timm/bigearthnet-v2-rgb

BigEarthNet v2.0 (reBEN) - Sentinel-2 RGB

Description

BigEarthNet v2.0 (reBEN, "refined BigEarthNet"), a multi-label land-cover classification dataset of Sentinel-2 image patches from 10 European countries. Each 1200 m x 1200 m patch (120x120 pixels at 10 m) is labelled with one or more of 19 land-cover classes. The labels come from the CORINE Land Cover 2018 map, using the 19-class BigEarthNet nomenclature.

This variant is RGB-only, stored as JPEG (quality 98, 4:4:4 chroma, ~40 dB+ PSNR vs. the tone-mapped values). See timm/bigearthnet-v2-rgb-nir-swir for lossless PNG RGB plus raw NIR / SWIR bands.

This is a derived version of the Sentinel-2 part of the official release. It contains only the 480,038 patches in metadata.parquet; the 69,450 patches flagged for seasonal snow or cloud / shadow are excluded, as in the standard benchmark. It uses the official train / validation / test splits (237,871 / 122,342 / 119,825). reBEN assigns these geographically to reduce spatial correlation between splits, so do not re-split randomly.

• Website: https://bigearth.net
• Source data: https://zenodo.org/records/10891137 (BigEarthNet-S2.tar.zst, metadata.parquet)
• Paper: https://arxiv.org/abs/2407.03653

Labels

Every patch has at least one label (mean 2.95). Class indices follow the short snake_case names in alphabetical order. The original names are in the table:

index	name	original CLC-based label	train	validation	test
0	agriculture_with_natural_vegetation	Land principally occupied by agriculture, with significant areas of natural vegetation	60,127	32,736	29,846
1	agro_forestry_areas	Agro-forestry areas	15,082	8,157	9,942
2	arable_land	Arable land	88,236	49,737	50,052
3	beaches_dunes_sands	Beaches, dunes, sands	738	426	152
4	broad_leaved_forest	Broad-leaved forest	64,164	35,387	36,377
5	coastal_wetlands	Coastal wetlands	670	610	117
6	complex_cultivation_patterns	Complex cultivation patterns	51,034	26,486	22,078
7	coniferous_forest	Coniferous forest	76,344	39,554	39,043
8	industrial_commercial_units	Industrial or commercial units	6,398	2,726	2,018
9	inland_waters	Inland waters	30,466	15,900	16,846
10	inland_wetlands	Inland wetlands	11,344	5,056	4,519
11	marine_waters	Marine waters	35,914	14,064	11,854
12	mixed_forest	Mixed forest	78,791	42,705	44,284
13	moors_heathland_sclerophyllous_vegetation	Moors, heathland and sclerophyllous vegetation	6,329	3,806	3,759
14	natural_grassland_sparse_vegetation	Natural grassland and sparsely vegetated areas	6,984	2,687	2,211
15	pastures	Pastures	42,968	25,915	26,722
16	permanent_crops	Permanent crops	15,190	8,688	5,710
17	transitional_woodland_shrub	Transitional woodland, shrub	64,198	36,429	40,523
18	urban_fabric	Urban fabric	33,626	17,308	12,824

RGB tone mapping

The source bands are 16-bit L2A surface reflectance (reflectance x 10000). RGB (B04, B03, B02) is converted to 8-bit with one fixed, global tone curve. Nothing is adjusted per image, so brightness stays radiometrically consistent across patches:

1. x = DN / 3500
2. highlight shoulder: identity for x <= 0.6 (DN <= 2100, ~98% of pixels). Above that, an extended-Reinhard roll-off (slope-continuous at the knee) reaches 1.0 at DN 12000. Snow, cloud and bright surfaces are compressed instead of clipped.
3. gamma: y = x ** (1 / 2.2), v = round(255 * y)

This follows the same idea as Sentinel Hub's "L2A True Color Optimized" rendering: display gamma plus highlight compression. It is not the linear, clipped ESA TCI product. On clean patches, 0% of pixels clip at 255. Mean / std of the tone-mapped RGB over train (0-1 scale): mean [0.3956, 0.4205, 0.3508], std [0.174, 0.141, 0.1355].

To map values back to approximate reflectance DN (exact up to 8-bit quantization of the forward curve):

import numpy as np

def inverse_tone_map(v):
    """8-bit tone-mapped value(s) -> approximate L2A reflectance DN (reflectance * 10000)."""
    scale, white, knee, gamma = 3500., 12000., 0.6, 2.2
    x = (np.asarray(v, dtype=np.float64) / 255) ** gamma  # undo gamma
    tw = (white / scale - knee) / (1 - knee)
    s = np.clip((x - knee) / (1 - knee), 0, None)  # undo extended-Reinhard shoulder above the knee
    a = 1 / tw ** 2
    t = (-(1 - s) + np.sqrt((1 - s) ** 2 + 4 * a * s)) / (2 * a)
    x = np.where(x <= knee, x, knee + (1 - knee) * t)
    return x * scale

Dataset Structure

• image - RGB (B04, B03, B02), 120x120 (10 m), tone-mapped to 8-bit (see below), JPEG q98 4:4:4.
• labels - sequence of ClassLabel indices of the positive classes.
• label_names - short names of the positive classes.
• image_id - reBEN patch_id, e.g. S2A_MSIL2A_20170613T101031_N9999_R022_T33UUP_26_57.
• country - country of the patch.
• tile, acquisition_time, satellite, relative_orbit, patch_row, patch_col - parsed from image_id (Sentinel-2 tile, sensing time (UTC), S2A / S2B, relative orbit, patch position in the tile grid).
• epsg, utm_x, utm_y - projection (UTM zone EPSG code) and top-left corner of the patch in metres, from the source GeoTIFF.
• lat, lon - WGS84 coordinates of the patch centre.
• s1_name - the paired Sentinel-1 patch in reBEN (BigEarthNet-S1).
• s2v1_name - the corresponding BigEarthNet v1.0 patch name.

Usage

This dataset follows the same multi-label format as timm/plant-pathology-2021 for use with timm 1.0.31 or greater:

python train.py --dataset hfds/timm/bigearthnet-v2-rgb --train-split train --val-split validation \
  --task multilabel --target-key labels --num-classes 19 \
  --model resnet50.ram_in1k --pretrained --img-size 128 --batch-size 256 --epochs 30 \
  --opt adamw --lr 1e-3 --weight-decay 0.05 --amp

The usual metrics are mAP (macro / micro average precision) and F1.

License

The source data is released under the Community Data License Agreement - Permissive, Version 1.0 (CDLA-Permissive-1.0). It contains modified Copernicus Sentinel data (2017-2018) and CORINE Land Cover 2018 data. This derived version (tone-mapped RGB, JPEG q98 (4:4:4)) is distributed under the same license.

Citation

@inproceedings{clasen2025reben,
  title = {reBEN: Refined BigEarthNet Dataset for Remote Sensing Image Analysis},
  author = {Clasen, Kai Norman and Hackel, Leonard and Burgert, Tom and Sumbul, Gencer and Demir, Beg{\"u}m and Markl, Volker},
  booktitle = {IEEE International Geoscience and Remote Sensing Symposium (IGARSS)},
  year = {2025},
  note = {arXiv:2407.03653}
}