Earth Science Informatics (2024) 17:2495–2508

https://doi.org/10.1007/s12145-024-01301-x

RESEARCH

Satellite image classification using deep learning approach

Divakar Yadav1 · Kritarth Kapoor2 · Arun Kumar Yadav2 · Mohit Kumar2 · Arti Jain3 · Jorge Morato4

Received: 6 December 2023 / Accepted: 24 March 2024 / Published online: 3 April 2024
© The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2024

Abstract
Our planet Earth comprises distinguished topologies based on temperature, location, latitude, longitude, and altitude, which
can be captured using Remote Sensing Satellites. In this paper, the classification of satellite images is performed based on their
topologies and geographical features. Researchers have worked on several machine learning and deep learning methods like
support vector machine, k-nearest neighbor, maximum likelihood, deep belief network, etc. that can be used to solve satellite
image classification tasks. All strategies give promising results. Recent trends show that a Convolutional Neural Network
(CNN) is an excellent deep learning model for classification purposes, which is used in this paper. The open-source EuroSAT
dataset is used for classifying the remote images which contain 27,000 images distributed among ten classes. The 3 baseline
CNN models are pre-trained, namely- ResNet50, ResNet101, and GoogleNet models. They have other sequence layers added
to them with respect to CNN, and data is pre-processed using LAB channel operations. The highest accuracy of 99.68%,
precision of 99.42%, recall of 99.51%, and F- Score of 99.45% are achieved using GoogleNet over the pre-processed dataset.
The proposed work is compared with the state-of-art methods and it is observed that more layers in CNN do not necessarily
provide a better outcome for a medium-sized dataset. The GoogleNet, a 22-layer CNN, performs faster and better than the
50 layers CNN- ResNet50, and 101 layers CNN- ResNet101.

Keywords Convolutional neural networks · Deep learning · EuroSAT · GoogleNet · ResNet50 · RestNet101 · Satellite images

Introduction land used for crop production. The manual classification of

these different land areas using image interpretation tech-
On the planet Earth, only about 29% of its total surface is as niques (Di et al. 2018) requires more time and field experts.
land (Reality 2023)- continents and islands, while most of its This is because localization costs are too high, and data
surface is water (71%)- saltwater and freshwater bodies, with analysts are willing to do less research, so automation is
about 2% of the water being frozen- ice and glaciers. Among necessary. Hence, an efficient and automatic Satellite Image
the land types, habitable land is the one where one can live Classification technique is to be deployed which involves
and produce on, for example, agricultural land. The agricul- learning various vegetation types, for example, agriculture,
tural land comprises around 70% pastureland and 30% of forests, etc., and studying urban- residential, and commercial
arable land. The pasture/grazing/range land is also known as determining various land uses in an area (Abdul Azeem et al.
meadows used for livestock rearing and animal product pro- 2023; Tehsin et al. 2023).
duction, and the arable/cropland is also known as cultivated Satellite imagery serves as a global effort to map world-
wide communities, for example, OpenStreetMap (Zhou et al.
Communicated by: H. Babaie 2022), Google Earth (Zhang et al. 2022), and Earth Explorer
(Agocs et al. 2022). These are the platforms where maps
B Arti Jain
are digitized through high-resolution images. The digi-
[email protected]
tized maps are living documents where new elements are
1 SOCIS, IGNOU, New Delhi, India added by remote mappers, and satellite images are acces-
2 Department of CSE, NIT Hamirpur, Hamirpur, H.P., India sible to researchers around the world. The researchers have
3 worked with satellite image classification datasets, for exam-
Department of CSE and IT, JIIT, Noida, India
ple, Landsat (Landsat 2023), Sentinel-2 (Sentinel 2023),
4 Department of CSE, Universidad Carlos III de Madrid,
In-orbit (In-orbit 2023), RSI-CB256 (RSI-CB256 2023),
Madrid, Spain

123
2496 Earth Science Informatics (2024) 17:2495–2508

and many more. Satellite imaging has a wide range of depicts applications of the proposed work in varied domains.
usages, for example, cartography and navigation (Dymkova Finally, “Conclusion and future work” concludes the paper.
2020), ecological surveillance (Kimothi et al. 2022), disaster
responses (Antzoulatos et al. 2022) and regulation enforce-
ment (de Freitas Preto et al. 2022). These applications of Related work
satellite images can be implemented using higher accuracy
models. In the field of Satellite Image Classification (Adegun et al.
In this paper, an exclusive method of Satellite Image 2023), many researchers have important contributions such
Classification is detailed which is based on Convolutional as Support Vector Machine (SVM) and Maximum Like-
Neural Network (CNN) (Rai et al. 2020). The CNN archi- lihood (Mondal et al. 2012), Naïve Bayes (NB) (Mao
tecture is commonly used in computer vision applications et al. 2015), K-Nearest Neighbor (KNN) and Decision Tree
(Bhatt et al. 2021). The three baseline CNN models, (DT) and Random Forest (RF) (Hernandez et al. 2018).
namely- GoogleNet (GoogLeNet 2023), ResNet-50 (ResNet- These classical methods have predicted elegant labels and
50 2023), and ResNet101 (ResNet-101 2023) are chosen as effects against versioning of the aforementioned study.
pre-trained models where sequence layers are added, and Apart from these classical approaches, several deep learn-
hyper-parameter tuning is performed. A pre-trained model is ing methods (Glinka et al. 2023; Soufi and Belouadha
a previously trained model over a dataset and has weights as 2023) are applied for satellite image classification tasks, for
well as biases representing dataset features. The pre-trained example, AlexNet (Unnikrishnan et al. 2018), GoogleNet
model saves time since the learned features are transferable (Ostankovich and Afanasyev 2018), InceptionV3 (Boonyuen
to distinguished datasets. The pre-trained GoogleNet is also et al. 2019), DenseNet (Tong et al. 2020), Deep Belief Net-
known as InceptionV1 which yields a significant decrease in work (Horvath et al. 2020), VGGNet (Srinivas and Shakira
error rate as compared to other models such as AlexNet, ZF- 2021), and ResNet-18 (Zhao et al. 2022).
Net, and VGGNet. For deeper architecture, the GoogleNet Usman et al. (2013) has worked with the K-means clus-
uses 1×1 convolutions in the middle of the network, and tering approach to detect high-resolution satellite pictures.
global average pooling at the end of the architecture. The pre- The extracted areas are classified using a minimal distance
trained ResNet is a particular type of CNN which is known judgment. It is compared to the training samples, to establish
as a Residual Network that stacks residual blocks to formu- which class each region belongs to. Farmlands, barren lands,
late networks. It has fewer filters and is lesser complex than and built-up areas are used in this work. The precision of
VGGNet. The ResNet-50 is a 50-layer CNN having 48 con- classification is reported as 0.8889 and Kappa value as 0.835
volutions, 1 (3×3) max pool having 2-sized stride, and 1 respectively. Mahmon and Ya’acob (2014) have examined
average pooling layer. And, the ResNet-101 is a CNN archi- the ability of Artificial Neural Networks (ANNs) to classify
tecture that comprises 101 layers of deep convolutions. The satellite images using different approaches. Since the number
following research contributions are discussed in this paper. of data training sets depends on ANN models and the number
of classes, the classification takes a different amount of time.
The review article looks at different techniques for classify-
• To classify satellite images using baseline CNN mod- ing satellite images compared to ANN based on accuracy. It
els, i.e., GoogleNet, ResNet-50, and ResNet101 over discusses the best Land Use and Land Cover (LULC) clas-
EuroSAT (EuroSat 2023) dataset. sification techniques and compares them with respect to the
• To improve the image quality of satellite images using benchmark. The LULC map of the study area is useful for
Contrast Limited Adaptive Histogram Equalization various facets, including deforestation, weather forecasting,
(CLAHE). urban planning, damage assessment, archaeology, and more.
• To apply CNN models on the EuroSAT that too with Pandya and Priya (2015) have performed the collection of
and without preprocessing, and evaluate using standard satellite images for agricultural usage. Their primary goal is
metrics. to create a model for a vegetation classification system that
can accept satellite forecaster data from various sources, cate-
gorize vegetation regions using image processing techniques,
The rest of the research paper is structured as follows. and assess useful forecast policies for decision-makers.
“Related work” discusses related literature for satellite image Schmedtmann and Campagnolo (2015) have worked on 12
classification. “Proposed methodology” details the proposed different classes within different types of croplands such as
methodology based on CNN models for satellite imagery. permanent grassland, feed crop, maize, and so on that are
“Results and analysis” illustrates results and analysis while listed in each class. Through their efforts, they have built a
highlighting outcomes and findings for the chosen dataset simple automatic control system that has incorporated auto-
using evaluation metrics. “Applications of proposed work” matic image classification. The data is classified using an

123
Earth Science Informatics (2024) 17:2495–2508 2497

SVM classifier with 10-fold cross-validation. The statistics ship detection is presented. The YOLOv5 outperforms the
for classification accuracy is calculated at 80%. Jog and Dixit other YOLO algorithms having an accuracy of 99%.
(2016) have compared supervised classifiers such as mini- Table 1 summarizes the collection of research articles
mum distance, SVM, maximum likelihood, and parallelism. that are reviewed during the satellite image classification-
It is found that the maximum likelihood classifier is bet- based research work, covering Reference Number (Ref),
ter than the minimum and parallel distance classifiers for Dataset(s), Method(s), Total Images, Total Classes, and
kappa-differentiated image types. For all image types, except F-Score, along with their Advantages and Disadvantages
Landsat Multi-Spectral Scanner (MSS) images, the overall respectively.
accuracy is greater than 88%, and kappa statistics is more In the stated literature, it has been observed that there
significant than 0.82 for the maximum likelihood classifier. are various studies on machine learning and deep learning
Pritt and Chern (2017) have applied a deep learning models for the satellite image classification task. They have
ensemble of Convolutional Neural Network (CNN) systems used varied datasets and have achieved significant F-Score.
and other Neural Networks (NNs) for classifying satellite However, none of them have worked on the RGB images
image objects. The integration of satellite metadata and of the EuroSat dataset. In this research, we have worked on
image features is executed for the IARPA Functional Map of the EuroSAT dataset having RGB images. Also, we have
the World (fMoW) dataset which is categorized into sixty- applied deep learning-based baseline Convolutional Neu-
three classes. The system has achieved accuracy (0.83), ral Networks models, namely- ResNet50, ResNet101 and
F1-score (0.797), and 15 classes have accuracies (0.95) or GoogleNet respectively.
more. Sowmya et al. (2017) have discussed four process-
ing stages for satellite image classification. These stages
are- pre-processing, enhancement, transformation, and clas- Proposed methodology
sification. The pre-processing stage includes geometric,
radiometric, and atmospheric corrections. The enhancement The proposed methodology involves the acquisition of a
stage includes radiometric, spatial, spectral, and geometric standard dataset that undergoes- with pre-processing and
enhancement. The transformation stage includes band arith- without pre-processing, and after that pre-trained models
metic operations like manipulation of multiple bands of data such as ResNet-50, ResNet-101, and GoogleNet are used as
i.e., singular multi-spectral image or multi-temporal image the backbone of the model. Then add an additional sequential
to generate new images from different sources highlight- layer, comprising a linear one with 256 nodes and a ReLU
ing features or properties of interest. The classification stage activation function. The final layer is replaced with a fully
includes categorizing image pixels or remotely sensed satel- connected layer with an output number of classes. Finally,
lite data in rder to achieve class labels or land cover themes. The trained model is evaluated on a validation set, and its
Neware and Khan (2018) have stated that agriculture is rec- performance is measured using metrics such as accuracy,
ognized in satellite images using supervised classification precision, recall, and F-measure. A detailed description of
approaches. They have discussed 4 classes, namely- vege- the proposed methodology is given in Fig. 1.
tation, non-vegetation, water lands, and urban regions. The
primary purpose of their study is to identify agriculture using Dataset acquisition
minimum distance, spectral angle mapping, parallelepiped,
and land cover signature for categorization. The best results EuroSat Dataset is an open-source, remote satellite dataset
are obtained using the Normalized Difference Vegetation that is freely available at Kaggle (EuroSat 2023). It con-
Index (NDVI) technique to locate the vegetation. sists of two sets of images, namely- RGB images and Band
Syrris et al. (2020) have created SatImNet. The SatImNet images from the satellite, known as Sentinel-2, where every
is a collection of open, trained, structured, and harmonized image comprises 64 x 64 pixels whose ground sampling dis-
data based on predetermined rules. Furthermore, CNN has tance is 10 meters. The RGB images that are collected from
classified satellite pictures. Varma et al. (2021) have explored the Sentinel dataset are known as EuroSAT, while all bands
supervised classification techniques for SIC tasks, namely- of the spectrum that are collected from the Sentinel dataset
minimum distance, parallelepiped, KNN, and maximum is known as EuroSATallBands. The multi-spectral contains
likelihood. They have analyzed these techniques and a com- thirteen spectral bands, along with visible, close to infrared,
parative analysis of their effectiveness is drawn. Patel et al. and short-wave infrared with 10 particular instructions.
(2022) have worked on deep learning methods for automatic In this research, the EuroSAT dataset is chosen, i.e., RGB
ship detection over the Airbus ship dataset containing 40k satellite images due to its better feasibility and lesser space
satellite images. A comparative analysis of different versions occupancy Fig. 2. The EuroSAT dataset comprises Joint
of YOLO algorithms- YOLOv3, YOLOv4, and YOLOv5 for Photographic Expert Group (JPG or JPEG) images from

123
 Table 1 Satellite images classification methods
2498

Ref Data-set(s) Method(s) Total Image Tot-al Cla-sses F-Score (%) Advantages Disadvantages

123
Negrel et al. (2014) UC Merced Land Use HOG+RGB, Fisher Vec- 2100 21 93.80 . Higher number of classes, . Lesser data, model not used
tor, Vectors of Locally applications, and coverage extensively
Aggregated Descriptors increase
(VLAD) and Vectors of
Locally Aggregated Tensors
(VLAT)
. Higher accuracy . Slightly outdated technique
. When compared to BoW, . Confusion matrix isn’t
these methods lead to large shared- precision, recall, and
performance improvement F-measure can’t be com-
puted
Lv et al. (2015) RADARSAT -2 PolSAR Deep Belief Network (DBN) 1660 10 81.52 . Higher accuracy (81.74) . Lesser training dataset
over SVM (76.79), NN gives lesser accuracy
(74.37), SEM (72.43)
. Hidden layers (02), LULC
mapping
Basu et al. (2015) SAT-4 SAT-6 (high- DBN with Restricted Boltz- 500,000 4 97.95 . Model trained and tested . Lower number of classes
resolution satellite mann Machines and Feature on two different datasets and for both SAT-4 and SAT-6,
multispectral datasets) extraction using Enhanced yielded high accuracy for lesser coverage and limited
Vegetation Index (EVI), both SAT 4 and SAT 6 application
Normalized Difference Veg-
etation Index (NDVI), and
Atmospherically Resistant
Vegetation Index (ARVI)
405,000 6 93.92 . High number of datasets . Performance depends on
images, model is trained the selected features
really well on both datasets
. Confusion Matrix is not
shared- precision, recall, and
F-measures can’t be com-
puted
Jog and Dixit (2016) Landsat 5 TM SVM and Maximum Likeli- - 10 96.59 . Maximum Likelihood . Maximum Likelihood is
hood yields higher accuracy slow, computational com-
plexity
. SVM minimizes no. of mis- . SVM depends on the selec-
classifications, independent tion of kernel and size of data
of feature dimensionality
Berriel et al. (2017) Google Static Maps Visual Geometry Group 245,768 2 95.17 . Very high amount of data, . Only two classes, very spe-
(Crosswalks-p: positive, Deep Learning Model model can be trained really cific applications
zebra crossing: negative) well
Earth Science Informatics (2024) 17:2495–2508
 Table 1 continued
Ref Data-set(s) Method(s) Total Image Tot-al Cla-sses F-Score (%) Advantages Disadvantages

. Very high accuracy . Dataset not available,

manually downloaded using
Google Static Maps API
Pritt and Chern (2017) Functional Map of the World Ensemble- DenseNet-16 53000 63 79.70 . Larger amount of data and . Some classes have rela-
ResNet-152 Inception-v3 larger number of classes, tively lower accuracy (81%)
Xception increase the model function-
ality
. 15 classes have accuracy . Most classes have accuracy
greater than 95% < 70%, unsuitable for all
classes
Helber et al. (2019) UCM RSSCN SIRI-WHU ResNet-50 2100 21 97.77 . Fine-tune CNN hyperpa- . Trouble with fewer avail-
Earth Science Informatics (2024) 17:2495–2508

Corel-1K Corel-1.5K rameters using SGDM able training samples

2800 07 92.76 . Classifies images in an
effective and efficient man-
ner
2400 12 94.02
1000 10 96.99
1500 15 99.66
Miranda et al. (2019) Sentinel-2 Custom CNN using kernel - 3 97.70 . Uses image enhancement . Lower no. of classes, cov-
size=2, ReLU and Normal- techniques and metrics- erage is lesser
ized Difference Vegetation Normalized Difference Veg-
Index for 5,000 epochs etation Index, Brightness,
GLCM homogeneity, and
Rectangular Fit to improve
accuracy
. Uses 5 bands instead of . Uses 5 bands instead
3(RGB), due to extra band of 3(RGB), computationally
8(Near-Infrared) and band more expensive than RGB
11(SWIR), more accurate
than RGB.
. Size of the dataset is not dis-
closed, hard to reproduce the
steps mentioned in the paper
Sasidhar et al. (2019) SAT-4 Custom CNN, 2 convo- 500,000 4 98.75 . Excellent accuracy . Uses 4 Bands, computa-
lutional 2D, max-pooling, tionally expensive than RGB
dropout, dense, Normalized
Difference Vegetation Index
(NDVI)

123
2499
 Table 1 continued
2500

Ref Data-set(s) Method(s) Total Image Tot-al Cla-sses F-Score (%) Advantages Disadvantages

123
SAT-6 405,000 6 99.25 . Higher number of images, . Lower number of classes,
model is trained really well lesser coverage, and limited
application
. Custom CNN, better flex- . Overfitting due to very high
ibility and easier to modify accuracy
the existing model
Song et al. (2019) WorldView-3 MSI Custom 1D CNN, convo- 50,000 13 89.18 . Large amount of data and . Uses 36 bands, computa-
lutional, kernel 3x1, batch number of classes, increases tionally expensive
normalization, ReLU, max- model functionality
pooling, softmax
. 1D CNN with additional . Exact dataset is hard to
information boosts the clas- obtain, and difficult to repro-
sifier’s performance duce results
. Confusion Matrix not
shared- precision, recall,
and F-measures can’t be
computed
Akshay et al. (2020) EuroSAT (Sentinel-2A) Custom CNN, convolu- 1000 10 88.86 . Multiple classification and . Lesser amount of data, the
tional, ReLU, pooling, extensive use model isn’t suitable for any
flatten- image preprocess- different dataset
ing: grayscale and nose
filtering, feature extraction-
local binary pattern
. 64x64 image patches don’t . Not fully optimized, can
allow finely graduated per- be improved using more
pixel segmentation, detect datasets, and other pre-
changes and keep maps up- processing techniques
to-date
. Good accuracy for a smaller
dataset
Tun et al. (2020) UC Merced Land Use Custom 5 layer CNN using 2100 21 70.49 . Higher number of classes, . Low validation accuracy
Adam Optimizer, ReLU applications, and coverage (70%)
activation, and Softmax increase
classifier for 50 epochs
. Using custom CNN, . 100 images per class, not
increase in no. of con- adequate for proper training
volutional and tuning of model for each class
hyperparameters, accuracy
highly improved
Earth Science Informatics (2024) 17:2495–2508
 Table 1 continued
Ref Data-set(s) Method(s) Total Image Tot-al Cla-sses F-Score (%) Advantages Disadvantages

Liu et al. (2020) SAT-4 SAT-6 (high- Custom CNN, 2 convo- 500,000 4 99.90 . Model trained and tested . Lower number of classes
resolution satellite lutions, 32 and 64 fea- on 2 different datasets, high for both datasets, lesser cov-
multispectral datasets) ture maps, 3×3 ReLU, 2x2 accuracy erage, and limited applica-
max-pooling, dropout, Soft- tion
max with cross-entropy and
Adadelta optimizer
405.000 6 99.84 . Higher no. of images, . Confusion matrix is not
model trained well on both shared- precision, recall, and
datasets F-measures can’t be com-
puted
. Custom CNN, better flex-
Earth Science Informatics (2024) 17:2495–2508

ibility and easier to modify

model
Raiyani et al. (2021) Sentinel-2 and Hollstein Custom CNN, convolutional 6,628,4 6 84.65 . High number of points in . Lower no. of classes, lesser
pixel-level 1D, dropout, pooling, flat- the dataset coverage
ten, dense, and softmax acti-
vation
78 points . CNN performs better than . Relatively lower accuracy
Random Forest, Extra Trees,
and Sen2Cor models
. Model not fully optimized,
scope of improvement
Shabbir et al. (2021) SIRI-WHU ResNet-50 CNN 2400 12 94.02 . Hyper-parameters finely . Lesser data, model not used
tuned using SGD with extensively
Momentum
. Higher Accuracy . Optimal datasets like UC
Merced Land Use are not
used

123
2501
2502 Earth Science Informatics (2024) 17:2495–2508

Fig. 1 Architecture of proposed

methodology With
Preprocessing

Model
GoogleNet Sequence Training and
Dataset ResNet-50
Layer Evaluation
ResNet-101

Without
Preprocessing

the Sentinel-2 satellite, having only 3 bands of images- multiple channels (“l”, “a”, and “b”). CLAHE is applied to
Red, Green, and Blue. It consists of 27,000 labeled and the “l”-channel. Then, the “a” and “b” channels are com-
geo-referenced images. As far as the distribution is con- bined with the CLAHE-enhanced “l”-channel. Finally, the
sidered, classes- “AnnualCrop”, “Forest”, “HerbaceousVeg- image is converted from LAB to RGB color space which
etation”, “Residential”, and “SeaLake” each have 3000 greatly improves the performance. The choice to incorpo-
images, whereas classes- “Highway”, “Industrial”, “Per- rate the LAB color model in the pre-processing stage stems
manentCrop”, and “River” each have 2500 images, and from its well-established advantages in accuracy when com-
“Pasture” class has 2000 images respectively. Thus, there pared to traditional color models such as RGB or CMYK.
is a class imbalance among three types of classes having The LAB color space is specifically designed to more closely
3000, 2500, and 2000 images while classes having the same align with human perception, making it particularly suitable
number of images are balanced. The training, validation, and for image processing tasks. For instance, in medical imaging,
testing comprise 912.5kB, 257.43kB, and 128.15kB Comma where subtle variations in color may indicate critical informa-
Separated Value (.csv) files respectively. tion, the LAB color model’s perceptual uniformity ensures
that the model captures and interprets these nuances more
Pre-processing effectively. This can lead to enhanced diagnostic capabilities
and improved accuracy in tasks like identifying anomalies in
The dataset is preprocessed to ensure that it is in the correct medical scans. By leveraging the LAB color model during
format for training the deep learning model. This includes the pre-processing stage, our methodology extends beyond a
using the LAB color model and dividing the LAB image into one-size-fits-all approach, highlighting its adaptability.

Fig. 2 Distribution of EuroSAT

dataset AnnualCrop

Forest

EuroSAT HerbaceousVegetation

Residential

SeaLake
EuroSAT Dataset
Highway

Industrial

EuroSATallBands PermanentCrop

3000 Images River

2500 Images

2000 Images Pasture

123
Earth Science Informatics (2024) 17:2495–2508 2503

Input Image Conv + ReLU Pooling

Fully Connected
Softmax
+ ReLU

Labels

Fig. 3 Architecture of CNN

Model architecture • Pooling Layer: The primary purpose of this layer is to

reduce the size of the convolution feature map, in order
A deep learning model is chosen for this study due to its
to lower computational costs. This is accomplished by
ability to learn complex, non-linear relationships in image
decreasing layer connections and running each feature
data. The pre-trained models such as ResNet-50, ResNet-
map separately. In max pooling, the largest element is
101, and Googlenet are used as the backbone of the model.
derived from the feature map. Average pooling calcu-
The final layer is replaced with a fully connected layer with
lates an average of the components in a pre-defined sized
the output number of classes.
image segment. The entire sum of components in the
predefined section is calculated using sum pooling. The
Convolutional neural network pooling layer is commonly used to link the convolutional
and fully connected layers.
Convolutional Neural Network (CNN) carries a series of pro-
• Fully Connected Layer: The fully connected (FC) layer
cessing layers as is shown in Fig. 3.
connects neurons between two layers, and is made up of
Every layer has its own family of convolution filters
weights, biases, and neurons. In this layer, input images
that accumulate picture functions. The CNN can be mainly
from previous layers are flattened and sent to FC. The
divided into the following layers:
flattened vector is then transmitted via a few more FC
layers, where the mathematical functional operations are
• Convolutional Layer: This is the first layer, which extracts usually done.
various features from the input images. This layer per- • Dropout: The dropout layer is used to avoid overfitting by
forms the convolution mathematical process between an removing a few neurons from the neural network during
input image and a filter of a given size, say MxM. Sliding the training phase, resulting in a smaller model.
the filter across the input image produces the dot product • Activation Functions: At the network’s end, it determines
between the filter and parts of the input image in terms which model information should fire in the forward direc-
of filter size (MxM). The output of this layer is a fea- tion and which should not. Some of the most commonly
ture map, which contains information about the image’s used activation functions include ReLU, Softmax, Tanh,
corners and edges. and Sigmoid.
Fully Connected
Zero Padding

INPUT OUTPUT
Conv Block

Conv Block

Conv Block

Conv Block
BatchNorm

Flattening
Avg Pool
MaxPool

ID Block

ID Block

ID Block

ID Block
CONV

ReLU

Stage 1 Stage 2 Stage 3 Stage 4 Stage 5

Fig. 4 Architecture of ResNet50

123
2504 Earth Science Informatics (2024) 17:2495–2508

As mentioned earlier, 3 different pre-trained models, ResNet101

namely- ResNet50, ResNet101, and GoogleNet are consid-
ered as baseline models, and then add an additional sequential ResNet-101 is a deep residual neural network that is devel-
layer in the CNN. The following are the different models: oped by Microsoft in 2015. It is a 101-layer neural network
that is trained on more than a million images and won the Ima-
ResNet50 geNet Large Scale Visual Recognition Challenge (ILSVRC)
in 2015. The architecture of ResNet-101 is similar to ResNet-
The architecture of ResNet-50 can be broken down into four 50 which can be broken down into four main parts: a stem,
main parts: a stem, four residual blocks, a global average four residual blocks, a global average pooling layer, and a
pooling layer, and a fully connected layer as is shown in fully connected layer.
Fig. 4. The main difference between ResNet-101 and ResNet-50
is that ResNet-101 has more layers in the residual blocks,
• Stem: The stem is a series of convolutional layers that which leads to a deeper network. The deeper architecture
reduce the spatial resolution of the input image while allows us to learn more complex features and improves the
increasing the number of feature maps. The stem starts performance of the classification task.
with a convolutional layer with 7x7 filters, a stride of 2,
and padding of 3, followed by a 3x3 max-pooling layer GoogleNet
with a stride of 2, and two more convolutional layers with
1x1 filters. GoogleNet architecture is quite complex and can be bro-
• Four Residual Blocks: The four residual blocks are com- ken down into several smaller sub-networks, called Inception
posed of several residual units, each unit contains two modules as is shown in Fig. 5. The model’s input dimensions
3x3 convolutional layers with a stride of 1 and the same are 224 x 224 x 3, and the output dimensions are 112 x 112
number of filters. All units have a shortcut connection x 64.
that skips one or more layers. The GoogleNet has gone through a 7 x 7 convolutional
• Global Average Pooling: The global average pooling layer with 64 filters, a stride of 2, and padding of 3. After
layer reduces the spatial resolution of feature maps to that, the model applies a max pooling of 3 x 3 with a stride
1x1 while preserving the number of feature maps. of 2 which leads to an output size of 56 x 56 x 64. Then a
• Fully Connected Layer: The fully connected layer has 3 x 3 convolutional layer with 192 filters, stride of 1, and
1000 outputs and a softmax function is applied to the padding of 1 is applied and the output size becomes 56 x 56
output in order to obtain the probability of each class. x 192. After that, a 3 x 3 max pooling layer with a stride

Input Conv1 Conv2 Inception1 Inception3 Inception8 Output

MaxPool MaxPool Inception2 Inception4 Inception9

MaxPool Inception5 MaxPool

Inception6 FC

Inception7 Softmax

MaxPool

Fig. 5 Architecture of GoogleNet

123
Earth Science Informatics (2024) 17:2495–2508 2505

Table 2 Hyperparameters used in the proposed work ages for each model, with and without pre-processing, under
S. No. Item Detail different training/testing data split ratios. The table includes
various training/testing data split ratios, expressed as a per-
1 Programming Language Python 3.9 centage such as 10% for training and 90% for testing, 20% for
2 Frameworks Tensorflow, Keras training and 80% for testing, and so on. It may be observed
3 Cloud Environment Google Colab that as the proportion of data allocated for training increases,
4 RAM 12 Gb the accuracy of the models also increases. This is expected,
5 Graphics 8 Gb (Nvidia K80/T4) as more training data typically leads to better model per-
6 Disk Space 25 Gb formance. GoogleNet consistently achieves high accuracy
7 Batch Size 32 across various split ratios, often outperforming ResNet-50
8 Learning Rate 0.0001 and ResNet-101. The highest accuracy of 99.68%, precision
of 99.42%, recall of 99.51%, and F- Score of 99.45% are
achieved with the 90/10 split of the data. The results sug-
gest that GoogleNet is a robust model across different data
of 2 is applied and the output size becomes 28 x 28 x 192.
split ratios for the satellite image classification task, while
Then, 4 Inception modules are applied, named 3a, 3b, 4a, 4b,
ResNet-101 also performs well.
4c, 4d, and 4e each one gives a different output size. Each
Table 4 details the standard evaluation metrics, i.e.,
Inception module is composed of multiple branches, each
accuracy, precision, recall, and F-Score for the ResNet-50,
branch has different filter sizes and output sizes and finally
RestNet101, and GoogleNet using the EuroSAT dataset with
are concatenated together. After the Inception modules, a 3
and without preprocessing for the highest performance cases.
x 3 max pooling is applied again with a stride of 2 and the
Table 5 compares the proposed satellite image classifica-
output size becomes 7 x 7 x 832. Then, 2 further Inception
tion task with the state-of-art methods. Table 5 compares the
modules- 5a and 5b are applied, followed by average pooling
performance of various methods, including state-of-the-art
with an output size of 1 x 1 x 1024, dropout with a drop rate
approaches and the proposed methodology, in satellite image
of 0.4, and a linear layer with 1000 outputs, resulting in an
classification. The table includes references to different stud-
output size of 1 x 1 x 1000. Finally, a softmax function is
ies, the datasets used, and the methods applied for satellite
applied to the output in order to obtain the probability of each
image classification. It may be observed from the table that
class. The relevant hyperparameters used in the experiments
the proposed methodology achieves high accuracy (99.68%),
are depicted in Table 2.
precision (99.42%), recall (99.51%), and F-score (99.45%)
and consistently outperforms state-of-the-art methods. The
proposed method’s consistent high performance suggests its
Results and analysis efficacy for this particular task.

This section details rigorous experiments and evaluated

results that have been undergone during the execution of the
Applications of proposed work
deep learning-based satellite image classification process.
The k-fold cross-validation technique is applied over
The proposed work has several key industrial applications.
the EuroSAT dataset (EuroSat 2023). Hence, the dataset is
Some of them are highlighted here.
subdivided into different training and test ratios, namely-
10/90, 20/80, 30/70, 40/60, 50/50, 60/40, 70/30, 80/20,
and 90/10 respectively. Table 3 presents the results of K- • With admission to such information, packages within the
Fold Cross-Validation on three different models (ResNet-50, domains of agriculture, catastrophe recuperation, climate
ResNet-101, and GoogleNet) along with pre-processing over exchange, city development, or environmental monitor-
the EuroSAT dataset. The table provides accuracy percent- ing may be found.

Table 3 K-Fold cross-validation

Approach 10/90 20/80 30/70 40/60 50/50 60/40 70/30 80/20 90/10
on ResNet50, RestNet101,
GoogleNet over EuroSAT ResNet-50 85.88 89.84 91.29 92.37 94.48 95.77 97.24 97.96 98.71
ResNet-101 85.96 89.53 93.75 95.01 95.45 96.26 97.32 98.37 99.01
GoogleNet 87.37 90.97 93.97 95.62 96.96 97.54 98.70 99.42 99.68

123
2506 Earth Science Informatics (2024) 17:2495–2508

Table 4 Evaluation metrics of

Approach Dataset Type Accuracy (%) Precision (%) Recall (%) F-Score (%)
deep learning models (without
preprocessing vs. with ResNet-50 Without Preprocessing 98.49 98.57 98.44 98.50
preprocessing)
ResNet-50 With Preprocessing 98.71 98.72 98.61 98.65
ResNet-101 Without Preprocessing 98.79 98.87 98.74 98.80
ResNet-101 With Preprocessing 99.01 99.02 98.91 98.93
GoogleNet Without Preprocessing 99.29 98.97 99.24 99.12
GoogleNet With Preprocessing 99.68 99.42 99.51 99.45

• The programs of image classification and image graph and 99.68% respectively are achieved. The proposed method
evaluation in remote sensing are important as they are performs better than the state-of-the methods thus achiev-
used in diverse implemented domains along with navy ing one of the highest accuracies for the EuroSAT dataset
and civil fields. classification. The results are analyzed for both- without pre-
• The computational classification of agricultural land is processing and with pre-processing, The GoogleNet model
based on the type of crops, e.g. Annual Crops (corn, peas), being the smallest among all the models (comprising 22
Pasture (land used for grazing), and Herbaceous Vegeta- layers) is able to achieve the highest evaluation metrics-
tion (plants that don’t have wooden stems). precision (99.42%), recall (99.51%), and F-Score (99.45%)
proving that large deep learning models don’t necessarily
imply better performance. In the future, the proposed dataset
may be used in lots of actual-global viewing applications-
Conclusion and future work
capacity land use and land cover alternate or collect geo-
graphical development.
In this research, the satellite image classification task is
performed using 3 different pre-trained CNN models over Author Contributions Kritarth Kapoor and Divakar Yadav have written
the EuroSAT dataset. The dataset is open-source containing the initial manuscript and performed experimentations. Arti Jain has
27,000 labeled and geo-referenced images from Sentinel- edited the manuscript and performed further experimentations. Arun
Kumar Yadav and Mohit Kumar have worked on Tables and Figures
2 that are distributed among 10 classes. Upon training the throughout. Jorge Morato has critically analyzed and upgraded the
EuroSAT dataset on 3 different models that are ResNet-50, manuscript.
ResNet-101, and GoogleNet accuracy of 98.71%, 99.01%,

Table 5 Comparison of the proposed methodology with recent state-of-art methods. Here, Acc. stands for Accuracy, Pr. stands for Precision, Rec.
stands for Recall, F1 stands for F-Score, Ref. stands for Reference
Ref. Dataset Method Acc. Pr. Rec. F1

Lv et al. (2015) RADARSA T-2 PolSAR Deep Belief Network 81.74 81.56 81.60 81.52
Jog and Dixit (2016) Landsat 5 TM SVM, Maximum Likelihood 97.40 97.00 96.20 96.59
Berriel et al. (2017) Google Static Maps Visual Geometry Group Deep 97.11 96.68 94.77 95.17
Learning Model
Pritt and Chern (2017) Functional Map of the World Ensemble- DenseNet-16, ResNet- 81.00 81.20 78.07 79.70
152, Inception-v3, Xception
Miranda et al. (2019) Sentinel-2 CNN (Kernel Size=2, ReLu, Nor- 97.66 97.74 97.66 97.70
malized Difference Vegetation
Index for 5,000 Epochs)
Akshay et al. (2020) EuroSAT (Sentinel-2A) Custom CNN, Preprocessing: 89.00 88.78 88.94 88.86
Grayscale, Noise Filtering; Feature
Extraction: Local Binary Pattern
Tun et al. (2020) UC Merced Land Use Custom 5 layer CNN (Adam Opti- 70.00 71.00 70.00 70.49
mizer, ReLU, Softmax, 50 Epochs)
Raiyani et al. (2021) Sentinel-2 and Hollstein pixel-level Custom CNN (Convolutional 1D, 84.00 85.10 84.20 84.65
Dropout, Pooling, Flatten, Dense,
and Softmax)
Shabbir et al. (2021) SIRI-WHU ResNet-50 CNN 94.03 94.03 94.19 94.02
Proposed EuroSAT GoogleNet with preprocessing 99.68 99.42 99.51 99.45

123
Earth Science Informatics (2024) 17:2495–2508 2507

Funding declaration None. M et al (2022) Flood hazard and risk mapping by applying an
explainable machine learning framework using satellite imagery
Data Availability No datasets were generated or analysed during the and gis data. Sustainability 14(6):3251
current study. de Freitas Preto M, Garcia AS, Nakai ÉS, Casarin LP, Vilela VMdFN,
Ballester MVR (2022) The role of environmental legislation and
land use patterns on riparian deforestation dynamics in an amazo-
Declarations
nian agricultural frontier (mt, brazil). Land Use Policy 118:106132
Rai AK, Mandal N, Singh A, Singh KK (2020) Landsat 8 oli satellite
image classification using convolutional neural network. Procedia
Competing interests The authors have no relevant financial or non- Comput Sci 167:987–993
financial interests to disclose. Bhatt D, Patel C, Talsania H, Patel J, Vaghela R, Pandya S, Modi
K, Ghayvat H (2021) Cnn variants for computer vision: history,
Ethical approval Not applicable. architecture, application, challenges and future scope. Electronics
10(20):2470
GoogLeNet Understanding googlenet model - cnn architecture - geeks-
forgeeks. Accessed on Jan 2023. https://www.geeksforgeeks
References ResNet-50 Resnet-50: The basics and a quick tutorial. Accessed on Jan
2023. https://datagen.tech/guides/computer-vision/resnet-50/
Reality BM (2023) How much land is there on earth, & what ResNet-101 Resnet-101 from kaggle. Accessed on Jan 2023. https://
is it used for? Accessed Jan 2023 https://bettermeetsreality. www.kaggle.com/datasets/pytorch/resnet101
com/how-much-land-is-there-on-earth-what-is-it-used-for/#: EuroSat Eurosat dataset from kaggle. Accessed on Jan 2023.
~:text=Of%20the%20land%E2%80%99s%20total%20surface https://www.kaggle.com/datasets/1f15fbfaff64d4fba50c4c333c
%2C%20about%2029%25%20of,water%20there%20is%20on 8b07831e0deff5d980d214343d4fc4842758b4
%20earth%20in%20this%20guide Adegun AA, Viriri S, Tapamo J-R (2023) Review of deep learning meth-
Di H, Wang Z, AlRegib G (2018) Real-time seismic-image inter- ods for remote sensing satellite images classification: experimental
pretation via deconvolutional neural network. In: 2018 SEG survey and comparative analysis. J Big Data 10(1):93
international exposition and annual meeting. OnePetro Mondal A, Kundu S, Chandniha SK, Shukla R, Mishra P (2012)
Abdul Azeem N, Sharma S, Hasija S (2023) Classification of satellite Comparison of support vector machine and maximum likelihood
images using an ensembling approach based on deep learning”. classification technique using satellite imagery. Int J Remote Sens
Arab J Sci Eng 1–16 GIS 1(2):116–123
Tehsin S, Kausar S, Jameel A, Humayun M, Almofarreh DK (2023) Mao Y-M, Zhang M-S, Wang G-L, Sun P-P (2015) Landslide hazards
Satellite image categorization using scalable deep learning. Appl mapping using uncertain naïve bayesian classification method. J
Sci 13(8):5108 Central South Univ 22(9):3512–3520
Zhou Q, Wang S, Liu Y (2022) Exploring the accuracy and completeness Hernandez R, Elias I, Wenzhong S (2018) A random forests classifica-
patterns of global land-cover/land-use data in openstreetmap. Appl tion method for urban land-use mapping integrating spatial metrics
Geogr 145:102742 and texture analysis. Int J Remote Sens 39(4):1175–1198
Zhang C, Dong J, Xie Y, Zhang X, Ge Q (2022) Mapping irrigated Glinka S, Bajer J, Wierzbicki D, Karwowska K, Kedzierski M (2023)
croplands in china using a synergetic training sample generating The use of deep learning methods for object height estimation in
method, machine learning classifier, and google earth engine. Int high resolution satellite images. Sensors 23(19):8162
J Appl Earth Obs Geoinf 112:102888 Soufi O, Belouadha F-Z (2023) Deep learning technique for image satel-
Agócs T, Vanin F, Laberinti P, Oetjen H, Serlenga D, Sole MP, Salenc lite processing. Intell Methods Eng Sci 2(1):27–34
C, Lamarre D, Kaspers M, Rodrigues G et al (2022) Far-infrared Unnikrishnan A, Sowmya V, Soman K (2018) Deep alexnet with
outgoing radiation understanding and monitoring (forum)–system reduced number of trainable parameters for satellite image classi-
overview and key technology developments of esa’s 9 th earth fication. Procedia Comput Sci 143:931–938
explorer. In: IGARSS 2022-2022 IEEE international geoscience Ostankovich V, Afanasyev I (2018) Illegal buildings detection from
and remote sensing symposium. IEEE, pp 7186–7189 satellite images using googlenet and cadastral map. In: 2018 inter-
Landsat Landsat data freely available from the u.s. geological survey. national conference on Intelligent Systems (IS). IEEE, pp 616–623
Accessed on Jan 2023. https://landsat.gsfc.nasa.gov/data/ Boonyuen K, Kaewprapha P, Weesakul U, Srivihok P (2019) Convolu-
Sentinel Sentinel-2 datasets in earth engine from google develop- tional neural network inception-v3: a machine learning approach
ers. Accessed on Jan 2023. https://developers.google.com/earth- for leveling short-range rainfall forecast model from satellite
engine/datasets/catalog/sentinel-2/ image. In: Advances in Swarm Intelligence: 10th International
In-orbit In-orbit satellite image datasets from kaggle. Accessed on Conference, ICSI 2019, Chiang Mai, Thailand, July 26–30, 2019,
Jan 2023. https://www.kaggle.com/datasets/benguthrie/inorbit- Proceedings, Part II 10. Springer, pp 105–115
satellite-image-datasets Tong Z, Li Y, Li Y, Fan K, Si Y, He L (2020) New network based
RSI-CB256 Rsi-cb256 satellite image classification from kag- on unet++ and densenet for building extraction from high resolu-
gle. Accessed on Jan 2023. https://www.kaggle.com/datasets/ tion satellite imagery. In: IGARSS 2020-2020 IEEE international
mahmoudreda55/satellite-image-classification geoscience and remote sensing symposium. IEEE, pp 2268–2271
Dymkova S (2020) Conjunction and synchronization methods of earth Horváth J, Montserrat DM, Hao H, Delp EJ (2020) Manipulation detec-
satellite images with local cartographic data. In: Systems of signals tion in satellite images using deep belief networks. In: Proceedings
generating and processing in the field of on board communications. of the IEEE/CVF conference on computer vision and pattern recog-
IEEE, pp 1–7 nition workshops. pp 664–665
Kimothi S, Singh R, Gehlot A, Akram SV, Malik PK, Gupta A, Bilandi N Srinivas M, Shakira DV (2021) A novel satellite image segmentation
(2022) Intelligent energy and ecosystem for real-time monitoring using vggnet. Palarchs J Archaeol Egypt/Egyptol 18(4):1596–
of glaciers. Comput Electr Eng 102:108163 1607
Antzoulatos G, Kouloglou I-O, Bakratsas M, Moumtzidou A, Gialam-
poukidis I, Karakostas A, Lombardo F, Fiorin R, Norbiato D, Ferri

123
2508 Earth Science Informatics (2024) 17:2495–2508

Zhao Y, Zhang X, Feng W, Xu J (2022) Deep learning classification Berriel RF et al (2017) Deep learning-based large-scale automatic
by resnet-18 based on the real spectral dataset from multispectral satellite crosswalk classification. IEEE Geosci Remote Sens Lett
remote sensing images. Remote Sens 14(19):4883 14(9):1513–1517
Usman B et al (2013) Satellite imagery land cover classification using Helber P, Bischke B, Dengel A, Borth D (2019) Eurosat: A novel
k-means clustering algorithm computer vision for environmental dataset and deep learning benchmark for land use and land cover
information extraction. Elixir Int J Comput Sci Eng 63:18671– classification. IEEE J Sel Top Appl Earth Obs Remote Sens
18675 12(7):2217–2226
Mahmon NA, and Ya’acob N (2014) A review on classification of satel- Miranda E et al (2019) Forest classification method based on convolu-
lite image using artificial neural network (ann). In: 2014 IEEE 5th tional neural networks and sentinel-2 satellite imagery. Int J Fuzz
control and system graduate research colloquium. IEEE, pp 153– Logic Intell Syst 19(4):272–282
157 Sasidhar TT, Sreelakshmi K, Vyshnav M, Sowmya V, Soman K (2019)
Pandya A, Priya RS (2015) Classification of vegetation area from satel- Land cover satellite image classification using ndvi and simplecnn.
lite images using image processing techniques. Int J Res IT Manag In: 2019 10th International Conference on Computing, Commu-
Eng 5(3):2249–1619 nication and Networking Technologies (ICCCNT). IEEE, pp 1–5
Schmedtmann J, Campagnolo ML (2015) Reliable crop identification Song Y, Zhang Z, Baghbaderani RK, Wang F, Qu Y, Stuttsy C, Qi
with satellite imagery in the context of common agriculture policy H (2019) Land cover classification for satellite images through 1d
subsidy control. Remote Sens 7(7):9325–9346 cnn. In: 2019 10th Workshop on Hyperspectral Imaging and Signal
Jog S and Dixit M (2016) Supervised classification of satellite images. Processing: Evolution in Remote Sensing (WHISPERS). IEEE, pp
In: 2016 Conference on Advances in Signal Processing (CASP). 1–5
IEEE, pp 93–98 Akshay S, Mytravarun T, Manohar N, Pranav M (2020) Satellite image
Pritt M and Chern G (2017) Satellite image classification with deep classification for detecting unused landscape using cnn. In: 2020
learning. In: 2017 IEEE applied imagery pattern recognition work- international conference on electronics and sustainable communi-
shop (AIPR). IEEE, pp 1–7 cation systems (ICESC). IEEE, pp 215–222
Sowmya D, Deepa Shenoy P, Venugopal K (2017) Remote sensing Tun NL, Gavrilov A, Tun NM (2020) Multi-classification of satellite
satellite image processing techniques for image classification: a imagery using fully convolutional neural network. In: 2020 Inter-
comprehensive survey. Int J Comput Appl 161(11):24–37 national Conference on Industrial Engineering, Applications and
Neware R, Khan A (2018) Identification of agriculture areas in satellite Manufacturing (ICIEAM). IEEE, pp 1–5
images using supervised classification technique. J Creat Behav Liu Q, Basu S, Ganguly S et al (2020) Deepsat v2: feature augmented
6:682–688 convolutional neural nets for satellite image classification. Remote
Syrris V, Pesek O, Soille P (2020) Satimnet: Structured and harmonised Sens Lett 11(2):156–165
training data for enhanced satellite imagery classification. Remote Raiyani K, Gonçalves T, Rato L, Salgueiro P et al (2021) Sentinel-2
Sens 12(20):3358 image scene classification: a comparison between sen2cor and a
Varma MKS, Raja K, Rao NK (2021) Analysis of satellite images machine learning approach. Remote Sens 13(2):300
using supervised image classification methods. Solid State Tech- Shabbir A, Ali N, Ahmed J, Zafar B, Rasheed A, Sajid M, Ahmed
nol 64(2):2616–2621 A, Dar SH (2021) Satellite and scene image classification based
Patel K, Bhatt C, Mazzeo PL (2022) Deep learning-based automatic on transfer learning and fine tuning of resnet50. Math Probl Eng
detection of ships: An experimental study using satellite images. 2021:1–18
J Imaging 8(7):182
Negrel R, Picard D, Gosselin P-H (2014) Evaluation of second-order
visual features for land-use classification. In: 2014 12th Interna- Publisher’s Note Springer Nature remains neutral with regard to juris-
tional Workshop on Content-Based Multimedia Indexing (CBMI). dictional claims in published maps and institutional affiliations.
IEEE, pp 1–5
Lv Q, Dou Y, Niu X, Xu J, Xu J, Xia F (2015) Urban land use and land Springer Nature or its licensor (e.g. a society or other partner) holds
cover classification using remotely sensed sar data through deep exclusive rights to this article under a publishing agreement with the
belief networks. J Sensors 2015 author(s) or other rightsholder(s); author self-archiving of the accepted
Basu S, Ganguly S, Mukhopadhyay S, DiBiano R, Karki M, Nemani manuscript version of this article is solely governed by the terms of such
R (2015) Deepsat: a learning framework for satellite imagery. In: publishing agreement and applicable law.
Proceedings of the 23rd SIGSPATIAL international conference on
advances in geographic information systems. pp 1–10

123