Integration of convolutional neural networks with parcel-based image analysis for crop type mapping from time-series images

Abstract

Timely and accurate crop mapping is crucial for yield prediction, food security assessment and agricultural management. Convolutional neural networks (CNNs) have become powerful state-of-the-art methods in many fields, including crop type detection from satellite imagery. However, existing CNNs generally have large number of layers and filters that increase the computational cost and the number of parameters to be learned, which may not be convenient for the processing of time-series images. To that end, we propose a light CNN model in combination with parcel-based image analysis for crop classification from time-series images. The model was applied on two areas (Manisa and Kırklareli) in Türkiye using Sentinel-2 data. Classification results based on all bands of the time-series data had overall accuracies (OA) of 89.3% and 88.3%, respectively for Manisa and Kırklareli. The results based on the optimal bands selected through the Support Vector Machine–Recursive Feature Elimination (SVM-RFE) method had OA of 86.6% and 86.5%, respectively. The proposed model outperformed the VGG-16, ResNet-50, and U-Net models used for comparison. For Manisa and Kırklareli respectively, VGG-16 achieved OA of 86.0% and 86.5%, ResNet-50 achieved OA of 84.1% and 84.8%, and U-Net achieved OA of 82.2% and 81.9% based on all bands. Based on the optimal bands, VGG-16 achieved OA of 84.2% and 84.7%, ResNet-50 achieved OA of 82.4% and 83.1%, and U-Net achieved OA of 80.5% and 80.2%. The results suggest that the proposed model is promising for accurate and cost-effective crop classification from Sentinel-2 time-series imagery.

Full text

Vol.:(0123456789)
Earth Science Informatics (2025) 18:303
https://doi.org/10.1007/s12145-025-01819-8
RESEARCH
Integration ofconvolutional neural networks withparcel‑based image
analysis forcrop type mapping fromtime‑series images
MuslumAltun1,2· MustafaTurker2
Received: 13 November 2024 / Accepted: 16 February 2025
© The Author(s) 2025
Abstract
Timely and accurate crop mapping is crucial for yield prediction, food security assessment and agricultural management.
Convolutional neural networks (CNNs) have become powerful state-of-the-art methods in many fields, including crop type
detection from satellite imagery. However, existing CNNs generally have large number of layers and filters that increase the
computational cost and the number of parameters to be learned, which may not be convenient for the processing of time-series
images. To that end, we propose a light CNN model in combination with parcel-based image analysis for crop classification
from time-series images. The model was applied on two areas (Manisa and Kırklareli) in Türkiye using Sentinel-2 data.
Classification results based on all bands of the time-series data had overall accuracies (OA) of 89.3% and 88.3%, respectively
for Manisa and Kırklareli. The results based on the optimal bands selected through the Support Vector Machine–Recursive
Feature Elimination (SVM-RFE) method had OA of 86.6% and 86.5%, respectively. The proposed model outperformed the
VGG-16, ResNet-50, and U-Net models used for comparison. For Manisa and Kırklareli respectively, VGG-16 achieved OA
of 86.0% and 86.5%, ResNet-50 achieved OA of 84.1% and 84.8%, and U-Net achieved OA of 82.2% and 81.9% based on
all bands. Based on the optimal bands, VGG-16 achieved OA of 84.2% and 84.7%, ResNet-50 achieved OA of 82.4% and
83.1%, and U-Net achieved OA of 80.5% and 80.2%. The results suggest that the proposed model is promising for accurate
and cost-effective crop classification from Sentinel-2 time-series imagery.
Keywords Parcel-wise· Light CNN model· Classification· Sentinel-2· Time-series· Crop detection
Introduction
With the recent developments in remote sensing, image pro-
cessing and analysis techniques, space imagery has become
an important data source in many fields. Agriculture is one
of the areas in which the advantages of remote sensing can
be used best as the crops grown within agricultural fields are
dynamic and therefore their detection and monitoring must
be carried out frequently (Moysiadis etal. 2020). In this
regard, in recent years, space imagery has been preferred
as an important data provider in studies related to the deter-
mination of crop pattern with high accuracy, monitoring of
temporal changes and effective management of regional or
global sustainable agricultural areas (Pandey and Pandey
2023; Xu etal. 2024). Today, remote sensing technology
is very widely used in crop detection and monitoring (Han
etal. 2024; Omia etal. 2023). Automatic image classifi-
cation is the most commonly used method of information
extraction from remote sensing data (Shi etal. 2025; Wang
etal. 2024).
Deep learning algorithms have picked up improvement
in latest period and have demonstrated extraordinary
achievement for integrating spatial and temporal resolution
components for crop classification. Traditional machine
learning methods typically rely on derived attribute data
for input (Khanzode and Ravindra DS 2020). In contrast,
deep learning techniques enable machines to process raw
data, like pixel brightness values from raw images, and
automatically extract the required features for detection
Communicated by: Hassan Babaie
* Muslum Altun
muslum.altun@hacettepe.edu.tr
Mustafa Turker
mturker@hacettepe.edu.tr
1 Graduate School ofScience andEngineering, Hacettepe
University, 06532Beytepe, Ankara, Türkiye
2 Department ofGeomatics Engineering, Hacettepe University,
06800Beytepe, Ankara, Türkiye
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 303 Page 2 of 28
or classification across multiple pixel levels (Archana
and Jeevaraj 2024; Zhu etal. 2017). As for classification
processes, multiple layers of representation in a network
expand directions of the input data that are substantial
for distinction and that prevent unrelated diversification
(Archana and Jeevaraj 2024). This is extremely useful for
crop classification due to the complicated associations of
interior biochemical methods, inherent associations among
environmental parameters, and inconstant crop growing
condition.
In recent years, convolutional neural networks (CNNs),
a network architecture for deep learning, has become the
well-established popular hot topic in object detection stud-
ies (Simon and Uma 2020; Othman etal. 2016). Because of
their dominance superiority in high level object extraction
and pattern recognition, the CNN models have evidenced
extensively perfect potential in many image processing fields
(Sargent etal. 2017; Zhao etal. 2017a, b). Agriculture is
among the usual application areas of deep learning methods.
In particular, crop detection from space imagery is among
the most common implementations of deep learning meth-
ods in agriculture.
So far many deep learning approaches have been pro-
posed based on pixel-wise or patch-wise classification
(Adrian etal. 2021; Bhosle and Musande 2022; Kussul
etal. 2017; Mazzia etal. 2020; Sagan etal. 2021; Yan and
Ryu 2021; Yang etal. 2020; Zhao etal. 2019). Pixel-wise
approaches that depend strictly on spectral quality can detect
land cover types, but can be inadequate to separate land uses
that frequently consist of multiple classes, and such diffi-
culties are especially important in agriculture (Zhao etal.
2016). Spatial image information, such as texture or con-
text, must be integrated to examine crop patterns through
moving convolutions (Yousaf etal. 2024; Schutt etal 2022;
Niemeyer etal. 2014). It can be argued that pixel-wise CNN
methods require that regular image convolutions are pre-
defined, while the actual object shapes are irregular in real
world (Yousaf etal. 2024). Although pixel-wise techniques
are easy to implement, regular-shaped and uniformly sized
image patches introduces additional challenges. One chal-
lenge is the misalignment between the fixed, regular shape
of the image patch and the actual irregular footprint of the
objects in the image. The other challenge is the pixel-level
mixed-class effect that arises from using regular image
patches. As a result, square shaped image patches cannot
cover the irregular shape objects and an amalgamation of
multiple class categories bring about serious errors in clas-
sification. Because of these challenges, classification accura-
cies do not always become satisfactory. Hence, pixel-wise
CNN approaches demonstrate disadvantages and difficul-
ties (Liu and Shi 2020; Qiu etal. 2020; Mei etal. 2017;
Lee and Kwon 2017; Liu etal. 2016). To overcome existing
challenges, development of object-wise CNN models based
on deep learning algorithms must be realized (Charisis and
Argyropoulos 2024; Tian etal. 2024; Patel etal. 2023).
Recently, object-wise deep learning methods have
been extensively implemented for land cover/use mapping
and object detection from space imagery. For example,
Zhou etal. (2021) proposed an object-wise deep-learning
approach to map local climate zones from Sentinel-2 images.
After the segmentation process, all segments were edited
in order to fit and match the convolution size defined for
the CNN algorithm. The parcel-based method was reported
to be superior by approximately 50%. Zhang etal. (2018a,
b) detected urban land use types using object-based and
pixel-based CNN models. The overall accuracy (OA) was
computed as 81% for the pixel-based method, while for
the object-based method it was 84%, 87% and 89%, for the
48 × 48, 128 × 128 and 176 × 176 size image patches, respec-
tively. Martins etal. (2020) proposed a multi-object-based
CNN method for the classification of large areas using high
resolution (1m) images. The method consists of image seg-
mentation with the mean shift algorithm and a multiple CNN
algorithm for object classification. Their method provided
the best results with the OA of ~ 87%. The model developed
by Zhang etal. (2018a, b) for land cover classification from
hyperspectral imagery effectively detected spectral-spatial
features and achieved higher accuracy when compared to
recently developed methods. To detect complex patterns
in urban areas from high resolution WorldView-2 imagery,
Zhao etal. (2017a, b) used an object-based deep learning
approach. They reported that building types that cannot be
detected with the pixel-wise method were detected with the
object-wise method. Jin etal. (2019) combined object-wise
approach with deep CNN for the classification of land use
types. They constructed rule set of objects based on seg-
mentation, and classification was performed based on the
segmentation objects. Their proposed method effectively
solves the problem of misclassification of typical features
and provides better classification accuracy than the use of a
method based on traditional CNN.
Considering the complexities of agricultural practices,
monitoring and assessing agricultural activities over exten-
sive areas using localized methods proves to be highly
challenging. At this juncture, remote sensing provides an
effective means for determining the spatial distribution of
agricultural lands, enabling comprehensive and large-scale
analysis, to produce field boundaries and high-quality in sea-
son crop maps. Accurate classification of satellite images is
of a great importance for the mapping of crop types and crop
rotations that are very important for crop management in a
spatial–temporal context.
In this study, we propose a light CNN model in combina-
tion with parcel-based image analysis to extract crop types
in agricultural areas using time-series Sentinel-2 imagery.
Parcel-based image anaysis accounts for the shape and size of
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 Page 3 of 28 303
agricultural fields to enhance the accuracy of crop type map-
ping. We use the existing agricultural field boundary data to
divide the image into field units. To generate sample image
patches for the training dataset we convert the fields with vari-
ous sizes and shapes into regular patches by filling the gaps
between the convolution window (bounding box) and the image
subset of the field. The experimental results on two crop-rich
agricultural areas showed that the proposed approach effectively
combined the CNN model with parcel-based image analysis
and achieved notably good results. Furthermore, the proposed
model outperformed the VGG-16, ResNet-50, and U-Net mod-
els used for comparison. The main contributions are as follows:
(i) development of a light CNN deep learning model, (ii) inte-
gration of CNN and parcel-based image analysis approach for
crop type mapping, (iii) generation of training image patches
based on agricultural fields with various sizes and shapes, and
(iv) demonstration of the potential of the proposed CNN model
in conjunction with parcel-based image analysis method in crop
classification from time-series of Sentinel-2 images.
The remainder of this article presents study areas and
dataset (Sect.2), a methodology that includes data pre-pro-
cessing, the structure of the proposed CNN model, feature
selection, image classification, and accuracy assessment
(Sect.3), results and discussions on the findings and com-
putation time (Sect.4), and conclusions (Sect.5).
Study area anddatasets
Study area
In this work, two study areas were selected in different
regions of Türkiye to demonstrate the success of the pro-
posed parcel-based CNN classification model. Both study
areas exhibit substantial heterogeneity and present signifi-
cant challenges in agricultural crop harmonies in terms of
diversity and quantity, and are thereby appropriate to test
the effectiveness of the generalisation capability of the
proposed agricultural crop detection deep learning model.
Study area 1 (Manisa District): The first study area (her-
after named as SA1) is located in the southern west part
of Türkiye, within the Manisa province and has the size
of 520 km2 (Fig.1). The coordinates of the study region
are as follows: 573369,16m, 4275559,88m (North West);
573369,16m, 4257782,09m (South West); 603175,50m,
4275559,88m (North East); 603175,50m, 4257782,09m
(South East), with the projection information of WGS 84,
UTM-Zone 35 N, and Central meridian 27. In this area,
most of the parcels are planted with summer crops, while
others consist of fallow lands and non-agricultural zones
that remain uncultivated. The main crops include wheat,
tomatoes, corn, cotton, grapes and alfalfa.
Fig. 1 The locations of the study areas SA1 and SA2 and the false color composites of Sentinel-2 images
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 303 Page 4 of 28
Study area 2 (Kırklareli District): The second study area
(hereafter named as SA2) is located in the Northern West
part of Türkiye, within the Kırklareli provice and has the size
of 5200 km2 (Fig.1). The coordinates of the study region
are as follows: 489948,60m, 4563385,82m (North West);
489948,60m, 4661892,24m (South West); 590713,10m,
4563385,82m (North East); 590713,10m, 4661892,24m
(South East) with the projection information of WGS 84,
UTM-Zone 35 N, and Central meridian 27. In this area the
main crops include barley, sunflower, wheat, rice, corn, triti-
cale and alfalfa. Different from SA1, this study area (SA2)
contains several different industrial crops, including sun-
flower, rapeseed, triticale and oats.
Sentinel‑2 data acqusition
Data acquisition for remote sensing studies can be costly
and time-consuming. The freely available Sentinel-2 satel-
lite images, provided by the European Space Agency (ESA)
through its website, plays a crucial role in mitigating these
challenges (ESA 2024). Sentinel-2 satellites are widely uti-
lized across various fields due to its broad range of spectral
bands and ability to provide different spatial resolutions.
The lower temporal resolution feature of Sentinel-2 com-
bined constellations satellite (Sentinel-2A and Sentinel-
2B) with global coverage and a 5-day revisit frequency
enables the rapid utilization of images within a short time
interval. Thanks to these features, agriculture is one of the
primary application areas of Sentinel-2 data (Wang etal.
2016; Spoto etal. 2012; Van and McVicar 2004; Siachalou
etal. 2015).
For SA1, eight Sentinel-2 images were acquired from
April 10, 2017 / Day-of-Year (DOY) 100 to November 16,
2017 / DOY 320. For SA2, six Sentinel-2 images were col-
lected from April 12, 2021 / DOY 102 to September 19,
2021 / DOY 262. Information about the time-series Senti-
nel-2 images used for the study areas are given in Table1.
Figure1a shows the false colour composite image of the
Sentinel-2 image dated July 2, 2017 for SA1. Figure1b
shows the false colour composite image of the Sentinel-2
image dated August 10, 2021 for SA2.
Ground truth data
Farmer Registration System (FRS) is a comprehensive
database managed by the Republic of Türkiye Ministry
of Agriculture and Forestry (RTMAF 2024). This system
is designed to monitor, audit, report, and evaluate agri-
cultural subsidies to provide the farmers. In this study,
we utilized the existent FRS data as ground truth for the
reference purposes. The FRS data are widely used as
ground truth data owing to their high quality and have
been extensively accepted as reference data for crop map-
ping (Altun and Turker 2022a; Altun and Türker 2022b).
The spatial data stored in the FRS comprises the loca-
tions of the parcels, while the attribute data contains land
registry information, including parcel IDs, parcel areas,
cultivated areas, crop types, province names, and district
names, among other relevant attributes. The field bounda-
ries and the image subsets selected from SA1 and SA2 are
shown in Fig.2.
Method
Figure3 illustrates the flowchart that outlines the steps of
the proposed method. The method comprises five principal
stages: (i) data pre-processing, (ii) samples generation for
training, validation and test, (iii) classification process,
(iv) variable selection, and (v) accuracy assessment. Ini-
tially, the necessary pre-processing operations are applied
to Sentinel-2 images and the required adjustments are per-
formed on the FRS data, to enable its suitability for use
as reference data. In the second step, the model is trained
with the training samples generated based on the agricul-
tural fields. In the third step, image classification is car-
ried out using the proposed CNN model and the models
used for comparison. In the next step, variable selection
process is carried out to select the optimal bands from
time-series Sentinel-2 images. As the final step, the model
train accuracy and the classified map accuracy assessments
are handled.
Table 1 Information about the time-series Sentinel-2 images used for SA1 and SA2
SA1 Acqusition time 2017–04–10 2017–05-03 2017–06-02 2017–07-02 2017–08-01 2017–09-07 2017–10-10 2017–11–16
Day of Year (DOY) 100 123 153 183 213 250 283 320
Cloud Cover (%) 0.06 3.69 0.33 0 0 0 0.52 0
SA2 Acqusition time 2021–04–12 2021–05–12 2021–06-06 2021–07-01 2021–08–10 2021–09–19
Day of Year (DOY) 102 132 157 182 222 262
Cloud Cover (%) 0.05 7.18 2.60 0 0.08 2.96
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 Page 5 of 28 303
Data pre‑processing
Sentinel‑2 data pre‑processing
Atmospheric correction is a critical pre-processing step in
optical satellite image analysis, ensuring consistency across
multiple acquisition dates and adjacent image tiles. It is
particularly crucial for deriving biophysical variables, con-
ducting multi-temporal studies, and enhancing the accuracy
of image classification processes. The Sentinel-2 Level-1C
(L1C) product delivers radiometrically corrected imagery,
while the Level-2A (L2A) product includes atmospheric
correction, producing surface reflectance data suitable for
applications such as vegetation analysis and agricultural
crop classification (Kganyago etal. 2020; Sola etal. 2018).
In the present case, Level-2A (L2A) images were used for
both study areas.
Four bands (B2-Blue, B3-Green, B4-Red and B8-NIR)
with 10m resolution and six bands (B5-Red Edge 1, B6-Red
Edge 2, B7-Red Edge 3, B8A-Narrow-NIR, B11-SWIR 1,
B12-SWIR 2) with 20m resolution were used. The 20m res-
olution bands were sharpened to 10m resolution using the
high-pass filter (HPF) image fusion algorithm (Chavez etal.
1991). The general idea of HPF is to pull out high-frequency
information from the PAN image through a high-pass filter
and add that to low resolution bands with a specified weight
(Gangkofner etal. 2008). Band B4 was used as the PAN-like
band for the bands B11 and B12, while B3 band was used
as the PAN-like band for the bands B5, B6, B7 and B8A
(Zheng etal. 2017). For resampling, the nearest neighbour
(NN) algorihtm was chosen. After resampling, six bands
were integrated with four bands (Blue, Green, Red, NIR)
with 10m resolution, and each resultant image contained ten
bands with 10m resolution. In addition to original bands,
the well-known Normalized Difference Vegetation Index
(NDVI) (Tucker 1979) calculated for each image date was
also used in the classification. NDVI extracted from multi-
temporal images have been extensively used for agricultural
crop classification (Inglada etal. 2015; Schultz etal. 2015;
Valero etal. 2016; Sun etal. 2019). Consequently, a total
of 88 band (8 dates × 10 bands + 1 NDVI) and 66 band (6
dates × 10 bands + 1 NDVI) composites were created for
SA1 and SA2, respectively.
Unlike most land cover types, agricultural areas exhibit
significant variability due to high frequency of dynamic
changes that occur over a short time period. Most crop types
have a vegetation period lasting only a few months. Dur-
ing this period, the plant's phenology undergoes substantial
changes multiple times between sowing and harvest. Fur-
thermore, the onset and completion of the crop life cycle
vary depending on the crop type. Different crop types have
distinct planting, growing, and harvesting periods. Monitor-
ing these transitional phases and assessing their impact on
crop classification using single-date or non-temporal satellite
Fig. 2 Illustration of the field boundaries (left column) and the image subsets (right column) from (a) SA1 and (b) SA2
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 303 Page 6 of 28
Fig. 3 The flowchart of the methodology (B-Blue, G-Green, R-Red,
VRE_1-Vegetation Red Edge 1, VRE_2-Vegetation Red Edge 2,
VRE_3-Vegetation Red Edge 3, NIR-Near Infrared, SWIR_1-Short-
wave Infrared 1, SWIR_2-Shortwave Infrared 2, FRS-Farmer Reg-
istration System, NDVI-Normalized Difference Vegetation Index,
RFECV-Recursive Feature Elimination with cross-validation, CNN-
Convolutional Neural Network)
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 Page 7 of 28 303
imagery can be highly challenging (Bargiel 2017; Defourny
etal. 2019). Time-series satellite data offers opportunities
for detecting agricultural dynamics by leveraging multi-tem-
poral classification approaches that integrate phenological
information. These dynamic variations introduce challenges
for classification models; however, they also serve as a valu-
able source of information for improving classification accu-
racy. Furthermore, time-series satellite imagery is essential
for extracting specific crop temporal feature patterns, as it
enables the utilization of phenological information over time
to enhance classification models and improve agricultural
analysis (Blickensdörfer etal. 2022; Gallo etal. 2023).
Ground truth data pre‑processing
Before using as ground truth data, the necessary pre-pro-
cessing operations were performed on FRS reference data.
These mainly included relating the non-spatial information
(parcel attribute information) with the metadata of the sat-
ellite images and excluding small parcels (under 1000 m2)
with no cultivated crops from further processings. The FRS
data belonging to SA1 and a subset of an image with the
field boundaries superimposed are shown in Fig.2a. Simi-
larly, Fig.2b shows the FRS data belonging to SA2 and a
subset of an image with the FRS data superimposed.
Train, validation andtest samples generation
Sample parcels were acquired from the FRS reference
ground truth data by means of random sampling. The refer-
ence dataset for training, validation and testing were created
by means of field samples and visual inspection of the time-
series images. Eight and fifteen crop types were defined for
SA1 and SA2, respectively. In SA1, corn (drying corn) is
planted in the spring and harvested in the autumn. In the
region, silage corn is planted as the second crop Thus, in
this work, drying corn is referred to as "corn" and silage
corn (late corn) is designated as "corn_2". Training dataset
was used to teach the models to identify desired patterns or
perform a particulartask. The validation dataset was used to
select the optimal parameters for the models. The test data-
set was used to assess the quality and reliability of the final
classification results (Yang etal. 2020). The sample divi-
sion ratio was determined by taking inspiration from similar
up-to date studies previously conducted in the deep learning
domain (Sahu and Dash 2024; Nayak etal. 2023). Here, the
training, validation, and test datasets were independent of
each other and randomly divided by a ratio of 75%, 15%
and 10% respectively. To acquire enough representative
samples, the sample size for each class was set differently
(Table2). In Table2, the second column represents the crop
type. In the third, fourth, fifth and sixth columns, the first
value represents the number of parcels and the second value
represents the number of pixels. A total number of 1024
and 15,768 sample parcels were acquired for SA1 and SA2,
respectively.
Due to their architectures, CNN deep learning models
operate on the input image with regular geometric shaped
convolutions. However, agricultural parcels have irregular
geometric shapes on earth surface. To eliminate this prob-
lem, we use a parcel-based strategy in the integration of FRS
parcel data and satellite imagery (Fig.4(a)). After clipping
the image falling within the boundaries of the sample par-
cels (Fig.4(b)), the gaps between the convolution window
(the bounding box of the clipped parcel image subset) and
the parcel image subsets are filled with the original pixel
values of the parcel image subsets. For this gap filling (pad-
ding) operation, the nearest pixels to the gap area within
the parcel image subsets are used. With this process, the
input image subsets evolve into regular geometric shapes and
become compatible with the model architecture (Fig.4(c)).
Figure4 shown the padding process on three parcels selected
from SA1. After processing all sample parcels in this way,
we used them as input data in the model training process.
Within the time series framework, we trained the proposed
model and the tested models using dynamic input patch
sizes since the parcels cover different areas and consist of
different number of pixels. In this respect, the process of
adjusting irregular shape image subsets into regular image
patches (padding) has been a crucial aspect of the proposed
approach.
CNN classification model
The proposed CNN model is illustrated in Fig.5. The model
stands out with its simple understandable and uncomplicated
architecture in terms of the parameters and parameter values.
It consists of 6 layers, including 2 convolutional layers, 2
pooling layers (two pairs of convolution and max-pooling
layers), 1 flattening layer and 1 fully connected layer. A con-
volution is applied on the image by moving a kernel of a size
3 × 3. After each convolutional layer, max pooling operation
is performed with a kernel of a size 2 × 2. For the first and
second layers, the number of filters were set to 32 and 64
respectively to learn the input train sample and contextual
pattern for each detected crop. The output of the last pooling
layer is flattened, on which single fully connected layers with
64 neurons are built. Rectified linear unit (ReLu) is used as
the activation function. To avoid the model from overfitting,
a regularisation technique ‘dropout’ is applied before the
fully connected layer (Li etal. 2021a, b, c; Srivastava etal.
2014; Phung and Rhee 2019). The dropout layer is not used
since the proposed model is simple and does not repeat the
essential steps of the main CNN architecture (it is used only
once). The last layer (output layer) produces the classes and
represents the probability vector with multi-class softmax
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 303 Page 8 of 28
activation function. The parameters of the proposed model
are given in Table3. The convolution size of 64 × 64 was
defined for both study areas, as all sample parcels can be
fully represented. Additionally, the CNN models used for
testing were trained on the same sample set to suit our pur-
poses, and experiments were conducted accordingly.
The efficiency of the proposed model was compared with
VGG-16, ResNet-50, and U-Net which are three well-known
deep neural network architectures used for segmentation,
image classification, and object detection. These models
have also been widely used in agricultural crop detection
and classification (Fathipoor etal. 2023; Hu etal. 2020; Xu
etal. 2022).
VGG-16 was developed by the Visual Geometry Group
at the University of Oxford in 2014. It includes 16 layers,
including 13 convolutional layers and 3 fully connected
layers. For the convolution step, the filters with a size of
3 × 3 are used and after each convolution layer comes a
pooling layer. VGG-16 has achieved great performance
in image classification tasks but is relatively slow due to
its large number of parameters (Simonyan and Zisser-
man 2015). It is one of the frequently used models for
agriculture (Thenmozhi and Reddy 2019; Thakur etal.
2023; Nowakowski etal. 2021).
ResNet-50 was developed by Microsoft Research in 2015.
It uses residual connections to help alleviate the vanishing
gradient problem. It contains a total of 50 layers, includ-
ing convolutional, pooling, and fully connected layers.
ResNet-50 has demonstrated state-of-the-art performance
across various computer vision applications, particularly in
image classification and object detection. (He etal. 2016)
and agricultural crop type mapping (Shah etal. 2023; Rani
and Singh 2022).
U-Net includes four encoder blocks and four decoder
blocks, which are interconnected through a bridge. In the
encoder part, the spatial sizes are halved while doubling
the number of filters at each encoder block. In the decoder
part, the spatial sizes are doubled and the number of fea-
ture channels is halved. U-Net was specifically designed to
learn effectively from a limited number of training samples.
(Ronneberger etal. 2015). Of these three models used for
comparison, U-Net is probably the most frequently used
one for crop mapping (Fan etal. 2022; Liu etal. 2022;
Wei etal. 2019).
Table 2 The number of sample
parcels and pixels for SA1 and
SA2
Study Area (Number of Parcels / Pixels)
Crop Type Training Data Validation Data Test Data Total Data
SA1 Wheat 141 / 340668 28 / 61515 19 / 41742 188 / 443925
Tomatoe 39 / 115920 8 / 24364 5 / 15227 52 / 155511
Corn 117 / 239724 23 / 55625 16 / 38695 156 / 334044
Corn_2 38 / 68949 7 / 19047 5 / 13605 50 / 101601
Cotton 42 / 135810 8 / 36103 6 / 27077 56 / 198990
Grapes 303 / 305964 61 / 78262 40 / 51320 404 / 435546
Alfalfa 11 / 54243 3 / 3040 2 / 2027 16 / 59310
Olive Trees 77 / 247338 15 / 63018 10 / 42012 102 / 352368
Total 768 / 1508616 153 / 342265 103 / 230414 1024 / 2081295
SA2 Barley 1206 / 125594 241 / 25098 161 / 16767 1608 / 167458
Sunflower 2111 / 225822 422 / 45143 282 / 30167 2815 / 301132
Wheat 2304 / 245393 461 / 49100 307 / 32698 3072 / 327190
Meadow Grass 430 / 26060 86 / 5212 58 / 3515 574 / 34787
Rice 545 / 47673 109 / 9535 73 / 6386 727 / 63593
Walnut 599 / 54321 119 / 10792 80 / 7255 798 / 72367
Poplar Grove 123 / 3698 24 / 722 17 / 511 164 / 4931
Rapeseed 1013 / 178868 202 / 35668 136 / 24014 1351 / 238549
Hungarian Vetch 129 / 8306 25 / 1610 18 / 1159 172 / 11074
Corn 1288 / 166285 258 / 33309 172 / 22206 1718 / 221799
Triticale 560 / 82848 111 / 16422 75 / 11096 746 / 110366
Grape 330 / 17908 66 / 3582 44 / 2388 440 / 23877
Forage Pea 93 / 11840 18 / 2292 12 / 1528 123 / 15660
Alfalfa 459 / 37339 92 / 7484 61 / 4962 612 / 49785
Oats 636 / 57732 127 / 11528 85 / 7716 848 / 76976
Total 11826 / 1289658 2361 / 257474 1581 / 172412 15768 / 1719544
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 Page 9 of 28 303
Feature selection
Selection of the appropriate features is important in image
classification of time series images. Through the optimal
features selection process, dimensionality is reduced while
the classes to be extracted are effectively distinguished with
sufficient accuracy (Jensen and Lulla 1987). In this work, to
reduce the number of bands and generate optimal band sets
from time-series Sentinel-2 images, we used the Support
Vector Machine–Recursive Feature Elimination (SVM-RFE)
Fig. 4 (a) FRS parcel data superimposed on the image, (b) Image extraction enclosing parcel area, (c) The image patches upon the padding process
Fig. 5 The proposed CNN model for agricultural crop classification
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 303 Page 10 of 28
method (Grabska etal. 2020). Recursive feature elimination
(RFE) eliminates the least effective feature (or features) until
the optimal number of features is attained (Chandrashekar
and Sahin 2014; Li etal. 2021a, b, c). The results of the vari-
able selection procedure are shown in Figs.6(a) and 6(b).
As shown in Fig.6, the most relevant bands were selected
according to the peak points in the generated charts. For SA1
and SA2 respectively, the optimal number of bands were
determined as 41 and 29.
Classification process
Model training was carried out by configuring the classi-
fier's parameters using the training dataset. After training the
CNN models, the final evaluation was performed using the
validation samples. The inference phase of the CNN models
entails predicting class labels associated with specific spec-
tral categories such as agricultural crop types, based on the
trained CNN architecture. This process utilizes the identi-
fied convolutional positions and input image patches. At the
model detection stage, the trained CNN models are directly
applied to predict the label corresponding to each specified
class (crop type). Subsequently, each identifited crop type is
labeled using the trained model’s output culminating in the
comprehensive classification of the entire image. In SA1, the
crop types are wheat, tomatoes, corn, corn-2, cotton, grapes,
alfalfa, and olive trees. In SA2, the crop types are barley,
sunflower, wheat, meadow grass, rice, walnut, poplar grove,
rapeseed, hungarian vetch, corn, triticale, grapes, forage pea,
alfalfa, and oats.
As also mentioned above, an image stack with 11 bands
was generated by compositing for each time-series date and
used accordingly during classification. Time-series Senti-
nel-2 images used for SA1 and SA2 are given in Table1.
For SA1, an image stack of 88 bands across eight time-series
dates was incorporated into the classification process. The
goal was to detect eight distinct agricultural crop types. For
SA2, a total of 66 bands across six time-series dates were
included in the classification process. The objective was to
detect fifteen challenging agricultural crop varieties. The
core element of the method was the proposed CNN clas-
sification model.
Hyperparameter settings
Several experiments were carried out to determine the num-
ber of training epochs and values for the model parameters.
The optimal number of training epochs were determined
through cross-validation, a highly recommended method
Table 3 The parameters of the proposed CNN model. (*Number of
input image band; **Number of classes)
No Layer Output Size Filter Size Stride Size
1 Input 64 × 64 × * - -
2 Convolution 1 64 × 64 × 32 3 × 3 -
3 Activation Function
(ReLu)
64 × 64 × 32 - -
4 Pooling (Max pooling) 32 × 32 × 32 -2 × 2
5 Convolution 2 32 × 32 × 64 3 × 3 -
6 Activation Function
(ReLu)
32 × 32 × 64 - -
7 Pooling (Max pooling) 16 × 16 × 64 -2 × 2
8 Flattening 1 × 1 × 16384 - -
9 Fully Connected 1 × 1 × ** - -
10 Softmax Activation 1 × 1 × ** - -
Fig. 6 The graphical results of the SVM-RFE method for (a) SA1 and (b) SA2
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 Page 11 of 28 303
for stopping network training (Furtuna etal. 2011; Ma etal.
2017; Ramezan etal. 2019). The models were trained in a
Python integrated development and learning environment for
10, 50, 100, 200, 300, and 500 epochs, and 100 was found to
be the optimal number for both study areas. Cross-entropy
loss was used when adjusting model weights during training.
Categorical cross-entropy is also known assoftmax loss. It
is asoftmax activationplus across entropy loss used for
multiclass classification (Shah etal. 2022; Chand 2021). We
used categorical cross-entropy loss because both study areas
contain multiple crop types.
During the implementation process, fine-tuning of the
proposed model was performed. Fine-tuning is a transfer
learning method that refines a pre-trained neural network
using different data. It updates a subset or all feature extrac-
tion layers in CNNs, balancing reduced training time with
improved performance (Wojciuk etal. 2024; Singh etal.
2023; Gill etal. 2023). It can be applied to the entire model
or specific layers, with optimization techniques significantly
influencing performance. Adam optimizer is a widely used
fine-tuning method. Due to its adaptive learning rates and
efficiency (Poojary & Pai 2019; Shahade etal. 2023), Adam
optimizer surpasses traditional methods like Stochastic Gra-
dient Descent (SGD) and Adagrad optimizers. Therefore, in
this work we selected the Adam optimizer method for the
fine-tuning of the proposed model. The learning rate was set
to 0.001 for 100 epochs.
Accuracy assessment
We evaluated the performance of the generated crop maps
using the model accuracy and the classified map accuracy
metrics. Model accuracy performance was automatically cal-
culated in the Python programming environment. For the
classified map accuracy assessment, we computed standard
error matrices (Congalton 2001) and calculated the Overall
Accuracy (OA) and Kappa coefficient values. Moreover,
we also evaluated the performance of each class based on
producer’s accuracy (PA) and user’s accuracy (UA) (Con-
galton 1991). For SA1 and SA2 respectively, a total num-
ber of 2232 and 15,471 sample points were seeded inside
the parcels using the stratified random sampling method.
To minimize or prevent bias in the models, training parcels
were excluded from the classified map accuracy assessment.
Therefore, only the validation and test parcels were utilized
for computing the post-classification accuracy. In addition,
we evaluated the statistical significance of the results using
McNemar’s test, which estimates the statistical significance
of differences between two classification outcomes and
helps determine the models providing superior perfor-
mance (Foody 2004; De Leeuw etal. 2006; Petropoulos
etal. 2012). The test is computed using Eq.1 (Foody 2004;
McNemar 1947; Yan etal. 2006).
where,
f1,2
is the number of samples correctly classified
by the first model but misclassified by the second model,
while
f2,1
refers to the opposite: the samples that are cor-
rectly classified by the second model but misclassified by
the first model. The test statistic follows a chi-squared (
x2
)
distribution, where the square of
Z
follows a
x2
distribution
with one degree of freedom (Foody 2004; Dingle Robertson
and King 2011). Thus, the McNemar test can be stated using
the chi-squared formula in Eq.2 (Foody 2004).
A difference between the results of two classification
models is deemed statistically significant at the 95% confi-
dence level (p ≤ 0.05) if the
Z
value exceeds 1.96, or if the
x2
value is greater than or equal to 3.84 (Mallinis etal. 2014;
Abdel-Rahman etal. 2014; Omer etal. 2015; Li etal. 2017).
Results anddiscussion
Results ofstudy area SA1
The train, validation and test accuracies achieved for all
bands and the selected optimal bands are given in Table4.
As shown in Table4, the proposed model provided the great-
est train and test accuracy for both all bands and optimal
bands, with the train accuracy values of 96.8% and 95.4%
for all and optimal bands, respectively and the test accu-
racy values of 82.6% and 81.8% for all and optimal bands,
respectively. For validation, the proposed model also
acquired the highest accuracy value of 84.7% for optimal
bands but achieved the second highest accuracy of 85.9%
after ResNet-50 (88.6%). These values clearly demonstrate
the better performance of the proposed model with respect
to VGG-16, ResNet-50, and U-Net.
The error matrices of the crop maps produced using the
proposed CNN model are given in Table14and Table15in
Appendix 1. Table5 displays the OA, PA, UA, and Kappa
values obtained for SA1 using the proposed model and the
tested VGG-16, ResNet-50, and U-Net models, based on all
bands. It can be easily observed that the proposed model
yielded an OA of 89.3%, and a Kappa value of 0.874. The
OA obtained by the proposed model thereby exhibited con-
siderable improvement compared to other tested models. On
the other hand, VGG-16, ResNet-50, and U-Net achieved
OAs of 86.0%, 84.1%, and 82.2% as well as Kappa values of
0.835, 0.814, and 0.792, respectively. Based on the proposed
(1)
Z
=
f
1,2−
f
2,1
√
f1,2 +f2,1
(2)
x
2=(f1,2 −f2,1)
2
(
f
1,2
+f
2,1)
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 303 Page 12 of 28
model, olive trees and wheat maintained the highest PAs
96.1% and 94.7%, respectively. Moreover, alfalfa, corn, and
corn-2 held PA values > 87%, while grapes and cotton held
PA values > 82%. Tomatoe was considered the poorest class
in terms of the average of PA and UA (~ 79%) due to con-
fusion with other classes. The comparison of the accuracy
metrics achieved by the proposed and tested models is pre-
sented in a chart in Fig.7.
Table6 shown the OA, PA, UA, and Kappa values
achieved by the proposed model and the tested models based
on the selected optimal bands. The same trends were per-
cieved when comparing the OA and Kappa values, where
the proposed model provided the highest OA of 86.6% and a
kappa value of 0.842, exhibiting considerable improvement
over the tested models. VGG-16, ResNet-50, and U-Net
achieved OAs of 84.2%, 82.4%, and 80.5%, and Kappa val-
ues of 0.814, 0.794, and 0.772, respectively. The crop types
most efficiently detected by the proposed model were wheat
and olive trees in terms of the average of PA and UA (~ 93%
and 92%, respectively). In addition, grapes, corn, and alfalfa
held PA values > 82%, while corn-2 held a PA value > 79%.
In this case, tomateo and cotton were considered the poor-
est classes in terms of the average of PA and UA (~ 74% and
78%, respectively) due to confusion with other classes. The
comparison of accuracy values achieved by the proposed and
tested models with the selected optimal bands is presented
in a chart in Fig.8.
The classification results obtained using the optimal
bands slightly decreased compared to those obtained using
all available bands. While the total number of bands was 88,
the optimal number of bands was determined to be 41. This
outcome was anticipated, as the reduction in the number of
bands typically leads to a corresponding decrease in classi-
fication accuracy. This overall trend is consistently observed
across the proposed and tested models implemented in SA1.
Figure9 provides visual inspections of the extracted crop
maps based on all bands using subset images of SA1. Since
the post-classification outputs are too massy to display,
only the results of typical scenarios are presented here. For
instance, Fig.9 shows the thematic labeled classification
maps, obtained from executed models to image patches,
along with several parcels belonging to the same spatial
area, and their respective boundaries.
Figure9 shown the comparative analysis of the results
in the selected four subset regions. The first row shows the
Sentinel-2 satellite image of the selected area, a false color
time series image taken on 2017–04–10. The remaining four
rows illustrate the classified images obtained using the pro-
posed model and the tested models.
As can be seen in b1 in the first row, the proposed model
correctly classified the parcel shown by an ellipse as tomato.
In contrast, the tested models misclassified this parcel as
non-cropland in the same scenario, as shown in (c1), (d1)
and (e1). For the parcel shown by a circle, the proposed (b1),
VGG-16 (c1), and U-Net (e1) models successfully identi-
fied corn_2, whereas ResNet-50 (d1) failed by mixing up
Table 4 Model set accuracy values of the proposed model and the
tested models for SA1
Model set Variable options
Model type All bands (%) Optimal
bands (%)
Train Proposed 96,8 95,4
VGG-16 92,3 91,9
ResNet-50 91,4 91,2
U-Net 93,1 92,7
Validation Proposed 85,9 84,7
VGG-16 82,5 82,3
ResNet-50 88,6 83,9
U-Net 84,7 82,2
Test Proposed 82,6 81,8
VGG-16 79,8 78,9
ResNet-50 75,3 75,1
U-Net 76,4 75,9
Table 5 Classification accuracy
comparison between the
proposed model and the tested
models with all bands for SA1
Crop Type Proposed VGG-16 ResNet-50 U-Net
PA UA PA UA PA UA PA UA
Wheat 94,7 94,7 89,8 91,9 87,2 90,3 83,3 89,5
Tomatoe 79,7 78,8 78,6 73,4 78,8 74,8 76,8 70,4
Corn 87,7 88,7 87,4 87,4 85,4 87,2 80,0 86,1
Corn_2 92,3 87,2 79,5 72,9 78,6 64,4 78,0 63,6
Cotton 82,0 78,7 81,3 76,4 81,1 72,4 80,0 72,4
Grapes 83,2 93,6 80,4 91,8 79,3 93,2 79,4 93,3
Alfalfa 94,2 82,5 88,0 77,8 79,1 75,2 79,1 71,6
Olive Trees 96,1 92,3 93,3 91,0 92,2 89,5 92,3 85,3
OA: 89,3 OA: 86,0 OA: 84,1 OA: 82,2
Kappa: 0,874 Kappa: 0,835 Kappa: 0,814 Kappa: 0,792
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 Page 13 of 28 303
with alfalfa. The proposed model (b1) and VGG-16 (c1)
accurately detected the crop types in the parcel shown by
a square, whereas ResNet-50 (d1) and U-Net (e1) failed in
correctly identifying the crop type of this parcel.
As shown in the second row of Fig.9, both the proposed
model (b2) and VGG-16 (c2) accurately classified the parcel
marked with a circle as cotton, whereas ResNet-50 (d2) and
U-Net (e2) fail to do so. For the parcel marked by a square,
the proposed (b2), VGG-16 (c2), and ResNet-50 (d2) mod-
els successfully detected the corn crop, whereas U-Net (e2)
failed by mixing up with other classes. The proposed model
also demonstrates superior crop identification for the parcel
marked with a circle and a square in the second row. How-
ever, in certain instances, such as the parcel marked with an
ellipse, the proposed model (b2) shows misclassification. It
is evident that ResNet-50 (d2) also failed. In the second row,
the models that performed accurately are VGG-16 (c2) and
U-Net (e2) models.
As seen in the third row, VGG-16 (c3) correctly detected
the crop types in all the marked parcels. In contrast,
ResNet-50 (d3) failed in correctly classifying these parcels.
On the other hand, the proposed model (b3) performed cor-
rect detection only for the parcel shown by a circle, and
misclassified the parcels shown by an ellipse and a square.
U-Net (e3) exhibited the same trend as the proposed model.
In the last row, both the proposed model (b4) and VGG-
16 (c4) inaccurately classified the parcel shown by an ellipse
as corn_2. ResNet-50 (d4) and U-Net (e4) correctly classi-
fied this parcel as wheat. The entire parcel shown by a circle
contains corn. But the proposed model (b4) and all tested
models (c4, d4 and e4) mixed corn and cotton, leading to
incorrect classification. On the other hand, the crop type of
Fig. 7 A chart that shows the comparison of the proposed and tested models with all bands for study area SA1
Table 6 Classification accuracy
comparison between the
proposed model and the tested
models with optimal bands for
study area SA1
Crop Type Proposed VGG-16 ResNet-50 U-Net
PA UA PA UA PA UA PA UA
Wheat 95,9 89,8 92,1 91,0 87,7 89,0 82,8 87,9
Tomatoe 76,1 72,7 76,9 63,1 75,6 71,1 73,4 64,5
Corn 84,3 86,7 83,7 83,9 80,4 83,1 78,1 82,6
Corn_2 79,6 86,4 78,1 66,8 77,3 70,7 70,2 65,9
Cotton 76,1 80,3 77,3 72,3 84,7 74,7 76,1 69,3
Grapes 82,5 91,9 78,3 93,8 79,7 92,1 79,8 91,3
Alfalfa 86,6 82,2 84,0 81,3 87,0 56,4 78,7 72,5
Olive Trees 95,3 88,6 91,2 95,3 83,7 89,7 91,9 85,6
OA: 86,6 OA: 84,2 OA: 82,4 OA: 80,5
Kappa: 0,842 Kappa: 0,814 Kappa: 0,794 Kappa: 0,772
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 303 Page 14 of 28
Fig. 8 A chart that shows the comparison of the proposed and tested models with the selected optimal bands for study area SA1
Fig. 9 The classification results
for representative subsets
(a1–a4) of several parcels were
achieved respectively by the
proposed (b1-b4), VGG-16
(c1-c4), ResNet-50 (d1 d4), and
U-Net (e1-e4) models, with all
bands for SA1. The parcels with
correct and wrong classification
results were labeled with yellow
and blue markers (e.g. circle,
ellipse, square), respectively
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 Page 15 of 28 303
the parcel shown by a yellow rectangle was correctly identi-
fied as alfalfa by all models.
Results ofstudy area SA2
Table7 shows the training, validation, and test accuracies
for SA2. For both all bands and the selected optimal bands,
the proposed model achieved the highest train, validation
and test accuracies. For all bands, the computed accura-
cies were 91.3%, 87.4%, and 86.1%, respectively for train,
validation, and test. The results obtained by the proposed
model based on the selected bands were also superior to
tested models, with the train, validation and test accuracies
of 91.1%, 86.7%, and 84.7%, respectively. These results
clearly indicate that the proposed model outperforms the
tested VGG-16, ResNet-50, and U-Net models.
The error matrices of the crop maps produced using the
proposed CNN model are given in Table16and Table17in
Appendix 2. A comparison of the classification accuracy
performance of each model based on all bands is shown
in Table8. The proposed model achieved an OA of 88.3%
and a Kappa value of 0.865, demonstrating a significant
improvement over the tested models. VGG-16, ResNet-50,
and U-Net respectively produced OAs of 86.5%, 84.8% and
81.9% along with Kappa values of 0.845, 0.826 and 0.799.
The classification result of the proposed model is observed
to be the best for the sunflower, corn and rice categories with
the PA values of 99.5%, 98.9%, 98.4%, respectively. The
classification performance for wheat, meadow grass, walnut,
triticale and forage pea is also desirable with PA values of
above 86%, and for rapeseed and oats with PA values of
above 78%. The overall performance for the poplar grove,
grape and alfalfa categories is deemed acceptable. Barley
is considered the weakest class in terms of PA (~ 61%), and
according to UA values hungarian vetch provided the poor-
est result of ~ 70%. A comparison of the accuracy values
obtained by the proposed and tested models is presented in
a graph in Fig.10.
A comparison of the classification performance of each
model based on the selected optimal bands is shown in
Table 7 Model set accuracy values of the proposed model and the
tested models for SA2
Model set Variable options
Model type All bands (%) Optimal
bands (%)
Train Proposed 91,3 91,1
VGG-16 89,7 89,3
ResNet-50 87,2 86,8
U-Net 85,9 85,6
Validation Proposed 87,4 86,7
VGG-16 80,8 80,3
ResNet-50 81,2 80,9
U-Net 79,4 79,2
Test Proposed 86,1 84,7
VGG-16 76,7 75,6
ResNet-50 75,5 73,9
U-Net 73,6 72,8
Table 8 Classification accuracy comparison between the proposed model and the tested models with all bands for SA2
Crop Type Proposed VGG-16 ResNet-50 U-Net
PA UA PA UA PA UA PA UA
Barley 60,5 78,5 66,1 59,7 87,4 87,1 81,8 75,9
Sunflower 99,5 99,8 98,3 98,5 96,0 99,9 89,5 95,9
Wheat 94,0 85,5 94,0 95,5 88,4 83,6 67,4 82,3
Meadow Grass 91,5 95,1 95,8 97,1 94,2 91,2 73,9 73,4
Rice 98,4 98,4 93,8 93,8 87,2 81,2 97,1 97,7
Walnut 88,2 87,1 81,4 85,4 79,6 89,8 73,3 89,1
Poplar Grove 72,1 91,3 95,0 94,6 68,8 50,3 95,6 93,3
Rapeseed 80,9 88,0 68,3 49,5 82,8 75,9 77,1 75,9
Hungarian Vetch 73,8 70,3 79,3 95,6 75,8 76,5 68,3 94,4
Corn 98,9 94,5 96,2 98,7 97,5 88,3 87,8 91,3
Triticale 86,8 92,1 84,3 90,9 74,3 73,4 91,1 94,6
Grape 69,7 78,8 84,8 50,9 83,4 76,1 79,4 92,0
Forage Pea 86,7 85,1 86,8 85,1 86,3 79,1 91,6 50,9
Alfalfa 68,3 87,1 78,8 90,4 75,0 81,7 91,7 88,7
Oats 78,0 92,8 60,0 79,5 84,0 92,6 86,1 67,5
OA: 88,3 OA: 86,5 OA: 84,8 OA: 81,9
Kappa: 0,865 Kappa: 0,845 Kappa: 0,826 Kappa: 0,799
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 303 Page 16 of 28
Table9. Similar trends were observed when comparing the
OA and Kappa values; the proposed model demonstrated
significant improvement over the tested models by provid-
ing the highest OA of 86.5% and a Kappa value of 0.847.
VGG-16, ResNet-50, and U-Net, respectively achieved OAs
of 84.7%, 83.1%, and 80.2%, and Kappa values of 0.825,
0.806, and 0.780. The classification achievement of the pro-
posed model is observed to be the best for the corn, rice,
sunflower, wheat and walnut categories with the PA values
above 90%. The classification performance for forage pea,
triticale, grape and barley, rapeseed, and oats categories is
also desireble with the PA values above 79%. The poorest
classification results were found in the poplar grove, hun-
garian vetch, meadow grass, and alfalfa categories with the
PA and UA values of about 70% and 75%, respectively. The
comparison of the accuracy metrics achieved by the pro-
posed and tested models with the selected optimal bands is
presented in a chart in Fig.11.
Fig. 10 A chart that shows the comparison of the proposed and tested models with all bands for SA2
Table 9 Classification accuracy
comparison between the
proposed model and the tested
models with optimal bands for
SA2
Crop Type Proposed VGG-16 ResNet-50 U-Net
PA UA PA UA PA UA PA UA
Barley 80,3 90,8 72,0 68,0 84,7 85,2 75,1 72,5
Sunflower 93,0 99,8 95,5 97,6 94,9 99,9 89,7 96,1
Wheat 92,4 76,7 93,0 84,5 86,7 82,4 69,0 72,0
Meadow Grass 71,0 88,5 95,6 90,1 90,8 87,0 75,0 53,2
Rice 93,5 88,6 72,7 93,3 80,5 78,0 79,1 95,8
Walnut 89,9 84,9 80,1 85,6 75,3 88,4 70,6 84,6
Poplar Grove 74,4 82,9 94,7 94,2 66,3 43,6 93,1 76,7
Rapeseed 79,6 86,8 77,0 56,4 80,0 71,2 77,5 73,2
Hungarian Vetch 74,3 73,4 79,3 95,6 74,3 73,3 59,9 85,5
Corn 98,5 94,1 94,2 93,3 97,1 87,2 93,2 88,2
Triticale 84,5 90,7 72,4 90,8 74,8 68,9 88,9 93,0
Grape 83,4 84,2 84,8 68,8 82,4 60,2 76,3 85,4
Forage Pea 84,7 86,2 84,5 86,1 84,5 78,9 87,0 57,7
Alfalfa 67,1 86,5 78,8 90,4 74,6 84,6 92,4 93,5
Oats 78,8 93,1 79,8 71,1 75,9 94,7 77,3 80,0
OA: 86,5 OA: 84,7 OA: 83,1 OA: 80,2
Kappa: 0,847 Kappa: 0,825 Kappa: 0,806 Kappa: 0,780
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 Page 17 of 28 303
The classification results derived using the optimal bands
showed a slight drop compared to those obtained from uti-
lizing all available bands. With a total of 66 bands initially
available, the optimal set was determined to consist of
29 bands. This result was expected, as a reduction in the
number of bands often results in a proportional decline in
classification accuracy. This comprehensive trend was sys-
tematically observed across the proposed and tested models
executed in SA2.
Figure12 illustrates visual inspections of the classified
maps based on all bands using subset images from SA2.
a1-a4 are false colour image subsets taken from the image
dated 2021–04–12. As can be seen in b1, the proposed
model correctly classified the parcel shown by an ellipse as
hungarian vetch. In contrast, the tested models misclassified
this parcel as forage pea as shown in (c1), (d1) and (e1).
Similarly, the crop type of the parcel shown by a square was
successfully detected by the proposed model as rapeseed,
whereas all tested models failed by mixing up barley, forage
pea and alfalfa, as shown in c1, d1, e1. Hence, the proposed
model stands out due to its success in this scenario.
In the qualitative evaluation of the second row, the crop
type of the parcel shown by a circle was correctly identified
as sunflower by the proposed model (b2), ResNet-50 (d2),
and U-Net (e2). VGG-16 (c2) on the other hand misclassi-
fied this parcel as oats. The crop type of the parcel shown
by a square is corn. The proposed model is the only model
that correctly classified this parcel, as shown in b2. In this
scenario, the tested models (c2, d2 and e2) either mixed the
classes or failed in detecting the crop type. For example,
VGG-16 mixed triticale, forage pea, and alfalfa (c2). On the
other hand, ResNet-50 and U-Net incorrectly identified the
crop type as hungarian vetch (d2) and forage pea (e2).
As it is shown in the third row, while VGG-16 (c3) and
ResNet-50 (d3) successfully classified the parcel shown by a
circle as rapeseed, the proposed model (b3) and U-Net (e3)
failed in this case wrongly classifying the parcel as barley.
However, the proposed model (b3) correctly assigned the
parcels shown by an ellipse and a square to alfalfa and triti-
cale categories. On the other hand, all tested models failed
in classifying these two parcels. Specifically, VGG-16 mixed
the sunflower and triticale categories, ResNet-50 mixed the
oats and triticale categories, and U-Net mixed the triticale
and rapeseed categories, as shown by an ellipse in c3, d3,
and e3. The failure of the tested models is also evident
in the square-marked parcel (c3, d3, and e3), where mis-
classification occurred between barley, rapeseed, triticale,
grape, alfalfa and oats. Hence, as shown with the ellipse
and square-marked parcels, the proposed model (b3) dem-
onstrated superior performance when compared to all tested
models (c3, d3 and e3).
In the last row, the parcel shown by a circle was expected to
be assigned the crop type of triticale. However, the proposed
model (b4) and VGG-16 (c4) achieved wrong labelling as bar-
ley, whereas ResNet-50 (d4) and U-Net (e4) correctly classi-
fied this parcel. For the parcel shown by an ellipse, only the
proposed model (b4) correctly identified the rapeseed crop,
whereas all tested models (c4, d4 and e4) failed to efficiently
classify this parcel. As shown by an ellipse, VGG-16 (c4) and
ResNet-50 (d4) mixed the rapeseed and sunflower categories,
while U-Net (e4) mixed the rapeseed, sunflower, barley, oats
and poplar grove categories. In the parcel shown by a square, a
Fig. 11 A chart that shows the comparison of the proposed and tested models with optimal bands for SA2
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 303 Page 18 of 28
similar trend to that observed in the parcel shown by an ellipse
is noted, as only one of the tested models correctly classified it.
Only U-Net (e4) efficiently identified the crop type of the parcel
shown by a square as barley. In contrast, ResNet-50 (d4) mis-
classified this parcel as triticale, while the proposed model (b4)
and VGG-16 (c4) mixed up the barley and triticale categories.
Discussion
The results obtained in this study indicate that the spatial
information extracted by the proposed CNN model is prom-
ising in parcel-based crop type classification. Among the crop
categories in SA1, wheat, corn and sunflower were the most
efficiently classified crop types. This is due to the fact that the
wheat and corn categories were very divergent classes. How-
ever, tomatoe was mixed with cotton. It is shown in Fig.13
that the NDVI spectral reflectance curves of these two crop
categories significantly overlap demonstrating very similar
spectral reflectances between April and November.
Among the crop categories in SA2, wheat, corn and sun-
flower exhibited significant divergence in the classified output.
Poplar grove overlapped with many classes including, meadow
grass, hungarian vetch and forage pea, as shown in Fig.14,
where significant spectral similarity between the NDVI spectral
reflectance curves of these crop categories is evident.
The distinct classes in both study aeas are evidence that
the crop types exhibit clearly different spectral values across
the time-series dates used. Spectral information for these
agricultural crop classes is crucial for accurate classification
in both study areas. Therefore, the ability of the proposed
model to extract spectral patterns significantly contributed to
the classification task. Figure15 presents a graph that illus-
trates the pattern of the average NDVI spectral reflectance
values of the classified crop types across the time-series
dates used in both study areas. For SA1 (a) and SA2 (b), the
highest average values were observed for alfalfa, with 0.819
and 0.706, respectively, while the lowest average values were
recorded for corn (0.370) and forage pea (0.385).
Fig. 12 The classification
results for representative subsets
(a1–a4) of several parcels were
achieved respectively by the
proposed (b1-b4), VGG-16
(c1-c4), ResNet-50 (d1-d4), and
U-Net (e1-e4) models, with all
bands for SA2. The parcels with
correct and wrong classification
results were labeled with yellow
and blue markers (e.g. circle,
ellipse, square), respectively
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 Page 19 of 28 303
The training sample parcels with the fewest numbers
belong to alfalfa and forage pea categories in SA1 and SA2,
respectively. With respect to the total number of pixels
within the sample parcels, alfalfa has the fewest in SA1.
In SA2, while forage pea has the fewest sample parcels,
poplar grove has the fewest total pixels. This discrepancy
is due to the varying parcel sizes. On the other hand, the
highest number of sample parcels and the total number of
pixels belong to the same crop categories in both study
areas. Specifically, the grape category has the highest num-
ber of sample parcels and total number of pixels in SA1,
while the wheat category holds this distinction in SA2. The
quantitative data were given above in Table2. The smallest
parcel is assigned to alfalfa category in SA1 and to hungar-
ian vetch category in SA2. Conversely, the largest parcels
are associated with the olive trees category in SA1 and the
rapeseed category in SA2.
Accurate classification and feature extraction of sat-
ellite imagery is considered one of the major challenges
within the remote sensing community (Huang and Wang
2006). The proposed model utilizes agricultural fields to
generate sample image patches for the training dataset
and requires padding operation. For the padding process,
compactly shaped parcels require less padding, while the
elongated parcels necessitate more padding. Figure4(b)
shown elongated parcels with relatively more compact
shapes. Although the padding process is performed using
the original pixel values that belong to parcel being con-
sidered (Fig.4(c)), excessive padding would also impact
the algorithm's processing time.
Fig. 13 Change in NDVI ratios of mixed crops according to time-series dates in SA1
Fig. 14 Change in NDVI ratios of mixed crops according to time-series dates in SA2
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 303 Page 20 of 28
Based on both the quantitative and qualitative evaluation
results and across both datasets (all bands and selected optimal
bands), the following assessments can be drawn. CNN deep
learning frameworks demonstrate superior robustness and
effectiveness compared to traditional classification models
(Wang etal. 2021; Alzubaidi etal. 2021; Archana and Jee-
varaj 2024; Ghafari etal. 2021). The proposed CNN model
partially addresses both the robustness and the time-consuming
nature associated with deep learning algorithms, at least for
the present. While discussions about robustness and process-
ing time remain ongoing, it is evident that performing optimal
band selection using the RFECV method can reduce processing
time without significantly affecting the model and classified
map accuracy. The image classification results obtained using
the selected optimal bands through the SVM-RFE technique
exhibit a relative decrease of approximately 1% to 3% compared
to the classification accuracy achieved when employing all
spectral bands. From other perspective, the computational time
is shorter as expected. When both computational time and clas-
sification success are considered together, it is evaluated that
the band combination obtained with the SVM-RFE technique
can be ideal in terms of time-performance for the analyses to be
performed (Dubey and Choubey 2024; Ramezan 2022).
The values of McNemar’s statistical test computed for SA1
are given in Table10. The matrix represents the chi-square
values computed for all pair combinations of classifiers used.
For this study area, across both the all-bands and optimal-
bands variable options, the computed OA values were close.
However, the computed chi-squared values indicate statis-
tically significant differences between the classifier pairs.
The VGG-16 and ResNet-50 pair yielded the lowest value
of 36.47 for the all-bands variable option, while the pair of
the proposed model and U-Net showed the largest signifi-
cant difference, with the highest value of 106.34. Notably, all
model pairs demonstrated significant differences, with the
values exceeding 3.84. The values of McNemar’s statistical
test computed for SA2 are given in Table11. As seen in the
table, all chi-squared values surpassed the threshold for sta-
tistical significance (3.84). The pair of ResNet-50 and U-Net
provided the lowest values for both the all-bands (9.93) and
optimal-bands (4.12) variable options. The proposed model,
when compared to tested other models in pairwise logic, pro-
duced the highest statistical values. Table12 gives the areas
of the crops computed from the classification results of the
proposed method based on all bands of time series images
together with the FRS data.
The VGG-16, ResNet-50, and U-Net models used for
comparison are quite complex, incorporating additional
Fig. 15 Pattern of average NDVI values of crop types for SA1 (a) and SA2 (b)
Table 10 Results of McNemar’s statistical test for SA1
Variable
options
Model type Proposed VGG-16 ResNet-50 U-Net
All bands Proposed × 61,35 82,76 106,34
VGG-16 - × 36,47 73,89
ResNet-50 - - × 40,18
U-Net - - - ×
Optimal
bands
Proposed × 58,24 77,11 102,16
VGG-16 - × 45,06 66,55
ResNet-50 - - × 51,29
U-Net - - - ×
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 Page 21 of 28 303
steps beyond those used in the proposed model, utilizing
a vast number of parameters, and featuring intricate archi-
tectures. Despite its simpler and more barer architecture
(as shown in Fig.5), the proposed model demonstrated
superior performance in comparison with more complex
models that require more parameter values.
In summary, the VGG-16, ResNet-50, and U-Net CNN
architectures are employed for computer vision tasks (Zhao
etal. 2024; Li etal. 2021a, b, c; Bhatt etal. 2021; Hu etal.
2023). They vary in their depth, architectural design, number
of parameters, and utilization of techniques, such as residual
connections and data augmentation. Each architecture pos-
sesses distinct strengths and limitations, and the selection of
a particular model often depends on the specific task require-
ments and available resources. The proposed model is mainly
aimed at classifying and detecting agricultural crops.
Computation time
All experiments were carried out on an Intel Xenon
W-10885M pocessor with 5.3GHz of clock and 32GB
RAM memory and NVIDIA Quadro RTX 5000 Max-Q
graphic card. Table13 shown the approximate computation
time (in second) comparison for the training process carried
out using different number of bands and 100 epochs. It can
be seen that due to its lighter design the proposed model has
significantly lower computational cost when compared to
models tested. Considering that the models were trained with
100 epochs and same optimizer and learning rates, the results
are highly correlated with the model depth. U-net was the
Table 11 Results of McNemar’s statistical test for SA2
Variable
options
Model type Proposed VGG-16 ResNet-50 U-Net
All bands Proposed × 58,46 76,28 97,85
VGG-16 - × 16,54 22,49
ResNet-50 - - × 9,93
U-Net - - - ×
Optimal
bands
Proposed × 31,74 53,88 84,16
VGG-16 - × 14,99 11,27
ResNet-50 - - × 4,12
U-Net - - - ×
Table 12 The crop areas
computed from the
classification results of the
proposed method based on all
bands of time series images in
comparison with the FRS data
Study area Crop type FRS data
area (da)
Classified
area (da)
Difference between FRS
and classified area (da)
Compat-
ibility
(%)
SA1 Wheat 5009,0 4550,7 458,3 90,9
Tomatoe 1928,1 1465,3 462,7 76,0
Corn 3908,3 3415,8 492,4 87,4
Corn_2 1135,9 865,5 270,3 76,2
Cotton 2327,7 1835,4 492,3 78,9
Grapes 5139,2 4424,8 714,3 86,1
Alfalfa 696,0 577,0 119,0 82,9
Olive trees 4371,0 4027,9 343,1 92,2
Total 24515,5 21162,8 3352,7 86,3
SA2 Barley 22765,2 19862,6 2902,5 87,3
Sunflower 106292,8 104113,8 2179,0 98,0
Wheat 117070,6 100680,7 16389,8 86,0
Meadow Grass 3563,4 3303,2 260,1 92,7
Rice 5444,8 4584,5 860,2 84,2
Walnut 12856,3 10889,3 1967,0 84,7
Poplar Grove 575,7 342,8 232,9 59,6
Rapeseed 21726,5 17240,0 4486,5 79,4
Hungarian Vetch 1506,6 1147,3 359,3 76,2
Corn 31112,3 28903,3 2208,9 92,9
Triticale 8723,2 6442,1 2281,1 73,9
Grape 2975,8 2373,2 602,6 79,8
Forage Pea 1308,0 1081,7 226,2 82,7
Alfalfa 5178,4 4057,2 1121,1 78,4
Oats 7773,2 6863,7 909,4 88,3
Total 348873,6 311886,3 36987,3 89,4
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 303 Page 22 of 28
most computationally demanding, as its deeper architecture
ultimately increased the computational cost. As expected, as
the number of features and the number of classes decrease,
accordingly the computation time decreases.
Conclusions
In this study, a light CNN deep learning model combined
with parcel-based image analysis approach was proposed
to extract crop types in agricultural areas from time-series
optical satellite imagery. The experiments were carried
out on two different agricultural areas with a variety of
crops in Türkiye using Sentinel-2 data.
The results showed that the proposed model had a better
crop type classification performance than VGG-16, ResNet-50,
and U-Net, with an increased OA of up to 8% compared to the
results of these models. The evaluation of the impact of the
optimal bands extracted from time-series images on the accu-
racy of crop classification suggests that the selected optimal
bands can be considered to be used in the classification. In this
respect, to make a decision, the analyst should take into account
the trade-off between computational cost and the achieved qual-
ity of crop type detection.
The results achieved in this study are promising for accurate
and cost-effective classification of parcel-level crop types from
time-series Sentinel-2 imagery. We conclude that the proposed
CNN model in conjunction with a parcel-based image analysis
is an effective and efficient method for accurate crop type clas-
sification using time-series Sentinel-2 imagery, and it is suitable
for different types of optical satellite images. However, more
experiments on different areas with wider variety of crop types
are needed to further explore the robustness and effectiveness
of the proposed of method. Furthermore, the use of different
vegetation indices, such as Difference Vegetation Index (DVI),
Radar Vegetation Index (RVI), etc. as additional bands in the
classification can be considered in future studies.
Appendix1: Error matrices computed
fromtheproposed model classication
forstudy area SA1
Explanatory notes:
Reference—from the FRS ground truth dataset, true
Classified—classification layer, prediction
PA- Producer’s Accuracy
UA- User’s Accuracy
OA- Overall Accuracy
Kappa- Kappa coefficient
Class number Class name
1 Wheat
2 Tomatoe
3 Corn
4 Corn_2
5 Cotton
6 Grapes
7 Alfalfa
8 Olive trees
Tables14 and 15.
Table 13 Computation time comparison of the models used
Study area Number
of bands
Processing Time (s)
Proposed VGG-16 ResNet-50 U-Net
SA1 88 8640 23041 33122 49561
41 7547 20128 28934 43,295
SA2 66 14535 38760 54180 73980
29 11474 31932 44636 60949
Table 14 The error matrix of
the crop map produced using
the proposed model with all
bands
Reference Total UA
12345678
1 437 3 4 5 2 0 1 9 461 94,7
Classified 2 0 134 13 0 7 16 0 0 170 78,8
3 0 6 314 0 21 13 0 0 354 88,7
4 11 4 3 144 1 0 1 1 165 87,2
5 1 6 10 0 178 30 0 1 226 78,7
6 0 6 8 0 4 324 0 4 346 93,6
7 4 0 2 5 1 1 66 1 80 82,5
8 8 9 4 2 3 5 2 397 430 92,3
Total 461 168 358 156 217 389 70 413 2232
PA 94,7 79,7 87,7 92,3 82,0 83,2 94,2 96,1
OA 89,3
Kappa 0,874
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 Page 23 of 28 303
Appendix2: Error matrices computed
fromtheproposed model classication
forstudy area SA2
Explanatory notes:
Reference—from the FRS ground truth dataset, true
Classified—classification layer, prediction
PA- Producer’s Accuracy
UA- User’s Accuracy
OA- Overall Accuracy
Kappa- Kappa coefficient
Class number Class name
1 Barley
2 Sunflower
3 Wheat
4 Meadow Grass
5 Rice
6 Walnut
7 Poplar Grove
8 Rapeseed
9 Hungarian Vetch
10 Corn
11 Triticale
12 Grape
13 Forage Pea
14 Alfalfa
15 Oats
Tables16 and 17.
Table 15 The error matrix of
the crop map produced using
the proposed model with the
selected optimal bands
Reference Total UA
12345678
1 426 8 11 11 3 0 2 13 474 89,8
Classified 2 0 131 19 0 12 18 0 0 180 72,7
3 0 9 308 0 24 14 0 0 355 86,7
4 9 5 2 121 1 0 1 1 140 86,4
5 1 7 8 0 176 26 0 1 219 80,3
6 0 9 11 0 5 318 0 3 346 91,9
7 2 0 1 8 1 1 65 1 79 82,2
8 6 3 5 12 9 8 7 389 439 88,6
Total 444 172 365 152 231 385 75 408 2232
PA 95,9 76,1 84,3 79,6 76,1 82,5 86,6 95,3
OA 86,6
Kappa 0,842
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 303 Page 24 of 28
Table 17 The error matrix of the crop map produced using the proposed model with the selected optimal bands
Reference Total UA
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
1 1538 0 43 0 0 12 0 31 0 0 0 0 66 3 0 1693 90,8
Classified 2 0 1641 0 2 0 0 0 0 0 0 0 0 0 0 0 1643 99,8
3 293 0 2180 101 0 62 0 52 9 0 75 0 53 7 7 2839 76,7
4 3 31 0 324 0 2 0 2 0 0 0 0 2 2 0 366 88,5
5 0 21 0 0 187 0 3 0 0 0 0 0 0 0 0 211 88,6
6 41 49 30 12 0 2473 0 21 73 0 22 5 134 33 17 2910 84,9
7 0 4 0 0 13 0 131 0 0 10 0 0 0 0 0 158 82,9
8 7 7 16 0 0 17 0 603 2 3 0 2 22 10 5 694 86,8
9 8 3 14 0 0 38 0 3 403 0 0 2 62 14 2 549 73,4
10 2 8 0 0 0 0 42 2 0 910 0 0 0 3 0 967 94,1
11 3 0 39 0 0 5 0 0 0 0 530 0 7 0 0 584 90,7
12 0 0 0 0 0 12 0 2 2 0 0 96 2 0 0 114 84,2
13 18 0 27 17 0 110 0 34 53 0 0 10 1990 38 10 2307 86,2
14 2 0 2 0 0 12 0 7 0 0 0 0 9 225 3 260 86,5
15 0 0 7 0 0 5 0 0 0 0 0 0 0 0 164 176 93,1
Total 1915 1764 2358 456 200 2748 176 757 542 923 627 115 2347 335 208 15471
PA 80,3 93,0 92,4 71,0 93,5 89,9 74,4 79,6 74,3 98,5 84,5 83,4 84,7 67,1 78,8
OA 86,5
Kappa 0,847
Table 16 The error matrix of the crop map produced using the proposed model with all bands
Reference Total UA
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
1 319 0 22 0 0 12 0 31 0 0 0 0 19 3 0 406 78,5
Classified 2 0 1854 0 2 0 0 0 0 0 0 0 0 0 0 0 1856 99,8
3 120 0 2286 0 0 62 0 52 9 0 75 0 53 7 7 2671 85,5
4 2 7 0 334 0 2 0 2 0 0 0 0 2 2 0 351 95,1
5 0 0 0 0 189 0 3 0 0 0 0 0 0 0 0 192 98,4
6 41 2 24 12 0 3042 0 21 73 0 22 15 187 33 17 3489 87,1
7 0 0 0 0 3 0 137 0 0 10 0 0 0 0 0 150 91,3
8 10 0 21 0 0 17 0 655 2 0 0 2 22 10 5 744 88,0
9 12 0 12 0 0 58 0 3 392 0 0 2 62 14 2 557 70,3
10 2 0 0 0 0 0 50 2 0 980 0 0 0 3 0 1037 94,5
11 2 0 41 0 0 5 0 0 0 0 643 0 7 0 0 698 92,1
12 0 0 0 0 0 12 0 2 2 0 0 67 2 0 0 85 78,8
13 17 0 15 17 0 221 0 34 53 0 0 10 2379 38 10 2794 85,1
14 2 0 2 0 0 12 0 7 0 0 0 0 9 238 3 273 87,1
15 0 0 7 0 0 5 0 0 0 0 0 0 0 0 156 168 92,8
Total 527 1863 2430 365 192 3448 190 809 531 990 740 96 2742 348 200 15471
PA 60,5 99,5 94,0 91,5 98,4 88,2 72,1 80,9 73,8 98,9 86,8 69,7 86,7 68,3 78,0
OA 88,3
Kappa 0,865
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 Page 25 of 28 303
Acknowledgements The authors would like to thank the The European
Space Agency (ESA) for providing satellite images free of charge and
Republic of Türkiye Ministry of Agriculture and Forestry for providing
Farm Registration System data of the study areas. The authors are thank-
ful to anonymous reviewers for their valuable contribution to the improve-
ment of the paper with their comments and suggestions. This study was
completed as a part of PhD dissertation studies of the first author.
Author contribution Muslum Altun and Mustafa Turker contributed to
the study conception and design. Experiments were performed by Mus-
lum Altun. The first draft of the manuscript was written by Muslum
Altun and Mustafa Turker, and both authors commented on previous
versions of the manuscript. Both authors read and approved the final
manuscript.
Funding Open access funding provided by the Scientific and Tech-
nological Research Council of Türkiye (TÜBİTAK). No financial
resources were used for this article.
Data availability No datasets were generated or analysed during the
current study.
Declarations
Conflicts of interest/Competing interests M. Altun and M. Turker de-
clare that they have no competing interests.
Open Access This article is licensed under a Creative Commons Attri-
bution 4.0 International License, which permits use, sharing, adapta-
tion, distribution and reproduction in any medium or format, as long
as you give appropriate credit to the original author(s) and the source,
provide a link to the Creative Commons licence, and indicate if changes
were made. The images or other third party material in this article are
included in the article’s Creative Commons licence, unless indicated
otherwise in a credit line to the material. If material is not included in
the article’s Creative Commons licence and your intended use is not
permitted by statutory regulation or exceeds the permitted use, you will
need to obtain permission directly from the copyright holder. To view a
copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
References
Abdel-Rahman EM, Mutanga O, Adam E, Ismail R (2014) Detect-
ing sirex noctilio grey-attacked and lightning-struck pine trees
using airborne hyperspectral data, random forest and support
vector machines classifiers. ISPRS J Photogramm Remote Sens
88:48–59. https:// doi. org/ 10. 1016/j. isprs jprs. 2013. 11. 013
Adrian J, Sagan V, Maimaitijiang M (2021) Sentinel sar-optical
fusion for crop type mapping using deep learning and google
earth engine. ISPRS J Photogramm Remote Sens 175:215–235.
https:// doi. org/ 10. 1016/j. isprs jprs. 2021. 02. 018
Altun M, Turker M (2022a) Integration of sentinel-1 and landsat-8
images for crop detection: the case study of Manisa, Tur-
key.Adv Remote Sensing2(1):23–33. https:// publi sh. mersin.
edu. tr/ index. php/ arsej/ artic le/ view/ 322
Altun M, Türker M (2022b) Kaynaştırılmış sentinel-1 sar ve land-
sat-8 optik veriden makine öğrenme algoritması ile tarımsal
ürün tespiti. Türk Uzaktan Algılama ve CBS Dergisi 3(1):1–
19. https:// doi. org/ 10. 48123/ rsgis. 999749
Alzubaidi L, Zhang J, Humaidi AJ, Al-Dujaili A, Duan Y, Al-
Shamma O, Santamaria J, Fadhel MA, Al-Amidie M, Farhan L
(2021) Review of deep learning: concepts, CNN architectures,
challenges, applications, future directions. J Big Data 8:53.
https:// doi. org/ 10. 1186/ s40537- 021- 00444-8
Archana R, Jeevaraj PE (2024) Deep learning models for digital
image processing: A review. Artif Intell Rev 57:11. https:// doi.
org/ 10. 1007/ s10462- 023- 10631-z
Bargiel D (2017) A new method for crop classification combining
time series of radar images and crop phenology information.
Remote Sens Environ 198:369–383. https:// doi. org/ 10. 1016/j.
rse. 2017. 06. 022
Bhatt D, Patel C, Talsania H, Patel J, Vaghela R, Pandya S, Modi K,
Ghayvat H (2021) CNN variants for computer vision: history,
architecture, application, challenges and future scope. Electronics
10(20):2470. https:// doi. org/ 10. 3390/ elect ronic s1020 2470
Bhosle K, Musande V (2022) Evaluation of CNN model by compar-
ing with convolutional autoencoder and deep neural network
for crop classification on hyperspectral imagery. Geocarto Int
37(3):813–827. https:// doi. org/ 10. 1080/ 10106 049. 2020. 17409 50
Blickensdörfer L, Schwieder M, Pflugmacher D, Nendel C, Erasmi
S, Hostert P (2022) Mapping of crop types and crop sequences
with combined time series of Sentinel-1, Sentinel-2 and Land-
sat 8 data for Germany. Remote Sens Environ 269:112831.
https:// doi. org/ 10. 1016/j. rse. 2021. 112831
Chand S (2021) Multiclass and multilabel classification of human cell
components using transfer learning of inceptionV3 model. In2021
International Conference on Computing, Communication, and
Intelligent Systems (ICCCIS), Greater Noida, February 19–20th,
India. https:// doi. org/ 10. 1109/ ICCCI S51004. 2021. 93971 65.
Chandrashekar G, Sahin F (2014) A survey on feature selection
methods. Comput Electr Eng 40(1):16–28. https:// doi. org/ 10.
1016/j. compe leceng. 2013. 11. 024
Charisis C, Argyropoulos D (2024) Deep learning-based instance
segmentation architectures in agriculture: A review of the
scopes and challenges. Smart Agricultural Technol 8:100448.
https:// doi. org/ 10. 1016/j. atech. 2024. 100448
Chavez PS, Sides SC, Anderson JA (1991) Comparison of three
different methods to merge multiresolution and multispectral
data: landsat TM and spot panchromatic. Photogrammetric Eng
Remote Sensing 57(3):295–303
Congalton RG (1991) A review of assessing the accuracy of clas-
sifications of remotely sensed data. Remote Sens Environ
37(1):35–46. https:// doi. org/ 10. 1016/ 0034- 4257(91) 90048-B
Congalton RG (2001) Accuracy assessment and validation of remotely
sensed and other spatial information. Int J Wildland Fire
10(4):321–328. https:// doi. org/ 10. 1071/ WF010 31
De Leeuw J, Jia H, Yang L, Liu X, Schmidt K, Skidmore AK (2006)
Comparing accuracy assessments to infer superiority of image
classification methods. Int J Remote Sens 27(1):223–232. https://
doi. org/ 10. 1080/ 01431 16050 02757 62
Defourny P, Bontemps S, Bellemans N, Cara C, Dedieu G, Guzzonato
E, Hagolle O, Inglada J, Nicola L, Rabaute T, Savinaud M,
Udroiu C, Valero S, Bégué A, Dejoux JF, Harti AE, Ezzahar
J, Kussul N, Labbassi K, Lebourgeois V, Miao Z, Newby T,
Nyamugama A, Salh N, Shelestov A, Simonneaux V, Traore PS,
Traore SS, Koetz B (2019) Near real-time agriculture monitoring
at national scale at parcel resolution: Performance assessment
of the Sen2-Agri automated system in various cropping systems
around the world. Remote Sens Environ 221:551–568. https://
doi. org/ 10. 1016/j. rse. 2018. 11. 007
Dingle Robertson L, King DJ (2011) Comparison of pixel- and object-
based classification in land cover change mapping. Int J Remote Sens
32(6):1505–1529. https:// doi. org/ 10. 1080/ 01431 16090 35717 91
Dubey RK, Choubey DK (2024) An efficient adaptive feature selection
with deep learning model-based paddy plant leaf disease clas-
sification. Multimed Tools Appl 83(8):22639–22661. https:// doi.
org/ 10. 1007/ s11042- 023- 16247-3
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 303 Page 26 of 28
ESA (2024) “Copernicus Dataspace Browser”. https:// brows er. datas
pace. coper nicus. eu/. Accessed 24 May 2024
Fan X, Yan C, Fan J, Wang N (2022) Improved U-net remote sensing
classification algorithm fusing attention and multiscale features.
Remote Sensing 14(15):3591. https:// doi. org/ 10. 3390/ rs141 53591
Fathipoor H, Shah-Hosseini R, Arefi H (2023) Crop and weed seg-
mentation on ground-based images using deep convolutional
neural network. ISPRS Ann Photogrammetry, Remote Sensing
Spatial Information Sci 10:195–200. https:// doi. org/ 10. 5194/
isprs- annals- X-4- W1- 2022- 195- 2023
Foody GM (2004) Thematic map comparison: evaluating the statisti-
cal significance of differences in classification accuracy. Photo-
grammetric Eng Remote Sens 70(5):627–633. https:// doi. org/ 10.
14358/ PERS. 70.5. 627
Furtuna R, Curteanu S, Cazacu M (2011) Optimization methodology
applied to feed-forward artificial neural network parameters. Int J
Quantum Chem 111(3):539–553. https:// doi. org/ 10. 1002/ qua. 22423
Gallo I, Ranghetti L, Landro N, La Grassa R, Boschetti M (2023) In-
season and dynamic crop mapping using 3D convolution neu-
ral networks and sentinel-2 time series. ISPRS J Photogramm
Remote Sens 195:335–352. https:// doi. org/ 10. 1016/j. isprs jprs.
2022. 12. 005
Gangkofner UG, Pradhan PS, Holcomb DW (2008) Optimizing the
high-pass filter addition technique for image fusion. Photogram-
metric Eng Remote Sensing 74(9):1107–1118. https:// doi. org/
10. 14358/ PERS. 74.9. 1107
Ghafari S, Tarnik MG, Yazdi HS (2021) Robustness of convolutional
neural network models in hyperspectral noisy datasets with loss
functions. Comput Electr Eng 90:107009. https:// doi. org/ 10.
1016/j. compe leceng. 2021. 107009
Gill KS, Anand V, Chauhan R, Kapruwan A, Hsiung PA (2023)
Hypothesis Classification of Weather on VGG19 CNN Model
Fine-Tuned with the Adam Optimizer. In2023 3rd International
Conference on Smart Generation Computing, Communication
and Networking (SMART GENCON), December 29–31st, Ban-
galore, India. https:// doi. org/ 10. 1109/ SMART GECON 60755.
2023. 10442 963
Grabska E, Frantz D, Ostapowicz K (2020) Evaluation of machine learning
algorithms for forest stand species mapping using Sentinel-2 imagery
and environmental data in the polish carpathians. Remote Sens Envi-
ron 251:112103. https:// doi. org/ 10. 1016/j. rse. 2020. 112103
Han H, Liu Z, Li J, Zeng Z (2024) Challenges in remote sensing
based climate and crop monitoring: navigating the complexities
using AI. J Cloud Computing 13(1):34. https:// doi. org/ 10. 1186/
s13677- 023- 00583-8
He K, Zhang X, Ren S, Sun J (2016) Deep residual learning for image
recognition. In2016 IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), Las Vegas, June 27–30th, USA.
https:// doi. org/ 10. 48550/ arXiv. 1512. 03385.
Hu WJ, Fan J, Du YX, Li BS, Xiong N, Bekkering E (2020) MDFC–
ResNet: An agricultural IoT system to accurately recognize crop
diseases. IEEE Access 8:115287–115298. https:// doi. org/ 10.
1109/ ACCESS. 2020. 30012 37
Hu K, Jin J, Zheng F, Weng L, Ding Y (2023) Overview of behavior
recognition based on deep learning. Artif Intell Rev 56(3):1833–
1865. https:// doi. org/ 10. 1007/ s10462- 022- 10210-8
Huang CL, Wang CJ (2006) A GA-based feature selection and param-
eters optimization for support vector machines. Expert Syst Appl
31(2):231–240. https:// doi. org/ 10. 1016/j. eswa. 2005. 09. 024
Inglada J, Arias M, Tardy B, Hagolle O, Valero S, Morin D, Ded-
ieu G, Sepulcre G, Bontemps S, Defourny P, Koetz B (2015)
Assessment of an operational system for crop type map produc-
tion using high temporal and spatial resolution satellite optical
imagery. Remote Sensing 7(9):12356–12379. https:// doi. org/ 10.
3390/ rs709 12356
Jensen JR, Lulla K (1987) Introductory digital image processing: A
remote sensing perspective. Prentice Hall Inc, USA
Jin B, Ye P, Zhang X, Song W, Li S (2019) Object-oriented method
combined with deep convolutional neural networks for land-use-
type classification of remote sensing images. J Indian Soc Remote
Sens 47(6):951–965. https:// doi. org/ 10. 1007/ s12524- 019- 00945-3
Kganyago M, Mhangara P, Alexandridis T, Laneve G, Ovakoglou G,
Mashiyi N (2020) Validation of sentinel-2 leaf area index (LAI)
product derived from SNAP toolbox and its comparison with
global LAI products in an African semi-arid agricultural land-
scape. Remote Sens Lett 11(10):883–892. https:// doi. org/ 10.
1080/ 21507 04X. 2020. 17678 23
Khanzode KCA, Ravindra DS (2020) Advantages and disadvantages of
artificial intelligence and machine learning: A literature review.
Int J Library Information Sci (IJLIS) 9(1):30–36. https:// doi. org/
10. 17605/ OSF. IO/ GV5T4
Kussul N, Lavreniuk M, Skakun S, Shelestov A (2017) Deep learning
classification of land cover and crop types using remote sensing
data. IEEE Geosci Remote Sens Lett 14(5):778–782. https:// doi.
org/ 10. 1109/ LGRS. 2017. 26811 28
Lee H, Kwon H (2017) Going Deeper with Contextual CNN for
Hyperspectral Image Classification. IEEE Trans Image Process
26(10):4843–4855. https:// doi. org/ 10. 1109/ TIP. 2017. 272558
Li X, Chen G, Liu J, Chen W, Cheng X, Liao Y (2017) Effects of
RapidEye imagery’s red-edge band and vegetation indices on
land cover classification in an arid region. Chin Geogra Sci
27(5):827–835. https:// doi. org/ 10. 1007/ s11769- 017- 0894-6
Li H, Zhang C, Zhang Y, Zhang S, Ding X, Atkinson PM (2021a)
A scale sequence object-based convolutional neural network
(SS-OCNN) for crop classification from fine spatial resolution
remotely sensed imagery. Int J Digital Earth 14(11):1528–1546.
https:// doi. org/ 10. 1080/ 17538 947. 2021. 19508 53
Li Q, Wong FKK, Fung T (2021b) Mapping multi-layered mangroves
from multispectral, hyperspectral, and LiDAR data. Remote Sens
Environ 258:112403. https:// doi. org/ 10. 1016/j. rse. 2021. 112403
Li Z, Liu F, Yang W, Peng S, Zhou J (2021c) A survey of convo-
lutional neural networks: analysis, applications, and prospects.
IEEE Trans Neural Networks Learning Syst 33(12):6999–7019.
https:// doi. org/ 10. 1109/ TNNLS. 2021. 30848 27
Liu S, Shi Q (2020) Local climate zone mapping as remote sensing
scene classification using deep learning: A case study of metro-
politan China. ISPRS J Photogramm Remote Sens 164:229–242.
https:// doi. org/ 10. 1016/j. isprs jprs. 2020. 04. 008
Liu X, Kang C, Gong L, Liu Y (2016) Incorporating spatial interaction
patterns in classifying and understanding urban land use. Int J
Geographic Information Sci 30(2):334–350. https:// doi. org/ 10.
1080/ 13658 816. 2015. 10869 23
Liu Z, Su B, Lv F (2022) Intelligent identification method of crop species
using improved U-Net network in UAV remote sensing image”.
Sci Program 9717843:1–9. https:// doi. org/ 10. 1155/ 2022/ 97178 43
Ma L, Li M, Gao Y, Chen T, Ma X, Qu L (2017) A novel wrapper
approach for feature selection in object-based image classification
using polygon-based cross-validation. IEEE Geosci Remote Sens
Lett 14(3):409–413. https:// doi. org/ 10. 1109/ LGRS. 2016. 26457 10
Mallinis G, Galidaki G, Gitas I (2014) A comparative analysis of EO-1
hyperion, Quickbird and Landsat TM imagery for fuel type
mapping of a typical mediterranean landscape. Remote Sensing
6(2):1684–1704. https:// doi. org/ 10. 3390/ rs602 1684
Martins VS, Kaleita AL, Gelder BK, Silveira HLD, Abe CA (2020)
Exploring multiscale object-based convolutional neural network
(Multi-OCNN) for remote sensing image classification at high
spatial resolution. ISPRS J Photogramm Remote Sens 168:56–
73. https:// doi. org/ 10. 1016/j. isprs jprs. 2020. 08. 004
Mazzia V, Khaliq A, Chiaberge M (2020) Improvement in land cover
and crop classification based on temporal features learning from
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 Page 27 of 28 303
Sentinel-2 data using recurrent-convolutional neural network
(R-CNN). Appl Sci 10(1):238. https:// doi. org/ 10. 3390/ app10 010238
McNemar Q (1947) Note on the sampling error of the difference
between correlated proportions or percentages. Psychometrika
12(2):153–157. https:// doi. org/ 10. 1007/ BF022 95996
Mei S, Ji J, Hou J, Li X, Du Q (2017) Learning sensor-specific spatial-
spectral features of hyperspectral images via convolutional neural
networks. IEEE Trans Geosci Remote Sens 55(8):4520–4533.
https:// doi. org/ 10. 1016/j. isprs jprs. 2020. 04. 008
Moysiadis V, Tsolakis N, Katikaridis D, Sørensen CG, Pearson S,
Bochtis D (2020) Mobile Robotics in Agricultural Operations:
A Narrative Review on Planning Aspects. Appl Sci 10(10):3453.
https:// doi. org/ 10. 3390/ app10 103453
Nayak JG, Patil LG, Patki VK (2023) Artificial neural network based
water quality index (WQI) for river Godavari (India). Materials
Today Proceedings 81:212–220. https:// doi. org/ 10. 1016/j. matpr.
2021. 03. 100
Niemeyer J, Rottensteiner F, Soergel U (2014) Contextual classifica-
tion of lidar data and building object detection in urban areas.
ISPRS J Photogramm Remote Sens 87:152–165. https:// doi. org/
10. 1016/j. isprs jprs. 2013. 11. 001
Nowakowski A, Mrziglod J, Spiller D, Bonifacio R, Ferrari I, Mathieu
PP, Herranz MG, Kim DH (2021) Crop type mapping by using
transfer learning. Int J Appl Earth Obs Geoinf 98:102313. https://
doi. org/ 10. 1016/j. jag. 2021. 102313
Omer G, Mutanga O, Rahman EMA, Adam E (2015) Exploring the
utility of the additional WorldView-2 bands and support vector
machines in mapping land use/land cover in a fragmented eco-
system. South Africa South African J Geomatics 4(4):414–433.
https:// doi. org/ 10. 4314/ sajg. v4i4.5
Omia E, Bae H, Park E, Kim MS, Baek I, Kabenge I, Cho BK (2023)
Remote Sensing in Field Crop Monitoring: A Comprehensive
Review of Sensor Systems, Data Analyses and Recent Advances.
Remote Sensing 15(2):354. https:// doi. org/ 10. 3390/ rs150 20354
Othman E, Bazi Y, Alajlan N, Alhichri H, Melgani F (2016) Using con-
volutional features and a sparse autoencoder for land-use scene
classification. Int J Remote Sens 37(10):2149–2167. https:// doi.
org/ 10. 1080/ 01431 161. 2016. 11719 28
Pandey PC, Pandey M (2023) Highlighting the role of agriculture and
geospatial technology in food security and sustainable develop-
ment goals. Sustain Dev 31(5):3175–3195. https:// doi. org/ 10.
1002/ sd. 2600
Patel J, Ruparelia A, Tanwar S, Alqahtani F, Tolba A, Sharma R,
Raboaca MS, Neagu BC (2023) Deep learning-based model
for detection of brinjal weed in the era of precision agriculture.
Computers, Mater Continua 77(1):1281–1301. https:// doi. org/
10. 32604/ cmc. 2023. 038796
Petropoulos GP, Kalaitzidis C, Vadrevu KP (2012) Support vector
machines and object-based classification for obtaining land-use/
cover cartography from hyperion hyperspectral imagery. Comput
Geosci 41:99–107. https:// doi. org/ 10. 1016/j. cageo. 2011. 08. 019
Phung VH, Rhee EJ (2019) “A High-Accuracy Model Average Ensem-
ble of Convolutional Neural Networks for Classification of Cloud
Image Patches on Small Datasets. Appl Sci 9(21):4500. https://
doi. org/ 10. 3390/ app92 14500
Poojary R, Pai A (2019) Comparative study of model optimization
techniques in fine-tuned CNN models. In2019 International
Conference on Electrical and Computing Technologies and
Applications ICECTA, November 19–21st, Ras Al Khaimah,
United Arab Emirates. https:// doi. org/ 10. 1109/ ICECT A48151.
2019. 89596 81
Qiu C, Tong X, Schmitt M, Bechtel B, Zhu XX (2020) Multilevel
feature fusion-based CNN for local climate zone classifica-
tion from Sentinel-2 images: benchmark results on the So2Sat
LCZ42 dataset. IEEE J Selected Topics Appl Earth Observations
Remote Sens 13:2793–2806. https:// doi. org/ 10. 1109/ JSTARS.
2020. 29957 11
Ramezan CA (2022) Transferability of Recursive Feature Elimination
(RFE)-Derived Feature Sets for Support Vector Machine Land
Cover Classification. Remote Sens 14(24):6218. https:// doi. org/
10. 3390/ rs142 46218
Ramezan CA, Warner TA, Maxwell AE (2019) Evaluation of sampling
and cross-validation tuning strategies for regional-scale machine
learning classification. Remote Sens 11(2):185. https:// doi. org/
10. 3390/ rs110 20185
Rani PAS, Singh NS (2022) Paddy leaf symptom-based disease clas-
sification using deep CNN with ResNet-50. Int J Adv Sci Com-
puting Eng 4(2):88–94. https:// doi. org/ 10. 62527/ ijasce. 4.2. 83
Ronneberger O, Fischer P, Brox T (2015) U-net: convolutional net-
works for biomedical image segmentation. In 18th International
Conferenceon Medical Image Computing and Computer-
Assisted Intervention MICCAI, Munich, October 5–9th, Ger-
many. https:// doi. org/ 10. 48550/ arXiv. 1505. 04597.
RTMAF (2024) General directorate of agricultural reform.https://
www. tarim orman. gov. tr/ TRGM/ Sayfa lar/ EN/ AnaSa yfa.
aspx.Accessed 24May 2024
Sagan V, Maimaitijiang M, Bhadra S, Maimaitiyiming M, Browna DR,
Sidikee P, Fritschif FB (2021) Field-scale crop yield prediction
using multi-temporal WorldView-3 and PlanetScope satellite
data and deep learning. ISPRS J Photogramm Remote Sens
174:265–281. https:// doi. org/ 10. 1016/j. isprs jprs. 2021. 02. 008
Sahu M, Dash R (2024) Cognitive land cover mapping: A three-layer deep
learning architecture for remote sensing data classification. Environ
Challenges 15:100876. https:// doi. org/ 10. 1016/j. envc. 2024. 100876
Sargent I, Hare J, Young D, Wilson O, dge C, Holland D, Atkinson
PM, (2017) Inference and discovery in remote sensing data with
features extracted using deep networks. In AI-2017 Thirty-Sev-
enth SGAI International Conference on Artificial Intelligence.
December 12–14th, Cambridge, United Kingdom. https:// doi.
org/ 10. 1007/ 978-3- 319- 71078-5_ 10.
Schultz B, Immitzer M, Formaggio AR, Sanches IDA, Luiz AJB, Atz-
berger C (2015) Self-guided segmentation and classification of
multi-temporal landsat 8 images for crop type mapping in South-
eastern Brazil. Remote Sens 7(11):14482–14508. https:// doi. org/
10. 3390/ rs711 14482
Schutt P, Rosu RA, Behnke S (2022). Abstract flow for temporal seman-
tic segmentation on the permutohedral lattice. In2022 Interna-
tional Conference on Robotics and Automation (ICRA).May
23–27th, Philadelphia, PA, USA. https:// doi. org/ 10. 1109/ ICRA4
6639. 2022. 98118 18
Shah SR, Qadri S, Bibi H, Shah SMW, Sharif MI, Marinello F (2023)
Comparing inception V3, VGG 16, VGG 19, CNN, and ResNet
50: A case study on early detection of a rice disease”. Agronomy
13(6):1633. https:// doi. org/ 10. 3390/ agron omy13 061633
Shah D, Xue ZY, Aamodt TM (2022) Label encoding for regression
networks. In: 10th International conference on learning represen-
tations, April 25–29th, Online. https:// arxiv. org/ abs/ 2212. 01927.
Accessed 10 Nov 2024
Shahade AK, Walse KH, Thakare VM, Atique M (2023) Multi-lingual
opinion mining for social media discourses: An approach using
deep learning based hybrid fine-tuned smith algorithm with adam
optimizer. Int J Information Manag Data Insights 3(2):100182.
https:// doi. org/ 10. 1016/j. jjimei. 2023. 100182
Shi X, Zhang Y, Liu K, Wen Z, Wang W, Zhang T, Li J (2025) State
space models meet transformers for hyperspectral image clas-
sification. Signal Process 226:109669. https:// doi. org/ 10. 1016/j.
sigpro. 2024. 109669
Siachalou S, Mallinis G, Strati MT (2015) A hidden markov mod-
els approach for crop classification: Linking crop phenology to
time series of multi-sensor remote sensing data. Remote Sens
7(4):3633–3650. https:// doi. org/ 10. 3390/ rs704 03633
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Earth Science Informatics (2025) 18:303 303 Page 28 of 28
Simon P, Uma V (2020) Deep learning based feature extraction for
texture classification. Procedia Computer Sci 171:1680–1687.
https:// doi. org/ 10. 1016/j. procs. 2020. 04. 180
Simonyan K, Zisserman A (2015) Very deep convolutional networks
for large-scale image recognition.3rd International Conference
on Learning Representations, San Diego, May 7–9th, USA,
https:// doi. org/ 10. 48550/ arXiv. 1409. 1556.
Singh V, Chug A, Singh AP (2023) Classification of beans leaf diseases
using fine tuned cnn model. Procedia Computer Sci 218:348–
356. https:// doi. org/ 10. 1016/j. procs. 2023. 01. 017
Sola IT, García AM, Pozo LS, Álvarez JM, González MAA (2018)
Assessment of atmospheric correction methods for Sentinel-2
images in Mediterranean landscapes. Int J Appl Earth Obs Geo-
inf 73:63–76. https:// doi. org/ 10. 1016/j. jag. 2018. 05. 020
Spoto F, Sy O, Laberinti P, Martimort P, Fernandez V, Colin O, Hoersch B,
Meygret A (2012) Overview of Sentinel-2. In2012 IEEE International
Geoscience and Remote Sensing Symposium, Munich, July 22-27th,
Germany. https:// doi. org/ 10. 1109/ IGARSS. 2012. 63511 95.
Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R
(2014) Dropout: A simple way to prevent neural networks from
overfitting.J Mach Learn Res15(1):1929–1958. https:// api.
seman ticsc holar. org/ Corpu sID: 68444 31. Accessed 10 Nov 2014
Sun C, Bian Y, Zhou T, Pan J (2019) Using of multi-source and multi-
temporal remote sensing data improves crop-type mapping in the
subtropical agriculture region. Sensors 19(10):2401. https:// doi.
org/ 10. 3390/ s1910 2401
Thakur PS, Sheorey T, Ojha A (2023) VGG-ICNN: A lightweight
CNN model for crop disease identification. Multimed Tools Appl
82(1):497–520. https:// doi. org/ 10. 1007/ s11042- 022- 13144-z
Thenmozhi K, Reddy US (2019) Crop pest classification based on
deep convolutional neural network and transfer learning. Comput
Electron Agric 164:104906. https:// doi. org/ 10. 1016/j. compag.
2019. 104906
Tian Y, Cao X, Zhang T, Wu H, Zhao C, Zhao Y (2024) CabbageNet:
Deep Learning for High-Precision Cabbage Segmentation in
Complex Settings for Autonomous Harvesting Robotics. Sen-
sors 24(24):8115. https:// doi. org/ 10. 3390/ s2424 8115
Tucker CJ (1979) Red and photographic infrared linear combinations
for monitoring vegetation. Remote Sens Environ 8(2):127–150.
https:// doi. org/ 10. 1016/ 0034- 4257(79) 90013-0
Valero S, Morin D, Inglada J, Sepulcre G, Arias M, Hagolle O, Dedieu
G, Bontemps S, Defourny P, Koetz B (2016) Production of a
dynamic cropland mask by processing remote sensing image
series at high temporal and spatial resolutions. Remote Sens
8(1):55. https:// doi. org/ 10. 3390/ rs801 0055
Van TGN, McVicar TR (2004) Determining temporal windows for crop
discrimination with remote sensing: A case study in South-East-
ern Australia. Comput Electron Agric 45(1–3):91–108. https://
doi. org/ 10. 1016/j. compag. 2004. 06. 003
Wang Q, Shi W, Li Z, Atkinson PM (2016) Fusion of Sentinel-2
images. Remote Sens Environ 187:241–252. https:// doi. org/ 10.
1016/j. rse. 2016. 10. 030
Wang P, Fan E, Wang P (2021) Comparative analysis of image clas-
sification algorithms based on traditional machine learning and
deep learning. Pattern Recogn Lett 141:61–67. https:// doi. org/
10. 1016/j. patrec. 2020. 07. 042
Wang S, Han W, Huang X, Zhang X, Wang L, Li J (2024) Trustwor-
thy remote sensing interpretation: Concepts, technologies, and
applications. ISPRS J Photogramm Remote Sens 209:150–172.
https:// doi. org/ 10. 1016/j. isprs jprs. 2024. 02. 003
Wei S, Zhang H, Wang C, Wang Y, Xu L (2019) Multi-temporal SAR
data large-scale crop mapping based on U-Net model. Remote
Sens 11(1):68. https:// doi. org/ 10. 3390/ rs110 10068
Wojciuk M, Swiderska CZ, Siwek K, Gertych A (2024) Improving
classification accuracy of fine-tuned CNN models: Impact of
hyperparameter optimization. Heliyon 10(5):26465. https:// doi.
org/ 10. 1016/j. heliy on. 2024. e26586
Xu W, Sun L, Zhen C, Liu B, Yang Z, Yang W (2022) Deep learn-
ing-based image recognition of agricultural pests. Appl Sci
12(24):12896. https:// doi. org/ 10. 3390/ app12 24128 96
Xu C, Li B, Kong F, He T (2024) Spatial-temporal variation, driv-
ing mechanism and management zoning of ecological resil-
ience based on RSEI in a coastal metropolitan area. Ecol Ind
158:111447. https:// doi. org/ 10. 1016/j. ecoli nd. 2023. 111447
Yan Y, Ryu Y (2021) Exploring google street view with deep learn-
ing for crop type mapping. ISPRS J Photogramm Remote Sens
171:278–296. https:// doi. org/ 10. 1016/j. isprs jprs. 2020. 11. 022
Yan G, Mas JF, Maathuis BHP, Xiangmin Z, Van PMD (2006) Com-
parison of pixel-based and object-oriented image classification
approaches - A case study in a coal fire area, Wuda, Inner Mon-
golia China. Int J Remote Sens 27(18):4039–4055. https:// doi.
org/ 10. 1080/ 01431 16060 07026 32
Yang S, Gu L, Li X, Jiang T, Ren R (2020) Crop classification method
based on optimal feature selection and hybrid CNN-RF net-
works for multi-temporal remote sensing imagery. Remote Sens
12(19):3119. https:// doi. org/ 10. 3390/ rs121 93119
Yousaf J, Yakoub S, Karkanawi S, Hassan T, Almajali E, Zia H, Ghazal
M (2024) Through-the-wall human activity recognition using
radar technologies: A review. IEEE Open J Antennas Propagation
5(6):1815–1837. https:// doi. org/ 10. 1109/ OJAP. 2024. 34590 45
Zhang C, Sargent I, Pan X, Li H, Gardiner A, Hare J, Atkinson PM
(2018a) An object-based convolutional neural network (OCNN)
for urban land use classification. Remote Sens Environ 216:57–
70. https:// doi. org/ 10. 1016/j. rse. 2018. 06. 034
Zhang M, Li W, Du Q (2018b) Diverse region-based CNN for
hyperspectral image classification. IEEE Trans Image Process
27(6):2623–2634. https:// doi. org/ 10. 1109/ TIP. 2018. 28096 06
Zhao B, Zhong Y, Zhang L (2016) A Spectral-structural bag-of-features
scene classifier for very high spatial resolution remote sensing
imagery. ISPRS J Photogramm Remote Sens 116:73–85. https://
doi. org/ 10. 1016/j. isprs jprs. 2016. 03. 004
Zhao W, Du S, Emery WJ (2017a) Object-based convolutional neural
network for high-resolution imagery classification. IEEE J Selected
Topics Appl Earth Observations Remote Sens 10(7):3386–3396.
https:// doi. org/ 10. 1109/ JSTARS. 2017. 26803 24
Zhao W, Du S, Wang Q, Emery WJ (2017b) Contextually guided very-
high resolution imagery classification with semantic segments.
ISPRS J Photogramm Remote Sens 132:48–60. https:// doi. org/
10. 1016/j. isprs jprs. 2017. 08. 011
Zhao H, Chen Z, Jiang H, Jing W, Sun L, Feng M (2019) Evalua-
tion of three deep learning models for early crop classifica-
tion using Sentinel-1A imagery time series - A case study in
Zhanjiang China. Remote Sens 11(22):2673. https:// doi. org/
10. 3390/ rs112 22673
Zhao X, Wang L, Zhang Y, Han X, Deveci M, Parmar M (2024) A
review of convolutional neural networks in computer vision. Artif
Intell Rev 57(4):99. https:// doi. org/ 10. 1007/ s10462- 024- 10721-6
Zheng H, Du P, Chen J, Xia J, Li E, Xu Z, Li X, Yokoya N (2017)
Performance evaluation of downscaling Sentinel-2 imagery for
land use and land cover classification by spectral-spatial features.
Remote Sens 9(12):1274. https:// doi. org/ 10. 3390/ rs912 1274
Zhou Y, Wei T, Zhu X, Collin M (2021) A parcel-based deep-learning
classification to map local climate zones from Sentinel-2 images.
IEEE J Selected Topics Appl Earth Observations Remote Sens
14:4194–4204. https:// doi. org/ 10. 1109/ JSTARS. 2021. 30715 77
Zhu XX, Tuia D, Mou L, Xia GS, Fraundorfer F (2017) Deep learning
in remote sensing: A review. IEEE Geosci Remote Sens Maga-
zine 5(4):8–36. https:// doi. org/ 10. 1109/ MGRS. 2017. 27623 07
Publisher's Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

1.
2.
3.
4.
5.
6.
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:
use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
use bots or other automated methods to access the content or redirect messages
override any security feature or exclusionary protocol; or
share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at
onlineservice@springernature.com