50% cloud probability |
domain, sector, stem |
str | provenance; split by domain |
Fidelity
- Labels are bit-exact. PNG, lossless, no remapping applied.
- RGB is JPEG q95. Measured cost: a model trained on lossless tiles and evaluated on these scores -0.03 mIoU (0.01 sigma against this study's 2.00 noise floor). q90 costs -0.23, q85 -0.37.
- Sentinel-2 is bit-exact after zstd, but temporally composited to 12 equal-width bins with >50% cloud-probability dates dropped first. Re-pack from the source if you need raw dates.
Load
import io, json, zstandard, numpy as np, pyarrow.parquet as pq
from PIL import Image
from huggingface_hub import hf_hub_download
p = hf_hub_download("FatimahEmadEldin/flair-multidomain-parquet", "D004-2021.parquet", repo_type="dataset")
t = pq.read_table(p)
r = t.slice(0, 1).to_pylist()[0]
rgb = np.array(Image.open(io.BytesIO(r["rgb"]))) # (H,W,3) uint8
lab = np.array(Image.open(io.BytesIO(r["lab"]))) # (H,W) 0-18
s2 = np.frombuffer(zstandard.ZstdDecompressor().decompress(r["s2"]),
dtype=np.uint16).reshape(json.loads(r["s2_shape"]))
COSIA classes
Value 0 is Building - a real class, not a no-data marker. Values 15-18 are weak labels that IGN's own baseline excludes.
| value | class | value | class | |
|---|---|---|---|---|
| 0 | Building | 10 | Plowed land | |
| 1 | Greenhouse | 11 | Vineyard | |
| 2 | Swimming pool | 12 | Deciduous | |
| 3 | Impervious surface | 13 | Coniferous | |
| 4 | Pervious surface | 14 | Brushwood | |
| 5 | Bare soil | 15 | Clear cut | |
| 6 | Water | 16 | Ligneous | |
| 7 | Snow | 17 | Mixed | |
| 8 | Herbaceous vegetation | 18 | Undefined | |
| 9 | Agricultural land |
- Downloads last month
- 21