giswqs/EuroSAT_MS

EuroSAT Multispectral (All 13 Sentinel-2 Bands)

Dataset Description

EuroSAT is a dataset for land use and land cover (LULC) classification using Sentinel-2 satellite imagery. This version contains all 13 Sentinel-2 spectral bands stored as uint16 arrays at 64x64 pixel resolution.

The dataset covers 10 land use/land cover classes across 26,998 geo-referenced images from 34 European countries.

• Source: https://zenodo.org/records/7711810
• DOI: 10.5281/zenodo.7711810
• License: MIT
• Paper: EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification
• RGB Version: giswqs/EuroSAT_RGB

Authors

Patrick Helber, Benjamin Bischke, Andreas Dengel, Damian Borth

Spectral Bands

Each image is a numpy array of shape (13, 64, 64) with dtype uint16. The 13 bands correspond to the Sentinel-2 spectral bands:

Index	Band	Sentinel-2 Band	Wavelength (nm)	Resolution (m)
0	B01	Coastal aerosol	443	60
1	B02	Blue	490	10
2	B03	Green	560	10
3	B04	Red	665	10
4	B05	Veg. Red Edge 1	705	20
5	B06	Veg. Red Edge 2	740	20
6	B07	Veg. Red Edge 3	783	20
7	B08	NIR	842	10
8	B08A	Narrow NIR	865	20
9	B09	Water Vapour	945	60
10	B10	SWIR - Cirrus	1375	60
11	B11	SWIR 1	1610	20
12	B12	SWIR 2	2190	20

Note: All bands are resampled to 10m resolution (64x64 pixels) in the original dataset.

Dataset Structure

Splits

Split	Examples
train	18,880
validation	5,405
test	2,713

Classes

Label	Class Name
0	AnnualCrop
1	Forest
2	HerbaceousVegetation
3	Highway
4	Industrial
5	Pasture
6	PermanentCrop
7	Residential
8	River
9	SeaLake

Features

• image: Array3D(shape=(13, 64, 64), dtype="uint16") — 13-band Sentinel-2 multispectral image
• label: ClassLabel — Integer class label (0–9)
• filename: Value("string") — Original filename with class directory prefix

Usage

from datasets import load_dataset
import numpy as np

dataset = load_dataset("giswqs/EuroSAT_MS")

# Access training split
train = dataset["train"]
sample = train[0]

# Get multispectral image as numpy array
image = np.array(sample["image"], dtype=np.uint16)  # shape: (13, 64, 64)
label = sample["label"]
filename = sample["filename"]

print(f"Image shape: {image.shape}, dtype: {image.dtype}")
print(f"Label: {label}, Filename: {filename}")

# Extract RGB bands (B04, B03, B02 = indices 3, 2, 1)
rgb = image[[3, 2, 1]]  # shape: (3, 64, 64)

# Compute NDVI
red = image[3].astype(np.float32)
nir = image[7].astype(np.float32)
ndvi = (nir - red) / (nir + red + 1e-8)

Citation

@article{helber2019eurosat,
  title={EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification},
  author={Helber, Patrick and Bischke, Benjamin and Dengel, Andreas and Borth, Damian},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing},
  volume={12},
  number={7},
  pages={2217--2226},
  year={2019},
  doi={10.1109/JSTARS.2019.2918242},
  publisher={IEEE}
}