EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification

Abstract

In this paper, we address the challenge of land use and land cover classification using Sentinel-2 satellite images. The Sentinel-2 satellite images are openly and freely accessible provided in the Earth observation program Copernicus. We present a novel dataset based on Sentinel-2 satellite images covering 13 spectral bands and consisting out of 10 classes with in total 27,000 labeled and geo-referenced images. We provide benchmarks for this novel dataset with its spectral bands using state-of-the-art deep Convolutional Neural Network (CNNs). With the proposed novel dataset, we achieved an overall classification accuracy of 98.57%. The resulting classification system opens a gate towards a number of Earth observation applications. We demonstrate how this classification system can be used for detecting land use and land cover changes and how it can assist in improving geographical maps. The geo-referenced dataset EuroSAT is made publicly available at https://github.com/phelber/eurosat.

Full text

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 1
EuroSAT: A Novel Dataset and Deep Learning
Benchmark for Land Use and Land Cover
Classification
Patrick Helber1,2 Benjamin Bischke1,2 Andreas Dengel1,2 Damian Borth2
1TU Kaiserslautern, Germany 2German Research Center for Artificial Intelligence (DFKI), Germany
{Patrick.Helber, Benjamin.Bischke, Andreas.Dengel, Damian.Borth}@dfki.de
Abstract—In this paper, we address the challenge of land use
and land cover classification using Sentinel-2 satellite images. The
Sentinel-2 satellite images are openly and freely accessible pro-
vided in the Earth observation program Copernicus. We present
a novel dataset based on Sentinel-2 satellite images covering 13
spectral bands and consisting out of 10 classes with in total 27,000
labeled and geo-referenced images. We provide benchmarks for
this novel dataset with its spectral bands using state-of-the-art
deep Convolutional Neural Network (CNNs). With the proposed
novel dataset, we achieved an overall classification accuracy of
98.57%. The resulting classification system opens a gate towards
a number of Earth observation applications. We demonstrate
how this classification system can be used for detecting land
use and land cover changes and how it can assist in improving
geographical maps. The geo-referenced dataset EuroSAT is made
publicly available at https://github.com/phelber/eurosat.
Index Terms—Remote Sensing, Earth Observation, Satellite
Images, Satellite Image Classification, Land Use Classification,
Land Cover Classification, Dataset, Machine Learning, Deep
Learning, Deep Convolutional Neural Network
I. INTRODUCTION
WE are currently at the edge of having public and
continuous access to satellite image data for Earth
observation. Governmental programs such as ESA’s Coper-
nicus and NASA’s Landsat are taking significant efforts to
make such data freely available for commercial and non-
commercial purpose with the intention to fuel innovation and
entrepreneurship. With access to such data, applications in
the domains of agriculture, disaster recovery, climate change,
urban development, or environmental monitoring can be real-
ized [37], [2], [3], [5]. However, to fully utilize the data for
the previously mentioned domains, first satellite images must
be processed and transformed into structured semantics [35].
One type of such fundamental semantics is Land Use and
Land Cover Classification [1], [29]. The aim of land use
and land cover classification is to automatically provide labels
describing the represented physical land type or how a land
area is used (e.g., residential, industrial).
As often in supervised machine learning, the performance
of classification systems strongly depends on the availability
of high-quality datasets with a suitable set of classes [21].
In particular when considering the recent success of deep
Convolutional Neural Networks (CNN) [12], it is crucial to
have large quantities of training data available to train such
Fig. 1: Land use and land cover classification based on
Sentinel-2 satellite images. Patches are extracted with the pur-
pose to identify the shown class. This visualization highlights
the classes annual crop, river, highway, industrial buildings
and residential buildings.
a network. Unfortunately, current land use and land cover
datasets are small-scale or rely on data sources which do not
allow the mentioned domain applications.
In this paper, we propose a novel satellite image dataset for
the task of land use and land cover classification. The proposed
EuroSAT dataset consists of 27,000 labeled images with 10
different land use and land cover classes. A significant differ-
ence to previous datasets is that the presented satellite image
dataset is multi-spectral covering 13 spectral bands in the visi-
ble, near infrared and short wave infrared part of the spectrum.
In addition, the proposed dataset is georeferenced and based on
openly and freely accessible Earth observation data allowing a
unique range of applications. The labeled dataset EuroSAT is
made publicly available at https://github.com/phelber/eurosat.
Further, we provide a full benchmark demonstrating a robust
classification performance which is the basis for developing
applications for the previously mentioned domains. We outline
how the classification model can be used for detecting land
use or land cover changes and how it can assist in improving
geographical maps.
We provide this work in the context of the recently pub-
lished EuroSAT dataset, which can be used similar to [18] as
a basis for a large-scale training of deep neural networks for

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 2
Fig. 2: This illustration shows an overview of the patch-based land use and land cover classification process using satellite
images. A satellite scans the Earth to acquire images of it. Patches extracted out of these images are used for classification.
The aim is to automatically provide labels describing the represented physical land type or how the land is used. For this
purpose, an image patch is feed into a classifier, in this illustration a neural network, and the classifier outputs the class shown
on the image patch.
the task of satellite image classification.
In this paper, we make the following contributions:
•We introduce the first large-scale patch-based land
use and land cover classification dataset based on
Sentinel-2 satellite images. Every image in the
dataset is labeled and geo-referenced. We release
the RGB and the multi-spectral (MS) version of
the dataset.
•We provide benchmarks for the proposed Eu-
roSAT dataset using Convolutional Neural Net-
works (CNNs).
•We evaluate the performance of each spectral band
of the Sentinel-2 satellite for the task of patch-
based land use and land cover classification.
II. RE LATED WORK
In this section, we review previous studies in land use and
land cover classification. In this context, we present remotely
sensed aerial and satellite image datasets. Furthermore, we
present state-of-the-art image classification methods for land
use and land cover classification.
A. Classification Datasets
The classification of remotely sensed images is a challeng-
ing task. The progress of classification in the remote sensing
area has particularly been inhibited due to the lack of reli-
ably labeled ground truth datasets. A popular and intensively
studied [6], [19], [20], [27], [29] remotely sensed image classi-
fication dataset known as UC Merced (UCM) land use dataset
was introduced by Yang et al. [29]. The dataset consists of 21
land use and land cover classes. Each class has 100 images and
the contained images measure 256x256 pixels with a spatial
resolution of about 30 cm per pixel. All images are in the RGB
color space and were extracted from the USGS National Map
Urban Area Imagery collection, i.e. the underlying images
were acquired from an aircraft. Unfortunately, a dataset with
100 images per class is small-scale. Trying to enhance the
dataset situation, various works used commercial Google Earth
images to manually create novel datasets [22], [27], [28],
[30] such as the two benchmark datasets PatternNet [39] and
NWPU-RESISC45 [36]. The datasets are based on very-high-
resolution images with a spatial resolution of up to 30 cm
per pixel. Since the creation of a labeled dataset is extremely
time-consuming, these datasets consist likewise of only a few
hundred images per class. One of the largest datasets is the
Aerial Image Dataset (AID). AID consists of 30 classes with
200 to 400 images per class. The 600x600 high-resolution
images were also extracted from Google Earth imagery.
Compared to the EuroSAT dataset presented in this work,
the previously listed datasets rely on commercial very-high-
resolution and preprocessed images. The fact of using com-
mercial and preprocessed very-high-resolution image data
makes these datasets unsatisfying for real-world Sentinel-
2 Earth observation applications as proposed in this work.
Furthermore, while these datasets put a strong focus on
strengthening the number of covered classes, the datasets
suffer from a low number of images per class. The fact of a
spatial resolution of up to 30 cm per pixel, with the possibility
to identify and distinguish classes like churches, schools etc.,
make the presented datasets difficult to compare with the
dataset proposed in this work.
A study closer to our work, provided by Penatti et al.
[20], analyzed remotely sensed satellite images with a spatial
resolution of 10 meters per pixel to classify coffee crops.
Based on these images, Penatti et al. [20] introduced the
Brazillian Coffee Scene (BCS) dataset. The dataset covers
the two classes coffee crop and non-coffee crop. Each class
consists of 1,423 images. The images consist of a red, green
and near-infrared band.
Similar to the proposed EuroSAT dataset, Basu et al. [1]
introduced the SAT-6 dataset relying on aerial images. This
dataset has been extracted from images with a spatial reso-
lution of 1 meter per pixel. The image patches are created
using images from the National Agriculture Imagery Program
(NAIP). SAT-6 covers the 6 different classes: barren land,
trees, grassland, roads, buildings and water bodies. The pro-
posed patches have a size of 28x28 pixels per image and
consist of a red, green, blue and a near-infrared band.
B. Land Use and Land Cover Classification
Convolutional Neural Networks (CNNs) are a type of Neu-
ral Networks [13], which became with the impressive results
on image classification challenges [12], [21], [23] the state-
of-the-art image classification method in computer vision and
machine learning.

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 3
Fig. 3: The diagram illustrates the EuroSAT dataset creation
process.
To classify remotely sensed images, various different feature
extraction and classification methods (e.g., Random Forests)
were evaluated on the introduced datasets. Yang et al. eval-
uated Bag-of-Visual-Words (BoVW) and spatial extension
approaches on the UCM dataset [29]. Basu et al. analyzed
deep belief networks, basic CNNs and stacked denoising au-
toencoders on the SAT-6 dataset [1]. Basu et al. also presented
an own framework for the land cover classes introduced in
the SAT-6 dataset. The framework extracts features from the
input images, normalizes the extracted features and used the
normalized features as input to a deep belief network. Besides
low-level color descriptors, Penatti et al. also evaluated deep
CNNs on the UCM and BCS dataset [20]. In addition to
deep CNNs, Castelluccio et al. intensively evaluated various
machine learning methods (e.g., Bag-of-Visual-Words, spatial
pyramid match kernels) for the classification of the UCM and
BCS dataset.
In the context of deep learning, the used deep CNNs have
been trained from scratch or fine-tuned by using a pretrained
network [6], [19], [31], [36], [16]. The networks were mainly
pretrained on the ILSVRC-2012 image classification chal-
lenge [21] dataset. Even though these pretrained networks
were trained on images from a totally different domain, the
features generalized well. Therefore, the pretrained networks
proved to be suitable for the classification of remotely sensed
images [17]. The presented works extensively evaluated all
proposed machine learning methods and concluded that that
deep CNNs outperform non-deep learning approaches on the
considered datasets [6], [17], [15], [27].
III. DATAS ET ACQUISITION
Besides NASA with its Landsat Mission, the European
Space Agency (ESA) steps up efforts to improve Earth ob-
servation within its Copernicus program. Under this program,
ESA operates a series of satellites known as Sentinels.
In this paper, we use mutli-spectral image data provided
by the Sentinel-2A satellite in order to address the challenge
of land use and land cover classification. Sentinel-2A is
one satellite in the two-satellite constellation of the identical
land monitoring satellites Sentinel-2A and Sentinel-2B. The
satellites were successfully launched in June 2015 (Sentinel-
2A) and March 2017 (Sentinel-2B). Both sun-synchronous
satellites capture the global Earth’s land surface with a Multi-
spectral Imager (MSI) covering the 13 different spectral bands
listed in Table I. The three bands B01, B09 and B10 are
intended to be used for the correction of atmospheric effects
(e.g., aerosols, cirrus or water vapor). The remaining bands are
primarily intended to identify and monitor land use and land
cover classes. In addition to mainland, large islands as well as
inland and coastal waters are covered by these two satellites.
Each satellite will deliver imagery for at least 7 years with a
spatial resolution of up to 10 meters per pixel. Both satellites
carry fuel for up to 12 years of operation which allows for
an extension of the operation. The two-satellite constellation
generates a coverage of almost the entire Earth’s land surface
about every five days, i.e. the satellites capture each point in
the covered area about every five days. This short repeat cycle
as well as the future availability of the Sentinel satellites allows
a continuous monitoring of the Earth’s land surface for about
the next 20 - 30 years. Most importantly, the data is openly
and freely accessible and can be used for any application
(commercial or non-commercial use).
We are convinced that the large volume of satellite data
in combination with powerful machine learning methods will
influence future research. Therefore, one of our key research
aims is to make this large amount of data accessible for
machine learning based applications. To construct an image
classification dataset, we performed the following two steps:
1) Satellite Image Acquisition: We gathered satellite images
of European cities distributed in over 34 countries as
shown in Fig. 5.
2) Dataset Creation: Based on the obtained satellite images,
we created a dataset of 27,000 georeferenced and labeled
image patches. The image patches measure 64x64 pixels
and have been manually checked.
A. Satellite Image Acquisition
We have downloaded satellite images taken by the satel-
lite Sentinel-2A via Amazon S3. We chose satellite images
associated with the cities covered in the European Urban
Atlas. The covered cities are distributed over the 34 Euro-
pean countries: Austria, Belarus, Belgium, Bulgaria, Cyprus,
Czech Republic (Czechia), Denmark, Estonia, Finland, France,
Germany, Greece, Hungary, Iceland, Ireland, Italy / Holy See,
Latvia, Lithuania, Luxembourg, Macedonia, Malta, Republic
of Moldova, Netherlands, Norway, Poland, Portugal, Romania,

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 4
(a) Industrial Buildings (b) Residential Buildings (c) Annual Crop (d) Permanent Crop (e) River
(f) Sea & Lake (g) Herbaceous Vegetation (h) Highway (i) Pasture (j) Forest
Fig. 4: This overview shows sample image patches of all 10 classes covered in the proposed EuroSAT dataset. The images
measure 64x64 pixels. Each class contains 2,000 to 3,000 image. In total, the dataset has 27,000 geo-referenced images.
Slovakia, Slovenia, Spain, Sweden, Switzerland, Ukraine and
United Kingdom.
In order to improve the chance of getting valuable image
patches, we selected satellite images with a low cloud level.
Besides the possibility to generate a cloud mask, ESA provides
a cloud level value for each satellite image allowing to quickly
select images with a low percentage of clouds covering the
land scene.
We aimed for the objective to cover as many countries
as possible in the EuroSAT dataset in order to cover the
high intra-class variance inherent to remotely sensed images.
Furthermore, we have extracted images recorded all over the
year to get a variance as high as possible inherent in the
covered land use and land cover classes. Within one class of
the EuroSAT dataset, different land types of this class are rep-
resented such as different types of forests in the forest class or
different types of industrial buildings in the industrial building
class. Between the classes, there is a low positive correlation.
The classes most common to each other are the two presented
agricultural classes and the two classes representing residential
and industrial buildings. The composition of the individual
classes and their relationships are specified in the mapping
guide of the European Urban Atlas [40]. An overview diagram
of the dataset creation process is shown in Fig. 3
B. Dataset Creation
The Sentinel-2 satellite constellation provides about 1.6
TB of compressed images per day. Unfortunately, supervised
machine learning is restricted even with this amount of data
by the lack of labeled ground truth data. The generation of
the benchmarking EuroSAT dataset was motivated by the
objective of making this open and free satellite data accessible
to various Earth observation applications and the observation
that existing benchmark datasets are not suitable for the
intended applications with Sentinel-2 satellite images. The
dataset consists of 10 different classes with 2,000 to 3,000
images per class. In total, the dataset has 27,000 images. The
patches measure 64x64 pixels. We have chosen 10 different
TABLE I: All 13 bands covered by Sentinel-2’s Multispectral
Imager (MSI). The identification, the spatial resolution and the
central wavelength is listed for each spectral band.
Band Spatial Central
Resolution Wavelength
m nm
B01 - Aerosols 60 443
B02 - Blue 10 490
B03 - Green 10 560
B04 - Red 10 665
B05 - Red edge 1 20 705
B06 - Red edge 2 20 740
B07 - Red edge 3 20 783
B08 - NIR 10 842
B08A - Red edge 4 20 865
B09 - Water vapor 60 945
B10 - Cirrus 60 1375
B11 - SWIR 1 20 1610
B12 - SWIR 2 20 2190
land use and land cover classes based on the principle that they
showed to be visible at the resolution of 10 meters per pixel
and are frequently enough covered by the European Urban
Atlas to generate thousands of image patches. To differentiate
between different agricultural land uses, the proposed dataset
covers the classes annual crop, permanent crop (e.g., fruit
orchards, vineyards or olive groves) and pastures. The dataset
also discriminates built-up areas. It therefore covers the classes
highway, residential buildings and industrial buildings. The
residential class is created using the urban fabric classes
described in the European Urban Atlas. Different water bodies
appear in the classes river and sea & lake. Furthermore, unde-
veloped environments such as forest and herbaceous vegetation
are included. An overview of the covered classes with four
samples per class is shown in Fig. 4.
We manually checked all 27,000 images multiple times and
corrected the ground truth by sorting out mislabeled images
as well as images full of snow or ice. Example images, which
have been discarded, are shown in Fig. 6. The samples are
intended to show industrial buildings. Clearly, no industrial

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 5
TABLE II: Classification accuracy (%) of different training-test splits on the EuroSAT dataset.
Method 10/90 20/80 30/70 40/60 50/50 60/40 70/30 80/20 90/10
BoVW (SVM, SIFT, k = 10) 54.54 56.13 56.77 57.06 57.22 57.47 57.71 58.55 58.44
BoVW (SVM, SIFT, k = 100) 63.07 64.80 65.50 66.16 66.25 66.34 66.50 67.22 66.18
BoVW (SVM, SIFT, k = 500) 65.62 67.26 68.01 68.52 68.61 68.74 69.07 70.05 69.54
CNN (two layers) 75.88 79.84 81.29 83.04 84.48 85.77 87.24 87.96 88.66
ResNet-50 75.06 88.53 93.75 94.01 94.45 95.26 95.32 96.43 96.37
GoogleNet 77.37 90.97 90.57 91.62 94.96 95.54 95.70 96.02 96.17
Fig. 5: EuroSAT dataset distribution. The georeferenced im-
ages are distributed all over Europe. The distribution is influ-
enced by the number of represented cities per country in the
European Urban Atlas.
building is visible. Please note, the proposed dataset has not
received atmospheric correction. This can result in images with
a color cast. Extreme cases are visualized in Fig. 7. With the
intention to advocate the classifier to also learn these cases,
we did not filter the respective samples and let them flow into
the dataset.
Besides the 13 covered spectral bands, the new dataset has
three further central innovations. Firstly, the dataset is not
based on non-free satellite images like Google Earth imagery
or relies on data sources which are not updated on a high-
frequent basis (e.g., NAIP used in [1]). Instead, an open
and free Earth observation program whose satellites deliver
images for the next 20 - 30 years is used allowing real-
world Earth observation applications. Secondly, the dataset
uses a 10 times lower spatial resolution than the benchmark
dataset closest to our research but at once distinguishes 10
classes instead of 6. For instance, we split up the built-up
class into a residential and an industrial class or distinguish
between different agricultural land uses. Thirdly, we release
the EuroSAT dataset in a georeferenced version.
With the release of the geo-referenced EuroSAT we aim to
make the large amount of Sentinel-2 satellite imagery accessi-
ble for machine learning approaches. There effectiveness was
successfully demonstrated in [32], [33], [34].
IV. DATASE T BENCHMARKING
As shown in previous work [6], [15], [17], [19], deep
CNNs have demonstrated to outperform non-deep learning
approaches in land use and land cover image classification.
Accordingly, we use the state-of-the-art deep CNN models
GoogleNet [25] and ResNet-50 [9], [10] for the classifica-
tion of the introduced land use and land cover classes. The
networks make use of the inception module [25], [26], [24],
[14] and the residual unit [9], [10]. For the proposed EuroSAT
dataset, we also evaluated the performance of the 13 spectral
bands with respect to the classification task. In this context,
we evaluate the classification performance using single-band
and band combination images.
TABLE III: Benchmarked classification accuracy (%) of the
two best performing classifiers GoogLeNet and ResNet-50
with a 80/20 training-test split. Both CNNs have been pre-
trained on the image classification dataset ILSVRC-2012 [21].
Method UCM AID SAT-6 BCS EuroSAT
ResNet-50 96.42 94.38 99.56 93.57 98.57
GoogLeNet 97.32 93.99 98.29 92.70 98.18
A. Comparative Evaluation
We respectively split each dataset in a training and a test set
(80/20 ratio). We ensured that the split is applied class-wise.
While the red, green and blue bands are covered by almost
all aerial and satellite image datasets, the proposed EuroSAT
dataset consists of 13 spectral bands. For the comparative
evaluation, we computed images in the RGB color space
combining the bands red (B04), green (B03) and blue (B02).
For benchmarking, we evaluated the performance of the Bag-
of-Visual-Words (BoVW) approach using SIFT features and a
trained SVM. In addition, we trained a shallow Convolutional
Neural Network (CNN), a ResNet-50 and a GoogleNet model
on the training set. We calculated the overall classification
accuracy to evaluate the performance of the different models
on the considered datasets. In Table II we show how the
approaches perform in case of different training-test splits for
the EuroSAT RGB dataset.
It can be seen that all CNN approaches outperform the
BoVW method and, overall, deep CNNs perform better than
shallow CNNs. Nevertheless, the shallow CNN classifies the
EuroSAT classes with a classification accuracy of up to
89.03%. Please note [6], [19], [22] for the benchmarking
performance of the other datasets on different training-test
splits.
Table III lists the achieved classification results for the two
best performing CNN models GoogLeNet and ResNet-50. In

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 6
Fig. 6: Four examples of bad image patches, which are
intended to show industrial buildings. Clearly, no industrial
building is shown due to clouds, mislabeling, dead pixels or
ice/snow.
Fig. 7: Color cast due to atmospheric effects.
this experiment, the GoogleNet and ResNet-50 CNN models
were pretrained on the ILSVRC-2012 image classification
dataset [21]. In all fine-tuning experiments, we first trained
the last layer with a learning rate of 0.01. Afterwards, we
fine-tuned through the entire network with a low learning
rate between 0,001 and 0,0001. With a finetuned network we
achieve a classification accuracy of about 2% higher compared
to randomly initialized versions of the networks which have
been trained on the EuroSAT dataset with the same training-
test split (see Table II).
The deep CNNs achieve state-of-the-art results on the UCM
dataset and outperform previous results on the other three
presented datasets by about 2-4% (AID, SAT-6, BCS) [6],
[19], [22]. Table III shows that the ResNet-50 architecture
performs best on the introduced EuroSAT land use and land
cover classes. In order to allow an evaluation on the class
level, Fig. 8 shows the confusion matrix of this best performing
network. It is shown that the classifier sometimes confuses the
agricultural land classes as well as the classes highway and
river.
B. Band Evaluation
In order to evaluate the performance of deep CNNs using
single-band images as well shortwave-infrared and color-
infrared band combinations, we used the pretrained ResNet-50
with a fixed training-test split to compare the performance
of the different spectral bands. For the single-band image
evaluation, we used images as input consisting of the in-
formation gathered from a single spectral band on all three
input channels. We analyzed all spectral bands, even the bands
not intended for land monitoring. Bands with a lower spatial
resolution have been upsampled to 10 meters per pixel using
cubic-spline interpolation [8]. Fig. 9 shows a comparison of
the spectral band’s performance. It is shown that the red,
green and blue bands outperform all other bands. Interestingly,
the bands red edge 1 (B05) and shortwave-infrared 2 (B12)
with an original spatial resolution of merely 20 meters per
pixel showed an impressive performance. The two bands even
outperform the near-infrared band (B08) which has a spatial
resolution of 10 meters per pixel.
In addition to the RGB band combination, we also analyzed
the performance of the shortwave-infrared and color-infrared
An. Crop
Forest
Herbaceous
Highway
Industrial
Pasture
Per. Crop
Residential
River
Sea & Lake
Predicted label
An. Crop
Forest
Herbaceous
Highway
Industrial
Pasture
Per. Crop
Residential
River
Sea & Lake
True label
0.98 0.0 0.0 0.0 0.0 0.0 0.02 0.0 0.0 0.0
0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
0.0 0.0 0.99 0.0 0.0 0.01 0.0 0.0 0.0 0.0
0.0 0.0 0.0 0.98 0.0 0.0 0.0 0.0 0.02 0.0
0.0 0.0 0.0 0.0 0.99 0.0 0.0 0.01 0.0 0.0
0.0 0.0 0.02 0.0 0.0 0.98 0.0 0.0 0.0 0.0
0.0 0.0 0.02 0.0 0.0 0.0 0.98 0.0 0.0 0.0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0
0.0 0.0 0.0 0.01 0.0 0.0 0.0 0.0 0.98 0.01
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0
0.0
0.2
0.4
0.6
0.8
1.0
Fig. 8: Confusion matrix of a fine-tuned ResNet-50 CNN on
the proposed EuroSAT dataset using satellite images in the
RGB color space.
TABLE IV: Classification accuracy (%) of a fine-tuned
ResNet-50 CNN on the proposed EuroSAT dataset with
the three different band combinations color-infrared (CI),
shortwave-infrared (SWIR) and RGB as input.
Band Combination Accuracy (ResNet-50)
CI 98.30
RGB 98.57
SWIR 97.05
B01
B02
B03
B04
B05
B06
B07
B08
B8A
B09
B10
B11
B12
Band
0
20
40
60
80
100
Accuracy
Fig. 9: Overall classification accuracy (%) of a fine-tuned
ResNet-50 CNN on the given EuroSAT dataset using single-
band images.
band combination. Table IV shows a comparison of the per-
formance of these combinations. As shown, band combination
images outperform single-band images. Furthermore, images
in the RGB color space performed best on the introduced land
use and land cover classes. Please note, networks pretrained
on the ILSVRC-2012 image classification dataset have initially
not been trained on images other than RGB images.
V. APPLICATIONS
The openly and freely accessible satellite images allow
a broad range of possible applications. In this section, we
demonstrate that the novel dataset published with this paper
allows real-world applications. The classification result with

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 7
an overall accuracy of 98.57% paves the way for these appli-
cations. We show land use and land cover change detection
applications as well as how the the trained classifier can assist
in keeping geographical maps up-to-date.
A. Land Use and Land Cover Change Detection
Since the Sentinel-2 satellite constellation will scan the
Earth’s land surface for about the next 20 - 30 years on a
repeat cycle of about five days, a trained classifier can be used
for monitoring land surfaces and detect changes in land use
or land cover. To demonstrate land use and land cover change
detection, we selected images from the same spatial region but
from different points in time. Using the trained classifier, we
analyzed 64x64 image regions. A change has taken place if
the classifier delivers different classification results for patches
taken from the same spatial 64x64 region. In the following, we
show three examples of spotted changes. In the first example
shown in Fig. 10, the classification system recognized that
the land cover has changed in the highlighted area. The left
image was acquired in the surroundings of Shanghai, China in
December 2015 showing an area classified as industrial. The
right image shows the same area in December 2016 revealing
that the industrial buildings have been demolished. The second
example is illustrated in Fig. 11. The left image was acquired
in the surroundings of Dallas, USA in August 2015 showing
no dominant residential buildings in the highlighted area.
The right image shows the same area in March 2017. The
system has identified a change in the highlighted area revealing
that residential buildings have been constructed. The third
example presented in Fig. 12 shows that the system detected
deforestation near Villamontes, Bolivia. The left image was
acquired in October 2015. The right image shows the same
region in September 2016 revealing that a large area has been
deforested. The presented examples are particularly of interest
in urban area development, nature protection or sustainable
development. For instance, deforestation is a main contributor
to climate change, therefore the detection of deforested land is
of particular interest (e.g., to notice illegal clearing of forests).
Fig. 10: The left image was acquired in the surroundings
of Shanghai in December 2015 showing an area classified
as industrial. The right image shows the same region in
December 2016 revealing that the industrial buildings have
been demolished.
Fig. 11: The left image was acquired in the surroundings of
Dallas, USA in August 2015 showing no dominant residential
buildings in the highlighted area. The right image shows the
same area in March 2017 showing that residential buildings
have been built up.
Fig. 12: The left image was acquired near Villamontes, Bolivia
in October 2015. The right image shows the same area in
September 2016 revealing that a large land area has been
deforested.
B. Assistance in Mapping
While a classification system trained with 64x64 image
patches does not allow a finely graduated per-pixel segmenta-
tion, it cannot only detect changes as shown in the previous
examples, it can also facilitate keeping maps up-to-date. This
is an extremely helpful assistance with maps created in a
crowdsourced manner like OpenStreetMap (OSM). A possible
system can verify already tagged areas, identify mistagged
areas or bring large area tagging. The proposed system is
based on the trained CNN classifier providing a classification
result for each image patch created in a sliding windows based
manner.
As shown in Fig. 13, the industrial buildings seen in the
left up-to-date satellite image are almost completely covered
in the corresponding OSM mapping. The right up-to-date
satellite image also shows industrial buildings. However, a
major part of the industrial buildings is not covered in the
corresponding map. Due to the high temporal availability of
Sentinel-2 satellite images in the future, this work together
with the published dataset can be used to build systems which
assist in keeping maps up-to-date. A detailed analysis of the
respective land area can then be provided using high-resolution
satellite images and an advanced segmentation approach [4],
[11].

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 8
Fig. 13: A patch-based classification system can verify already tagged areas, identify mistagged areas or bring large area
tagging as shown in the above images and maps. The left Sentinel-2 satellite image was acquired in Australia in March 2017.
The right satellite image was acquired in the surroundings of Shanghai, China in March 2017. The corresponding up-to-date
OpenStreetMap (OSM) mapping images show that the industrial areas in the left satellite image are almost completely covered
(colored gray). However, the industrial areas in the right satellite image are not properly covered.
VI. CONCLUSION
In this paper, we have addressed the challenge of land
use and land cover classification. For this task, we presented
a novel dataset based on remotely sensed satellite images.
To obtain this dataset, we have used the openly and freely
accessible Sentinel-2 satellite images provided in the Earth
observation program Copernicus. The proposed dataset con-
sists of 10 classes covering 13 different spectral bands with in
total 27,000 labeled and geo-referenced images. We provided
benchmarks for this dataset with its spectral bands using state-
of-the-art deep Convolutional Neural Network (CNNs). For
this novel dataset, we analyzed the performance of the 13
different spectral bands. As a result of this evaluation, the
RGB band combination with an overall classification accuracy
of 98.57% outperformed the shortwave-infrared and the color-
infrared band combination and leads to a better classification
accuracy than all single-band evaluations. Overall, the avail-
able free Sentinel-2 satellite images offer a broad range of
possible applications. This work is a first important step to
make use of the large amount of available satellite data in
machine learning allowing to monitor Earth’s land surfaces
on a large scale. The proposed dataset can be leveraged for
multiple real-world Earth observation applications. Possible
applications are land use and land cover change detection or
the improvement of geographical maps.
ACKNOWLEDGMENT
This work was partially funded by the BMBF project
DeFuseNN (01IW17002). The authors thank NVIDIA for the
support within the NVIDIA AI Lab program.
REFERENCES
[1] S. Basu, S. Ganguly, S. Mukhopadhyay, R. DiBiano, M. Karki, and
R. Nemani. Deepsat: a learning framework for satellite imagery. In Pro-
ceedings of the 23rd SIGSPATIAL International Conference on Advances
in Geographic Information Systems, page 37. ACM, 2015.
[2] B. Bischke, P. Bhardwaj, A. Gautam, P. Helber, D. Borth, and A. Dengel.
Detection of Flooding Events in Social Multimedia and Satellite Imagery
using Deep Neural Networks. In MediaEval, 2017.
[3] B. Bischke, D. Borth, C. Schulze, and A. Dengel. Contextual Enrichment
of Remote-Sensed Events with Social Media Streams. In Proceedings of
the 2016 ACM on Multimedia Conference, pages 1077–1081. ACM, 2016.
[4] B. Bischke, P. Helber, J. Folz, D. Borth, and A. Dengel. Multi-Task
Learning for Segmentation of Buildings Footprints with Deep Neural
Networks. In arXiv preprint arXiv:1709.05932, 2017.
[5] B. Bischke, P. Helber, C. Schulze, V. Srinivasan, and D. Borth. The
Multimedia Satellite Task: Emergency Response for Flooding Events. In
MediaEval, 2017.
[6] M. Castelluccio, G. Poggi, C. Sansone, and L. Verdoliva. Land use
classification in remote sensing images by convolutional neural networks.
arXiv preprint arXiv:1508.00092, 2015.
[7] G. Cheng, J. Han, and X. Lu. Remote sensing image scene classification:
Benchmark and state of the art. Proceedings of the IEEE, 2017.
[8] C. De Boor, C. De Boor, E.-U. Math´
ematicien, C. De Boor, and
C. De Boor. A practical guide to splines, volume 27. Springer-Verlag
New York, 1978.
[9] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image
recognition. In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 770–778, 2016.
[10] K. He, X. Zhang, S. Ren, and J. Sun. Identity mappings in deep residual
networks. In European Conference on Computer Vision, pages 630–645.
Springer, 2016.
[11] M. Kampffmeyer, A.-B. Salberg, and R. Jenssen. Semantic segmentation
of small objects and modeling of uncertainty in urban remote sensing im-
ages using deep convolutional neural networks. In The IEEE Conference
on Computer Vision and Pattern Recognition (CVPR) Workshops, June
2016.
[12] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification
with deep convolutional neural networks. In Advances in neural infor-
mation processing systems, pages 1097–1105, 2012.
[13] Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard,
W. Hubbard, and L. D. Jackel. Backpropagation applied to handwritten
zip code recognition. Neural computation, 1(4):541–551, 1989.
[14] M. Lin, Q. Chen, and S. Yan. Network in network. arXiv preprint
arXiv:1312.4400, 2013.
[15] F. P. Luus, B. P. Salmon, F. van den Bergh, and B. Maharaj. Multiview
deep learning for land-use classification. IEEE Geoscience and Remote
Sensing Letters, 12(12):2448–2452, 2015.
[16] Z. Ma, Z. Wang, C. Liu, and X. Liu. Satellite imagery classification
based on deep convolution network. World Academy of Science, Engi-
neering and Technology, International Journal of Computer, Electrical,
Automation, Control and Information Engineering, 10(6):1113–1117,
2016.
[17] D. Marmanis, M. Datcu, T. Esch, and U. Stilla. Deep learning earth
observation classification using imagenet pretrained networks. IEEE
Geoscience and Remote Sensing Letters, 13(1):105–109, 2016.
[18] K. Ni, R. Pearce, K. Boakye, B. Van Essen, D. Borth, B. Chen, and
E. Wang. Large-scale deep learning on the yfcc100m dataset. arXiv
preprint arXiv:1502.03409, 2015.

JOURNAL OF L
A
T
E
X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 9
[19] K. Nogueira, O. A. Penatti, and J. A. dos Santos. Towards better exploit-
ing convolutional neural networks for remote sensing scene classification.
Pattern Recognition, 61:539–556, 2017.
[20] O. A. Penatti, K. Nogueira, and J. A. dos Santos. Do deep features
generalize from everyday objects to remote sensing and aerial scenes
domains? In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition Workshops, pages 44–51, 2015.
[21] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma,
Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-
Fei. ImageNet Large Scale Visual Recognition Challenge. International
Journal of Computer Vision (IJCV), 115(3):211–252, 2015.
[22] G. Sheng, W. Yang, T. Xu, and H. Sun. High-resolution satellite scene
classification using a sparse coding based multiple feature combination.
International journal of remote sensing, 33(8):2395–2412, 2012.
[23] K. Simonyan and A. Zisserman. Very deep convolutional networks for
large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
[24] C. Szegedy, S. Ioffe, V. Vanhoucke, and A. Alemi. Inception-v4,
inception-resnet and the impact of residual connections on learning. arXiv
preprint arXiv:1602.07261, 2016.
[25] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,
V. Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In
Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 1–9, 2015.
[26] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna. Rethinking
the inception architecture for computer vision. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, pages 2818–
2826, 2016.
[27] G.-S. Xia, J. Hu, F. Hu, B. Shi, X. Bai, Y. Zhong, and L. Zhang.
Aid: A benchmark dataset for performance evaluation of aerial scene
classification. arXiv preprint arXiv:1608.05167, 2016.
[28] G.-S. Xia, W. Yang, J. Delon, Y. Gousseau, H. Sun, and H. Maˆ
ıtre.
Structural high-resolution satellite image indexing. In ISPRS TC VII
Symposium-100 Years ISPRS, volume 38, pages 298–303, 2010.
[29] Y. Yang and S. Newsam. Bag-of-visual-words and spatial extensions
for land-use classification. In Proceedings of the 18th SIGSPATIAL
international conference on advances in geographic information systems,
pages 270–279. ACM, 2010.
[30] L. Zhao, P. Tang, and L. Huo. Feature significance-based multibag-of-
visual-words model for remote sensing image scene classification. Journal
of Applied Remote Sensing, 10(3):035004–035004, 2016.
[31] Kashif Ahmad, Konstantin Pogorelov, Michael Riegler, Nicola Conci,
and H Pal. Cnn and gan based satellite and social media data fusion for
disaster detection. In Proc. of the MediaEval 2017 Workshop, Dublin,
Ireland, 2017.
[32] Guanzhou Chen, Xiaodong Zhang, Xiaoliang Tan, Yufeng Cheng, Fan
Dai, Kun Zhu, Yuanfu Gong, and Qing Wang. Training small networks for
scene classification of remote sensing images via knowledge distillation.
Remote Sensing, 10(5):719, 2018.
[33] Subhankar Roy, Enver Sangineto, Nicu Sebe, and Beg¨
um Demir.
Semantic-fusion gans for semi-supervised satellite image classification.
In 2018 25th IEEE International Conference on Image Processing (ICIP),
pages 684–688. IEEE, 2018.
[34] Patrick Helber, Benjamin Bischke, Andreas Dengel, Damian Borth.
Introducing EuroSAT: A Novel Dataset and Deep Learning Benchmark
for Land Use and Land Cover Classification In Geoscience and Remote
Sensing Symposium (IGARSS), 2018 IEEE International. IEEE, 2018.
[35] Lanqing Huang, Bin Liu, Boying Li, Weiwei Guo, Wenhao Yu, Zenghui
Zhang, and Wenxian Yu. Opensarship: A dataset dedicated to sentinel-1
ship interpretation. IEEE Journal of Selected Topics in Applied Earth
Observations and Remote Sensing, 11(1):195–208, 2018.
[36] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote sensing image
scene classification: benchmark and state of the art. Proceedings of the
IEEE, 105(10):1865–1883, 2017.
[37] Moacir Ponti, Arthur A Chaves, F´
abio R Jorge, Gabriel BP Costa,
Adimara Colturato, and Kalinka RLJC Branco. Precision agriculture:
Using low-cost systems to acquire low-altitude images. IEEE computer
graphics and applications, 36(4):14–20, 2016.
[38] Weixun Zhou, Shawn Newsam, Congmin Li, and Zhenfeng Shao.
Patternnet: a benchmark dataset for performance evaluation of remote
sensing image retrieval. ISPRS Journal of Photogrammetry and Remote
Sensing, 2018.
[39] Weixun Zhou, Shawn Newsam, Congmin Li, and Zhenfeng Shao.
Patternnet: a benchmark dataset for performance evaluation of remote
sensing image retrieval. ISPRS Journal of Photogrammetry and Remote
Sensing, 2018.
[40] European Commission. Mapping guide for a European urban atlas. https:
//ec.europa.eu/regional policy/sources/tender/pdf/2012066/annexe2.pdf,
2012.