calebrob6/similar-but-different

Similar But Different — Sentinel-2 patches with deceptive RGB

A curated set of 12,474 32×32 Sentinel-2 L2A multispectral patches (12 bands resampled to 10 m) selected because they participate in at least one pair of patches that look nearly identical in true-colour RGB but differ substantially in near-infrared (NIR) reflectance.

This makes the dataset a small but useful stress-test for models that rely on the visible bands alone: many "obvious" classes (forest vs. cropland, water vs. dark roof) become genuinely hard when only RGB is available.

How it was built

1. Randomly sampled 32×32 windows from 2627 Sentinel-2 L2A scenes across 82 hand-picked global seeds (forest, urban, agriculture, water, desert, wetland, mountain, savanna).
2. Computed per-channel RGB means on the small 8-bit visual asset and used sklearn.neighbors.NearestNeighbors to find every pair within an L2 distance of 2 DN.
3. For pairs that passed the RGB filter, downloaded the B05–B07 (red edge) and B08 (NIR) bands and computed nir_mean_diff = |B08_a - B08_b|.
4. Looked up each patch's ESA WorldCover 10 m 2021 land-cover distribution and assigned the mode class as the patch label.
5. Kept all patches that appear in any pair satisfying min_mode_purity > 50% AND nir_mean_diff > 1000 (uint16 surface reflectance units) AND wc_mode_label != Unknown.
6. Wrote each patch as a 12-band uint16 deflate-compressed GeoTIFF in <wc_mode_label>/<patch_id>.tif.

Full pipeline source: https://github.com/calebrob6/find_similar_patches (forthcoming).

File layout

similar-but-different/
├── README.md                          # this file
├── splits.json                        # stratified 80/10/10 train/val/test
├── landcover_distribution.parquet     # per-patch WorldCover percentages
├── Grassland/                  # 1,942 patches
├── Cropland/                  # 1,887 patches
├── Permanent_water/                  # 1,635 patches
├── Herbaceous_wetland/                  # 1,541 patches
├── Shrubland/                  # 1,436 patches
├── Tree_cover/                  # 1,380 patches
├── Bare_sparse_veg/                  # 1,143 patches
├── Mangroves/                  # 802 patches
├── Built-up/                  # 392 patches
├── Moss_and_lichen/                  # 316 patches

Each <class>/<patch_id>.tif is a single 12-band uint16 GeoTIFF, 32×32 @ 10 m, georeferenced in the source scene's UTM CRS, nodata=0, compress=deflate. The GeoTIFF tags carry patch_id, scene_id, wc_mode_label, wc_mode_pct, and the patch centre in UTM coordinates.

Band order

#	Band	Wavelength (nm)	Native res (m)
1	B01 (coastal aerosol)	443	60
2	B02 (blue)	490	10
3	B03 (green)	560	10
4	B04 (red)	665	10
5	B05 (red edge 1)	705	20
6	B06 (red edge 2)	740	20
7	B07 (red edge 3)	783	20
8	B08 (NIR)	842	10
9	B8A (narrow NIR)	865	20
10	B09 (water vapour)	945	60
11	B11 (SWIR 1)	1610	20
12	B12 (SWIR 2)	2190	20

Non-10 m bands were resampled to 10 m with bilinear interpolation. (B10 cirrus is not provided in Sentinel-2 L2A so it is not included.)

Class distribution

Land cover (WorldCover mode)	n patches
Grassland	1,942
Cropland	1,887
Permanent water	1,635
Herbaceous wetland	1,541
Shrubland	1,436
Tree cover	1,380
Bare/sparse veg.	1,143
Mangroves	802
Built-up	392
Moss and lichen	316

Splits

A stratified-by-class 80/10/10 random split (seed = 42) is provided in splits.json. Rare classes are guaranteed at least one patch per split when possible.

split	n
train	9,978
val	1,248
test	1,248

The same split file also stores inverse-frequency class weights (clipped at 10.0) for use with weighted cross-entropy.

Quick start

import rasterio, numpy as np, json
from datasets import load_dataset

# Native HF datasets ImageFolder mode (one folder per class):
ds = load_dataset(
    "imagefolder",
    data_dir="<path to the cloned repo>",
)
# Or just download the TIFs directly:
from huggingface_hub import snapshot_download
local = snapshot_download(repo_id="calebrob6/similar-but-different", repo_type="dataset")

# Read a TIF:
with rasterio.open(f"{local}/Tree_cover/<some_patch_id>.tif") as src:
    arr = src.read()                  # (12, 32, 32) uint16
    print(src.descriptions[7])        # 'B08 NIR 842nm'
    print(src.tags())                 # {'patch_id': ..., 'wc_mode_label': ...}

# Stratified split:
splits = json.loads(open(f"{local}/splits.json").read())
train_ids = sum(splits["train"].values(), [])

Baseline results

ResNet18 fine-tuning (ImageNet weights, AdamW + cosine LR, 50 epochs, class- weighted CE) on this split:

Bands used	in_channels	Test accuracy	Macro F1	Weighted F1
RGB (B04, B03, B02)	3	0.843	0.571	0.852
RGB + NIR (+ B08)	4	0.886	0.638	0.892
All 12 bands	12	0.886	0.693	0.897

NIR alone closes most of the gap; the full 12-band stack mainly helps the long-tail classes (Mangroves, Bare/sparse veg., Built-up).

License

Released under CC BY 4.0, matching the upstream Sentinel-2 L2A data (Copernicus open access) and ESA WorldCover (CC BY 4.0).

Attribution: Contains modified Copernicus Sentinel-2 L2A data (2024–2025); land-cover labels derived from © ESA WorldCover 2021 v200.

Citation

@misc{similar-but-different-2025,
  title  = {Similar But Different: a Sentinel-2 dataset of RGB-deceptive patches},
  author = {Robinson, Caleb},
  year   = {2025},
  url    = {https://huggingface.co/datasets/calebrob6/similar-but-different}
}

Parquet shard layout (recommended for loading)

For fast loading, the dataset is also packaged as per-class parquet shards under shards/<class>.parquet. Each row contains the raw 12-band GeoTIFF as tif bytes plus the metadata columns.

import io, pandas as pd, rasterio
shard = pd.read_parquet("shards/Tree_cover.parquet")
row = shard.iloc[0]
with rasterio.open(io.BytesIO(row["tif"])) as src:
    arr = src.read()                                # (12, 32, 32) uint16
    print(src.descriptions[7])                      # 'B08 NIR 842nm'

Or with the HF datasets library:

from datasets import load_dataset
ds = load_dataset("calebrob6/similar-but-different",
                  data_files="shards/*.parquet")