s42979-024-03380-5

SN Computer Science (2024) 5:1073
https://doi.org/10.1007/s42979-024-03380-5
ORIGINAL RESEARCH
Multispectral Satellite Data Classification Using Rank Correlation

Similarity Index and Automatic Labeling Technique
Arindom Ain1 · Minakshi Gogoi2 · Dibyajyoti Chutia3
Received: 9 June 2023 / Accepted: 1 October 2024

© The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2024
Abstract
Multispectral satellite image classification is crucial in remote sensing and image analysis, serving various applications
like water area extraction, urban region classification, and vegetation monitoring. Model-based classification, particularly
supervised learning approaches, heavily relies on manually annotated training samples, which are labor-intensive and subject
to human bias. However, solely depending on features derived from individual pixels may fail to capture spatial contextual
information effectively and manage mixed pixels and various types of noise, resulting in classification inaccuracies. To
address these challenges, this research proposes a novel unsupervised learning approach and automatic labeling technique
(ALT). This approach utilizes collective information from groups of pixels instead of relying solely on individual pixels for
identification and extraction of meaningful unique patterns which serves as the most significant factor for satellite image
classification. A new similarity index, the patch similarity index (PSI), is generated by incorporating the rank correlation
similarity index (RCSI). The PSI is used to determine the similarity factor both between the patches and within the patch.
Subsequently, a clustering algorithm is employed to group pixels into four distinct classes: vegetation, water, built-ups, and
open land. Additionally, ALT streamlines the post-classification process by reducing the need for manual intervention in
mapping clusters to the actual class label. The proposed method offers a more accurate and reliable unsupervised classifica-
tion for multispectral satellite imagery, promising enhanced efficiency and reduced manual labor.
Keywords Rank Correlation · Multispectral · Patch similarity index · Frequent masking · Automatic labeling technique
Introduction transforming multi-band raster imagery into a single-band

raster with distinct categorical classes representing different
With the continuous advancements in remote sensing tech- land cover types.
nology, the classification of multispectral satellite images
has become a crucial aspect in various applications [1], Motivation
including water area extraction [2], urban region classifica-
tion [3], and vegetation monitoring. This process involves Numerous techniques have been proposed for multispec-
tral satellite image classification [4–8]. The model-based
* Arindom Ain approach has emerged as a comprehensive methodology
[email protected] in multispectral satellite image classification [7–9]. This
Minakshi Gogoi approach requires the development and utilization of
[email protected] explicit models that encapsulate the relationships between
Dibyajyoti Chutia spectral characteristics and land cover types. These models
[email protected] are often crafted based on prior knowledge of the study area
and domain expertise, making them particularly effective
1
Dept. of Computer Science and Engineering, GIMT, Under in capturing complex patterns within the satellite imagery.
Assam Science and Technology University, Guwahati,
Assam, India One of the prevailing paradigms within the model-based
2 approach is supervised learning. These learning methods,
Dept. of Computer Science and Engineering, Girijananda
Chowdhury University, Guwahati, Assam, India including Support Vector Machines [10–12], Random For-
3 ests [13], and Convolutional Neural Networks [9, 14] are
NESAC, Umium, Shillong, India
SN Computer Science
Vol.:(0123456789)
1073 Page 2 of 20 SN Computer Science (2024) 5:1073
designed to categorize pixels, based on a labeled training Objectives

dataset. These algorithms leverage annotated examples of
different land cover types to train the model, enabling it to The principal objective of this research is to separate the
generalize and make predictions on new, unseen data. The pixels of the multispectral satellite data into four groups
effectiveness of supervised learning methods in satellite representing vegetation area, water area, built-up areas,
image classification relies heavily on well-annotated train- and open land areas. To accomplish this, the objective is
ing datasets that include spectral bands, textures, and rele- divided into two parts. Firstly, to develop a new unsuper-
vant features [15]. Furthermore, successful learning neces- vised learning approach, which will ensure the capture
sitates access to extensive datasets. However, the human of meaningful unique patterns of the pixel necessary for
annotation of labeled datasets is a labor-intensive and the classification process without guidance. And secondly
time-consuming task [16]. It introduces subjectivity and to develop a technique that will automatically map the
potential challenges in obtaining large and diverse labeled cluster of pixels to the corresponding class label without
datasets for satellite imagery [17, 18]. The complexity is additional human effort.
heightened by spatial and spectral variability and makes
manual annotation challenging. This motivates the explo-
ration of alternative, less human-dependent methods, such Contribution
as unsupervised learning where the classification is done
without explicit guidance or labeled examples. One widely In response to challenges faced in capturing the proper
used unsupervised learning technique is clustering [19–21], spatial and spectral information of the pixels in satellite
which aims to partition the data into distinct groups based imagery in the classification process and manual labeling
on similarities. Though this algorithm explores the inher- of clusters under human supervision in the post-classifica-
ent structure within the data to group similar instances, it tion stage, we propose an unsupervised learning approach
widely depends on the extraction of hidden patterns or the and Automatic Labeling Technique (ALT) in this research.
relationship of the input data to the classification algorithm. The unsupervised approach includes a patch similarity
In the realm of unsupervised classification of multispectral index (PSI) to classify the target pixels into respective
satellite images, pixel-based methods [22, 23] have long four classes. Instead of using pixels directly for classifi-
been employed to assign individual pixels to specific land cation, this research groups the pixels into patches. The
cover classes without the need for labeled training data. In patch consists of center pixel and neighborhood pixels.
pixel-based approaches, spectral information of individual The center pixel is the target pixel to be classified and
pixels is utilized, and clustering algorithms are applied to it is surrounded by its neighborhood pixels. The idea is
group pixels with similar spectral signatures into distinct to determine the influence of the neighborhood pixels on
classes. However, pixel-based methods often face chal- the target pixels and use this influence in the classifica-
lenges, particularly in the presence of mixed pixels where tion process. The PSI is used to determine the similarity
the land cover characteristics vary within a single pixel. factor both between the patches and within the patch. It
This can lead to inaccuracies and misclassifications, espe- incorporates rank-based correlation of the patch by gener-
cially in heterogeneous landscapes. This method may also ating a rank correlation similarity index (RCSI) matrix. By
struggle to capture spatial contextual information, which introducing rank-based correlation, it intends to improve
is crucial for distinguishing between classes with similar the robustness of positional variability and handle hetero-
spectral properties. Moreover, positional variability where geneous land cover within patches. Calculating rank-based
pixels of the same type of land might look different due correlation coefficients for each target pixel within a patch
to noise factors like cloud shadow, organic and inorganic transforms raw pixel values into ranked values. This cor-
constituents in water, high rise building shadow, for which relation introduces a level of consistency and stability by
they struggle to maintain consistent spatial relationships considering the relative order of pixel values, thus mitigat-
among the pixels. Consequently, this inconsistency affects ing the impact of noise and positional variations. Notably,
the method’s capability to accurately depict land cover pat- the approach ensures that the method is less sensitive to
terns. This encourages us to explore a new unsupervised the heterogeneity of land cover classes within a patch and
learning approach that will extract complex information recognizes the significance of both inter-patch and intra-
considering the relationships of pixels and unique patterns patch relationships. This serves as a pre-processing step
of the dataset in classification. Furthermore, as discussed that refines the input data before applying it to clustering
earlier in this section, the post-classification stage requires algorithms, ultimately resulting in a more accurate and
manual annotation of clusters, prompting our motivation to reliable unsupervised classification for multispectral sat-
develop a method for automatically mapping labels without ellite imagery. Furthermore, the proposed ALT approach
dependence on human intervention.
SN Computer Science
SN Computer Science (2024) 5:1073 Page 3 of 20 1073
improves the post-classification stage by automatically where CDF is the cumulative distribution function of the
mapping the identified clusters to the corresponding standard normal distribution. |τ | is the absolute value of
labels. This approach streamlines the classification process Kendall’s tau coefficient.
without relying on additional human effort to identify and
map pixels to corresponding classes.
Image Data
Manuscript Structure Overview
The image sets used in this research were collected from
The manuscript is thoughtfully structured into five main sec- the Operational Sentinel-2 Imager on 14th February 2023,
tions, each dedicated to a specific aspect of the research. obtained from the United States Geological Survey (USGS)
Section2 provides a detailed background study and material Earth Resources Observation https://e arthe xplor er.u sgs.g ov/.
used in this research. Section 3 explains the methodology The Sentinel-2B satellite provides both spatial and spec-
employed in our research, offering clarity on the approach tral information. To ensure accurate analysis, images were
taken. Section 4 presents the results of the research and pro- atmospheric corrected and converted to reflectance using
vides a comparative discussion of these findings. Finally, the the SCP plugin of QGIS software, which is commonly used
manuscript concludes with a conclusion section that sum- in atmospheric correction [28]. In this research, six spectral
marizes the main points and conclusions of the research. bands were selected from the Sentinel-2B imagery. These
bands include Band 2 (492.1 nm), Band 3 (559.0 nm), Band
4 (664.9 nm), Band 8 (832.9 nm), Band 11 (1610.4 nm),
Background Study and Band 12 (2186.7 nm). These bands were chosen from
the available 13 spectral bands as they are commonly found
Correlation Coefficient in optical sensors and provide valuable information for our
analysis. To ensure consistent cell resolution across the data,
To measure the association between two measured quanti- the spatial resolution of the shortwave bands (Band 11, and
ties, Kendall’s τ coefficient is commonly used. It suits small Band 12) at 20 m was resampled to 10 m using the Bilinear
sample sizes and data with tied ranks [24]. It considers Resampling technique [29], where weighted average values
pairs of observations and determines the strength of asso- of the four nearest pixels values are used to calculate the
ciation based on the patterns of concordance and (discord- ) new pixel of the output raster. All six selected bands have
ance between the pairs. For a set of observations a1 , b1 , the same spatial resolution but different spectral properties.
( )
…, an , bn of the joint random The maps generated are projected in the World Geodetic
[ variables A and B, the num-
ber of concordant pair Nc = (x, y)|ax < ay and bx < by ] and System 1984 (EPSG:7780).
[
number of discordant pair Nd = (x, y)|ax < ay and bx > by ]
The Kendall τ coefficient is given by Eq. 1
Test Area
N − Nd
𝝉= c (1)
n The test scene for this research was collected from Tumak-
Furthermore, tied pairs occur when ax = ay and bx = by, uru district (TD) and Bengaluru Urban district (BUD) in
and they are neither concordant nor discordant. To handle the state of Karnataka [Fig. 1(a)], India, having a coordinate
tied pairs, Stuart et al. [25, 26] proposed a method by modi- of 13.34°N, 77.1°E [30] and 12.9700°N, 77.6536°E [31]
fying the coefficient as shown in Eq. 2 respectively respectively. Two random regions within these
districts were selected as the test areas, each marked by a
2(Nc − Nd ) red rectangular box. The first test area (TA1) selected from
𝝉= (2)
n2 (m−1) TD, covered an area of 1601 hectares, while the second test
m
area (TA2) from BUD covered an area of 1502 hectares [see
here r and c are the number of rows and columns, respec- Fig. 1(b) and Fig. 1(c)]. These test areas were chosen based
tively, and m = min(r, c). on their complex surface features, encompassing various
To determine the probability p [27], associated with each land cover types such as water, open land, vegetation, and
possible value of 𝜏 is calculated based on an approximation built-up areas. It is worth mentioning that these regions also
using the Gaussian distribution given by Eq. 3. contained certain sources of noise, including high-rise build-
ing shadows, cloud shadows, and organic and inorganic con-
Pvalue = 2 ∗ (1 − CDF(|𝝉|)) (3) stituents in water. The selection of scenes with such noise
elements was intentional to evaluate the robustness of the
proposed methods in handling challenging scenarios.
SN Computer Science
Fig. 1 a Karnataka District b TA1 c TA2
Reference data to verify the classification accuracy points, we identified vegetation pixels (V) as label 1, built-
up pixels (BU) as label 2, water pixels (W) as label 3, and
Earlier research studies have revealed the effectiveness of open land pixels (OL) as label 4. Consequently, a refer-
high-resolution images in obtaining the labels of different enced map was generated for TA1 and TA2, corresponding
land covers of satellite imagery [32, 33]. In this research, to their respective spatial extents. The reference image
we employed Google Maps having a resolution of about of TA1 and TA2 is shown in Fig. 2 (a) and Fig. 2 (b)
0.15 m [34] with a similar time frame, to extract the cor- respectively. It is important to note that the labeled maps
responding labels for the test areas. Both TA1 and TA2 obtained are considered ground truth for this research and
were delineated on the Google map. As the target classes are only utilized for validating our classification results.
are visually distinguishable, we used manual digitization The reference map for TA1 consists of a total of 758 poly-
of the four land cover classes referring to high-resolution gons and 150,495 pixels, while for TA2, it comprises 395
Google Earth imagery for generating reference data, for polygons and 21,823 pixels. The distribution of pixels
each of the test sites, corresponding to their respective among the different land cover classes in the reference
spatial extents. To assign labels to the selected sampling data is presented in Table 1.
SN Computer Science
Fig. 2 Reference Image of [a

TA1 b TA2]
Table 1 Number of pixels Class Number of Pixels Table 2 Confusion matrix for four class classification
associated with each class in
reference image of TA1 and TA1 TA2 Predicted Values
TA2 Class 1 Class 2 Class 3 Class 4
Vegetation 63,139 11,853
Built-up 21,657 1992 Actual Class 1 A11 A12 A13 A14
Water 8301 1832 Values Class 2 A21 A22 A23 A24
Open Land 57,398 6146 Class 3 A31 A32 A33 A34
Class 4 A41 A42 A43 A44
Accuracy Assessment
calculations for each class. The TP, FP, FN, and TN values
To assess the accuracy of our proposed method for mapping for the four classes are computed using the confusion matrix,
four classes, we evaluated by counting the number of pixels as illustrated in Table 3. To evaluate the performance of the
that were correctly classified and incorrectly classified. To unsupervised learning-based classification results, several
measure the performance, we utilized two widely used met- other metrics are employed for each class, including overall
rics: overall accuracy (OA) and kappa coefficient (KC) [35]. accuracy, precision, recall, and F1-score, as defined Table 4.
These metrics were calculated based on the confusion matrix Additionally, the output is analyzed using the Kappa coef-
(CM) method, which has been extensively employed in vari- ficient to further assess the classification results.
ous studies [35–38]. The confusion matrix consists of four
classes, namely True Positive (TP), False Negative (FN),
False Positive (FP), and True Negative (TN). The interpreta- Methodology
tion of these classes is as follows:
The methodology proposed in this research paper consists of
• TP: The number of times the classifier correctly predicts three key stages for the classification of multispectral satel-
the positive class. lite images:
• FN: Counts the number of times where the classifier
incorrectly predicts the positive class as negative. Patch Similarity Index (PSI) Generation:
• FP: Counts the number of times the classifier incorrectly
predicts a negative class as positive. • Patch Generation: A set of’N’ group of pixels
• TN: Counts the number of times the classifier correctly referred to as patch is generated from the 3D raster
predicts the negative class. ( Rraster ) representing the multispectral image.
• Rank Correlation Similarity Index (RCSI) Calcula-
The confusion matrix (CM) for mapping the target tion: A RCSI matrix is computed for each patch to
class and reference map in the four-class classification is quantify the correlation between the target pixel and
presented in Table 2. In contrast to binary classification its neighboring pixels.
problems [35], multi-class classification requires separate
SN Computer Science
Table 3 TP, FN, FP and TN values of four class classification

TP FN FP TN
Class 1 A11 A12 + A13 + A14 A21 + A31 + A41 A22 + A23 + A24 + A32 + A33 + A34 + A42 + A43 + A44
Table 4 Metrics
Evaluation Metrics Description Expression
Precision The ratio of predictions as the positive class was positive TP

TP+FP
Recall Measures what fraction of all positive samples were correctly predicted as positive by the classifier TP
TP+FN
F1-score F1-score Harmonic mean of precision and recall precision∗recall
2 ∗ precision+recall
Overall Accuracy Overall accuracy represents the fraction of the total samples that were the classifier correctly classified TP+TN
TP+TN+FP+FN
• PSI Generation: The RCSI matrix obtained is used a total of 4500 randomly distributed sample values from each
to generate the PSI matrix for further classification class, based on the reference data. These samples were collected
process. separately from each band of the RSbands. We then calculated
the mean reflectance value for each spectral band from both the
Clustering: A clustering algorithm is employed to group
test scene, corresponding to the target class. To visualize and
the data points based on PSI values. This step aims to
analyze the spectral characteristics, we plotted a spectral curve,
identify distinct clusters representing the four land cover
as shown in Fig. 4. The Y-axis represents the reflectance values
classes.
of the target classes (pixel values), while the X-axis represents
Automatic Labeling: The clusters obtained in the previous
the corresponding spectral bands. The spectral curve highlights
step are assigned to one of the four target classes using an
several key observations-
ALT and Frequent Masking (FM) technique.
• Water pixels primarily absorb energy in the SWIR
The overall architecture of the proposed methodology is
region, but they also exhibit a moderate reflectance in
illustrated in Fig. 3. The input to this approach is a radio-
the GREEN and NIR bands.
metrically corrected multispectral image from the Sentinel
• Vegetation pixels show high reflectance in the NIR wave-
2B satellite. Therefore, no additional radiometric correction
length and moderate reflectance in the GREEN bands,
is applied as part of the methodology.
while they absorb energy in the SWIR wavelengths.
• Built-up features reflect more energy in the visible wave-
Patch Similarity Index (PSI) Generation
length range, whereas open land reflects more in the NIR
and SWIR1 bands. Both built-up and open land pixels
This sub section provides a comprehensive breakdown of
show energy absorption in the SWIR2 wavelength.
the various steps involved in generating the PSI, including
spectral curve generation, deriving indices, spectral analysis,
Deriving indices from RS bands
spectral-spatial feature extraction, and the formulation of the
RCSI and PSI matrices.
Spectral indices have been widely used in previous studies to
investigate various fields such as agriculture, water resources,
Spectral curve generation from the six spectral bands
and urban development [32, 33, 36, 37, 39]. These indices are
computed by combining reflectance values from multiple spec-
Classification of satellite images or separating pixels into clus-
tral bands (RSbands) into a single value. In this research, several
ters commonly relies on the spectral bands available in the satel-
indices were derived based on specific formulas, as presented in
lite imagery. We stack together the six spectral bands to create
Table 5. It is important to note that these indices were selected
a 3D spectral raster called RSbands. This allows us to analyze the
based on their capability to highlight specific land cover classes.
spectral characteristics of the target class pixels. To investigate
All the indices are then stacked together to form a 3D indices
the spectral properties of the target class pixels, we collected
raster called RI bands. Each index is considered as an individual
SN Computer Science
Fig. 3 The architecture of the proposed methodology
band. Following the steps, for selecting target pixel values as dis- Spectral analysis of the RS bands and the derived indices
cussed in section A.a, we plotted another spectral curve shown bands:
in Fig. 5 based on the RI bands. The spectral indices are shown
in the X-axis and the corresponding pixel values for each of Analyzing the spectral curve is a useful approach to gain insights
the target classes are shown in the Y-axis. The spectral curve into the properties of land surface pixels [36, 39]. Different land
provides insights into the behavior of the target classes across cover types exhibit distinct spectral properties. It is observed
the derived indices. In Figs. 6 and 7, we display the output of the that vegetation has high reflectance in the near-infrared (NIR)
derived indices for TA1 and TA2, respectively. The analysis of region and water pixels absorb more energy with low reflectance
the spectral indices reveals interesting patterns. The BSI, NDWI- in NIR and shortwave infrared (SWIR) wavelengths. As each
B, AEWInsh, MNDWI, BU, and NBI indices exhibit positive class label interacts with light energy differently, noticeable vari-
values for water regions and negative values for other classes. ations are observed in the normalized difference trends shown
On the other hand, indices such as BU, NDBI, and NBI indices in Fig. 4 and Fig. 5. Figure 4 illustrates that the spectral curve of
effectively highlight urban or built-up areas. Additionally, the built-up areas and open land share a similar trend in the RSbands.
EVI, NDVI, NDWIns, and NDWI-B indices emphasize vegeta- However, a significant disparity can be observed in the reflec-
tion regions, while the BSI, NDVI, NDBI, and NBI indexes pre- tance curves of indices NDBI, BSI, and NBI, as shown in Fig. 5.
dominantly highlight open land areas, suppressing other indices. Furthermore, the spectral curve of water exhibits low variance
The derived spectral indices provide valuable information for compared to the indices curves, which display high negative and
characterizing different land cover classes and are utilized for positive values for water pixels across different indices. While
further analysis and classification tasks. the vegetation curve in RSbands relies solely on the NIR band for
SN Computer Science
Fig. 4 Spectral graph for the 6

optical bands
Table 5 Spectral indices Spectral‑Spatial Feature Extraction

Indices Formulae
While spectral features can partially distinguish different classes
NDWI-B [40] (blue − nir)∕(blue + nir)
based on the quantity of reflection, they often result in mislead-
AEWInsh [41] 4 ∗ (green − swir1) − (0.25 ∗ nir + 2.75 ∗ swir1)
ing pixels due to similar spectral properties of shadow pixels,
MNDWI [42] (green − swir1)∕(green + swir1)
dark object pixels, and mixed pixels. This research adopts a strat-
NDWIns [33] (green − 2 ∗ nir)∕(green + nir)
egy of grouping pixels into patches. Each patch comprises a
NDSI [33] (swir2 − green)∕(swir2 + green)
center pixel along with its surrounding neighborhood pixels. The
BU [43] (swir1−nir) (nir−red)
(swir1+nir)
− (nir+red) center pixel serves as the focal point for classification, while its
NDBI [44] (swir1 − nir)∕(swir1 + nir) neighboring pixels provide contextual information for the classi-
NBI [44] (swir1 ∗ red)∕nir fication task. To accomplish this, we leverage the neighborhood
EVI [45] 2∗(nir−red)
information of the target pixel. The stacked raster data Rraster is
(nir+6∗red−7.5∗blue)+1
NDVI [46] (nir − red)∕(nir + red) divided into 3D patches using a stride of 1, as shown in Equa-
BSI [47] (swir1+red)−(nir+red) tion 4. Each patch has dimensions (m ∗ n ∗ d), representing the
(swir1+red)+(nir+red)
row, column, and number of bands of the patch, respectively. To
account for edge pixels in the patches, a padding technique is
normalized difference computation, the RI bands demonstrates applied. The most used padding techniques include zero padding
varying reflectance values. Therefore, it is valuable to include and reflect padding [48–50]. The zero-padding technique fills
the spectral information from indices in conjunction with the the outer padded area with zeros which may introduce unrelated
spectral bands to effectively discriminate between the target data into the input. So, in this research, we have used reflection
classes. To achieve this, we stack the RSbands with the RI bands, padding techniques [51] that reflect the input values instead of
resulting in a consolidated 3D raster dataset denoted as Rraster. replicating the same value across the border axis to introduce
It has an image size of ( Rrow ∗ Rcolumn∗Rbands), encompassing more texture information than object information. The pseu-
both the spectral bands and derived indices. This integration docode for patch generation is given Algorithm 1.
allows for comprehensive spectral information utilization in the
separation of target classes.
SN Computer Science
Fig. 5 Spectral graph for the 11

derived bands
Fig. 6 Spectral indices of TA1: a NDBI-B b NDVI c NDSI d NDBI e NBI f MNDWI g EVI h BU i BSI j AWEInsh k NDWIns
SN Computer Science
Fig. 7 Spectral indices of TA2: a NDBI-B b NDVI c NDSI d NDBI e NBI f MNDWI g EVI h BU i BSI j AWEInsh k NDWIns
hand, pixel pairs from different classes tend to have correlation

Rraster = P1 , P2 , P3 , ..., PN (4)
values close to zero. To quantify the influence of neighborhood
where ‘N’ is the number of pixels in the raster and pixels on the center pixel, an RCSI matrix, RCSI N is generated
P1 , P2 , P3 , … , PN are patches of Rraster. for each patch of size (h, h, d), where h is an odd number and h
Each
( patch consists ) of a center pixel Cx where, does not equal to 1. The RCSI matrix incorporates modified val-
x= , 2 , d and set of neighborhood pixels Cy given
(m−1) (n−1) ues of the center pixel and its neighboring pixels. This process
2
involves two steps. Firstly, the value of the center pixel for each
by Eq. 5.
patch is computed using Equation 6. Then, the spectral values
}| of each neighborhood pixel are sorted in ascending order based
Cy=(p,q,r) = Cy1 , Cy2 , Cy3 , ..., Cy(m∗n−1) |y! = x (5)
| on the reflectance values of the center pixel. For each patch, the
where (1 ≤ p ≤ m), (1 ≤ q ≤ n), (1 ≤ r ≤ d) and Cy1 , Cy2 , … Kendall 𝝉 c rank correlation and pvalues between the center pixel
are the set of neighborhood pixels except the center pixel. Cx and its corresponding neighborhood pixels Cy, as described
in Eq. 2 and Eq. 3 respectively in this paper, are calculated and
represented in matrix form in Eq. 7.
Formulation of RCSI Matrix [ k−1 ]/
∑
( )
Cx = C m−1 , n−1 ,d (k + 1) (6)
The strength of the correlation between two variables can be d=0 2 2
determined by the likelihood of their values changing together

in the same or opposite direction. In the context of this research, where k = number of bands
the variables refer to the reflectance of pixels. Within a patch,
there is likely to exist a strong relationship between the reflec- ⎡ 𝝉 11 , p11 𝝉 12 , p12 … … 𝝉 1h , p1h ⎤
⎢ 𝝉 ,p 𝝉 ,p … … 𝝉 2h , p2h ⎥
tance of a center pixel and the corresponding surrounding pixels. 𝐑𝐂𝐒𝐈N = ⎢ 21 21 22 22 ⎥ (7)
For a specific band and class, the reflectance of the center pixel ⎢ ⋮ ⋮ Cx , 0 ⋮ ⎥
⎣ 𝝉 h1 , ph1 𝝉 h2 , ph2 … … 𝝉 hh , phh ⎦
is directly or inversely proportional to the reflectance of each
neighboring pixel. The reflectance of pixel pairs belonging to where 'N' is the number of patches in Rraster. The pvalue for
the same class generally tends to exhibit a high correlation for the center pixel in each patch is zero, indicating the highest
corresponding bands, regardless of their location. On the other
SN Computer Science
correlation within the patch. Secondly, the 𝜏c − statistic of (1 − pvalue ), resulting in the final normalized RCSI matrix for
each cell is divided by the corresponding significance quantity each patch.
Algorithm 1 PatchGeneration ( )
Formulation of PSI matrix − 1 to 1, as described in Equation 10 . Standardization was

preferred since it is not influenced by outliers, as there is
The PSI matrix is calculated by taking the dot product of the no predefined range for the transformed features.
RCSI matrix of each patch with the corresponding patch of
Rraster , as shown in Eq. 8.
Rraster = (x − 𝜇)∕s. (10)
where ‘x’ is the sampling pixel, µ is the mean of those sam-
PSIN = RCSIN ⋅ PN (8)
ples, and ‘s’ is the standard deviation of the samples. Moreo-
Where N is the number of patches in Rraster ver, many algorithms were proposed by several researchers
The pseudocode for generating the PSI matrix is given to group the unknown dataset into different clusters [52, 53].
in Algorithm 2. K-Means algorithm [54] is one of the most used unsuper-
Now, the Rraster is represented by Eq. 9. vised learning algorithms. So, in this research, we have used
the K-Means algorithm to assign pixels to 4 groups based on
Rraster = PSI1 , PSI2 , PSI3 , … , PSIN (9) the similarity of the pixel’s properties. Since the K-means
algorithm is not well-suited for handling high-dimensional
Grouping data points using unsupervised approach data due to its simplicity and limited computational capa-
bilities [55], it is necessary to reduce the dimensionality of
In satellite datasets, outliers are often present, which are the data before applying the clustering algorithm. Previous
anomalous values compared to the surrounding pixels. studies, such as [56], have employed principal component
These outliers are primarily caused by undetected cloud analysis (PCA) for dimensionality reduction in satellite
cover and rain. Additionally, the input data exhibit a wide image analysis. In this research, we utilized PCA to reduce
range of scales, varying from high positive values to high the dimensionality of the scaled Rraster data to a specific
negative values. To address this, data scaling is performed number of principal components. We have used the first
using standardization techniques. We have utilized stand- thirty percent of the principal components as input data to
ard scaling to transform the values of Rraster to a range of the algorithm. The centroid in our approach is not by the
SN Computer Science
mean of neighboring pixels alone. Instead, the correlated clusters by minimizing the in-cluster sum of squares. Pixels
influencing factors of the neighboring contribute to the act as with similar properties are clustered together, resulting in a
the centroid. Each center pixel is assigned to one of the four classification map Rm∗n
clustered
containing four distinct clusters.
Algorithm 2 PatchSimilarityIndexGeneration( )
Assigning Labels to Clusters generating Bmask. Threshold values, as suggested in the relevant
literature, are employed to separate target pixels and suppress
Assigning labels to clusters often involves subjective human unwanted pixels for each class label. These threshold values are
observation, leading to potential errors and misleading classifi- presented in Table 6. The mask maps for water, vegetation, open
cations. Traditional methods [6, 57] for creating and validating land, and built-up areas are depicted in Fig. 8 (a–d) for TA1 and
labeled maps are prone to human subjectivity and errors. Gen- 8 (e–h) for TA2, respectively. It is important to emphasize that
erating labeled data for satellite images is particularly challeng- the primary focus is on the major mask, disregarding small clas-
ing due to their complex spectral properties. In this research, we sification errors or misclassified pixels. The concept involves
proposed a method that utilizes spectral indices derived from overlaying individual mask maps of the corresponding target
target image data to generate binary classification maps for spe- pixels onto the clustered image to assign specific target labels to
cific target classes. The MNDWI index [42] is employed to dis- each corresponding clustered pixel. For instance, for water iden-
tinguish water and non-water pixels in the test image using a tification and automatic labeling, a Rm∗n
merged
matrix is generated
threshold, resulting in a binary water mask (Wmask), where water through pixel-wise multiplication of the Rclustered and Wmask. The
m∗n
pixels are assigned a value of 1 and non-water pixels 0. Simi- Rm∗n matrix contains positive pixels likely to be represented
merged
larly, EVI index are used to generate a vegetation mask (Vmask) as water and 0 values for others. But it may have misleading
to separate vegetation and non-vegetation pixels, with values of
Table 6 Threshold value for indices
1 and 0 assigned accordingly. To extract the open land pixels and
suppress non open land pixels in the test image to generate Class Threshold Label 1 Label 0
Omask , the BSI index is utilized, assigning values of 1 and 0,
Water [40] NDWI > 0 Water pixel Other pixels
respectively. Although the NBI index is used to extract built-up
Vegetation [45, NDVI > 0.3 Vegetation pixel Other pixels
regions, it is also associated with open land pixels in the classi- 46]
fication. To address this, we identify positive pixels resulting Open land [47] BSI > 0.2 Open land pixel Other pixels
from the difference between the NBI and BSI, assigning them a Built- ups [44] [(NBI > 0.1) Built-up pixel Other pixels
value of 1 for built-up pixels while assigning 0 to other pixels -(BSI > 0.2)] > 0
SN Computer Science
Fig. 8 TA1 [a Water mask, b Vegetation mask, c Open Land mask, d Built-up mask] TA2 [e Water mask, f Vegetation mask, g Open Land
mask, h Built-up mask]
pixels. So, to handle this a new masking technique known as label, and Bmask , for the built-up labels. The pseudocode for
Frequent Masking (FM) method is proposed. In this method the assigning labels to clusters can be found in Algorithm 3. It is
frequency of all unique non-zero pixels in Rm∗n merged
is calculated. worth noting that there can be multiple cluster labels with the
The cluster label (number) associated with the highest frequency same frequency can be found in Rm∗n merged
, indicating ambiguity in
is identified. Then all the pixel associated with the correspond- the assigned labels, in that case, the clustering algorithm is re-
ing cluster number in Rm∗nclustered
is identified and marked as water evaluated to find corresponding clusters, till all the labels are
labels. The same procedure is applied using Vmask, to identify the identified and properly labeled.
vegetation cluster label, Omask, to detect the open land cluster
Algorithm 3 AutomaticLabelingTechnique( )
SN Computer Science
Results and discussion
Kappa
ficient
Coef-
0.762
0.531
0.468
Comparative analysis of different classification
methods with the proposed method
Overall Accuracy
To conduct a comprehensive comparative analysis, we evaluated
several state-of-the-art classification approaches, focusing on
their effectiveness in classifying land cover types. A commonly
0.856
0.655
employed method is pixel-based classification, widely used in
0.6
numerous studies [4, 5, 58–61]. In this method, raw pixel values
are directly fed into the clustering algorithm to classify the four
F1-Score
land cover classes. This is referred to as Pixel-Based Classifica-
0.705
0.963
0.761
0.637
0.275
0.964
0.84
0.89
0.97
0.82
0.31
0.72
tion (PiBC) in this research. Another approach involves group-
ing pixels into chunks, rather than treating them individually,
to classify different land cover types [62, 63]. It is important
0.7662
Recall
0.563
0.967
0.664
0.486
0.741
0.971
0.613
to note that this approach relies solely on multispectral bands
0.83
0.84
0.97
0.84
from satellite imagery and does not incorporate derived spectral
indices. Additionally, these chunks use the reflectance values of
Precision
the pixels within the group for classification, without consider-
0.2001
0.944
0.959
0.891
0.169
0.957
0.894
ing the influence of neighboring pixels on the target pixels. This
TA2
0.85
0.94
0.97
0.92
0.8
method is distinctly different from the proposed approach and
is referred to as Simple Chunk-Based Classification (CuBC ).
Coefficient
Some researchers have integrated dimensionality reduction
techniques into the CuBC approach to improve classification
Kappa
0.609
0.557
0.64
accuracy [64–66]. This variant is termed Chunk-Based Clas-
sification with PCA (CuBCpca) in this research. The above
approaches were evaluated on two feature sets: RSbands (multi-
Accuracy
spectral bands) and Rraster (a combination of RSbands and RI bands,

Overall
0.734
0.698
0.76
the latter representing derived spectral indices) across two test
scenes, TA1 and TA2. The quantitative classification results
for RSbands are presented in Table 7, while those for Rraster are
Table 7 Classification accuracies of different approaches using RSbands feature set
F1-Score
in Table 8. In this section, we thoroughly analyze and compare
0.75
0.76
0.95
0.76
0.67
0.97
0.88
0.62
0.39
0.97
0.89
the performance of these classification approaches against the

0.5
proposed method using various evaluation metrics such as pre-

cision, recall, F1-score, overall accuracy (OA), and the Kappa
Recall
Coefficient (KC). The classification accuracy of the proposed

0.74
0.82
0.99
0.75
0.61
0.74
0.97
0.84
0.55
0.57
0.98
0.86
methods is summarized in Table 9, and the comparative perfor-

mance of the different approaches against the proposed method
Precision
is depicted in Figs. 9 and 10 for TA1 and TA2, respectively.

For both RSbands and Rraster , the PiBC showed moderate per-
0.76
TA1
0.71
0.97
0.77
0.74
0.38
0.98
0.94
0.71
0.29
0.96
0.93
formance. It achieved high precision and recall for the "Water"

class, indicating accurate identification. However, the "Built-
Class Label
up" class exhibited low precision, with values of 0.29 (RSbands,

TA1), 0.34 (Rraster, TA1), 0.16 (RSbands, TA2), and 0.18 (Rraster,
TA2), indicating a high number of false positives and poor class
BU
BU
BU
OL
OL
OL
W
W
W
V
V
V
separation. The recall and F1-scores for the "Built-up" class

were moderate, pointing to incomplete classification. Addition-
CuBCpca
ally, there was relatively low recall for "Open Land," indicat-
CuBC
PiBC
ing some misclassification of pixels in both feature sets across

test scenes. The overall accuracy for this method was 69.83%
Classification
(RSbands, TA1), 70.72% (Rraster, TA1), 60% (RSbands, TA2), and

RSbands
63.20% (Rraster, TA2). The Cohen’s Kappa Coefficient values,
SN Computer Science
Table 8 Classification accuracies of different approaches using Rraster feature set

Classification Class Label TA1 TA2
Precision Recall F1-Score Overall Accuracy Kappa Precision Recall F1-Score Overall Kappa
Coef- Coef-
ficient Accuracy ficient
Rraster PiBC V 0.94 0.83 0.88 0.707 0.572 0.9 0.65 0.75 0.632 0.504
W 0.98 0.97 0.97 0.96 0.97 0.96
BU 0.34 0.69 0.46 0.18 0.74 0.29
OL 0.71 0.57 0.63 0.93 0.53 0.67
CuBC V 0.93 0.86 0.9 0.722 0.591 0.89 0.66 0.76 0.637 0.51
W 0.96 0.98 0.97 0.96 0.97 0.96
BU 0.34 0.62 0.44 0.19 0.77 0.31
OL 0.73 0.59 0.66 0.94 0.52 0.67
CuBCpca V 0.77 0.82 0.8 0.801 0.695 0.8 0.97 0.87 0.88 0.816
W 0.98 0.98 0.98 0.95 0.98 0.97
BU 0.79 0.87 0.83 0.93 0.84 0.89
OL 0.81 0.74 0.77 0.96 0.8 0.87
0.5575 ( RSbands, TA1), 0.5723 ( Rraster , TA1), 0.468 ( RSbands, Kappa Coefficient also improved, reinforcing the effectiveness
TA2), and 0.504 (Rraster, TA2), indicated moderate agreement of this approach. In contrast, the proposed method, depicted in
between the predicted and true labels. The classification map Figs. 9 (g) and 10(g) for TA1 and TA2 respectively, outper-
for PiBC method is shown in Figs. 9 (a) and (d) and Fig. 10(a) formed all other approaches. It achieved high precision, recall,
and (d) and for CuBC method is depicted in Figs. 9 (b) and and F1-scores across all four classes. The “Water” and “Vegeta-
(e) and Fig. 10(b) and (e) for TA1 and TA2 respectively. The tion” classes were classified with high accuracy, and the method
CuBC method demonstrated improved performance over the significantly improved the classification of the “Built-up” and
pixel-based approach for both RSbands and Rraster feature sets. It “Open Land” classes, which had underperformed in previ-
achieved higher precision, recall, and F1-scores for the “Vegeta- ous approaches. The overall accuracy of the proposed method
tion” and “Water” classes, indicating better classification accu- was the highest, 90.30% (TA1) and 91.79% (TA2), indicating
racy. However, the performance for the “Built-up” and “Open a significant enhancement in classification performance. The
Land” classes remained relatively low. The overall accuracy Cohen’s Kappa Coefficient values were the highest for proposed
increased to 73.44% ( RSbands, TA1), 72.29% ( Rraster , TA1), method, suggesting substantial agreement between the predicted
65.57% (RSbands, TA2), and 63.70% (Rraster , TA2), indicating and actual classifications. The results demonstrate that the pro-
enhanced classification performance. The Cohen’s Kappa Coef- posed method effectively captures the characteristics of the tar-
ficient also improved, with values of 0.6098 ( RSbands, TA1), get classes and significantly improves classification accuracy
0.5917 ( Rraster , TA1), 0.5315 ( RSbands , TA2), and 0.5108 compared to the other approaches.
( Rraster , TA2), suggesting a stronger agreement between pre-
dicted and true labels. However, the lower accuracy of the Analyzing the Effectiveness of Proposed Approach
CuBC method when using Rraster compared to RSbands suggests
that the increase in dimensionality due to the addition of indi- To illustrate the importance and effectiveness of the
ces negatively impacts the clustering algorithm, as k-means proposed approach, we conducted ablation research by
does not perform well in high-dimensional spaces. Therefore, progressively modifying the proposed architecture and
dimensionality reduction techniques like PCA are essential for excluding key components to evaluate their impact on
further improving performance. The CuBCpca method, shown classification performance. The results of this research
in Figs. 9 (c) and (f) and Fig. 10(c) and (f) for TA1 and TA2 are summarized in Table 10, where the performance is
respectively, achieved consistent precision, recall, and F1-scores measured by OA and KC for two test areas, TA1 and TA2.
for the “Vegetation” and “Water” classes. Although the “Built- This analysis provides insights into the role of each com-
up” and “Open Land” classes showed slight improvements, ponent– RSbands (multispectral bands), RI bands (the spec-
they approached near-complete classification. The overall tral indices), and PCA in the proposed approach. In the
accuracy increased to 76% ( RSbands, TA1), 80.17% ( Rraster , first approach, we excluded the RI bands and only retained
TA1), 85.69% (RSbands, TA2), and 88.05% (Rraster, TA2), dem- the RSbands . This configuration achieved an OA of 0.834
onstrating enhanced classification performance. The Cohen’s
SN Computer Science
and KC of 0.738 compared to the proposed approach that
Kappa Coefficient
includes both RI bands and RSbands . These results indicate
that although using RSbands with dimensionality reduc-
tion improves accuracy, the absence of spectral indices
0.871
restricts the approach from achieving optimal perfor-
mance. Therefore, the combination of multispectral bands
Overall Accuracy and spectral indices is essential for optimal classifica-
tion performance. In the second approach, both PCA and
RI bands were excluded, leaving only the RSbands . This led to
a significant drop in performance, with an OA of 66.16%
0.917
and a KC of 0.5070 for TA1, and an OA of 61.83% with

a KC of 0.4855 for TA2. This sharp decline highlights
PCA's critical role in reducing noise and enhancing rel-
F1-Score
evant features. In the third configuration, we retained both

0.92
0.98
0.84
0.92
RSbands and RI bands but excluded PCA. This resulted in an

OA of 64.64% and a KC of 0.4876 for TA1, and 63.12%
Recall
with a KC of 0.5016 for TA2. Although this approach per-

0.88
0.96
0.81
0.97
formed slightly better than using RSbands alone, the results

remained suboptimal. This underlines the importance of
Precision
PCA in enhancing classification accuracy when using

TA2
0.97
0.99
0.87
0.88
both multispectral bands and spectral indices together.

The fourth approach evaluated the use of only RI bands
excluding RSbands . The results showed an OA of 84.21%
Kappa Coefficient
and a KC of 0.7225 for TA1, and an OA of 83.92% with

a KC of 0.7396 for TA2. These results suggest that spec-
tral indices alone provide substantial classification power.
0.850
However, the performance still fell short compared to the

proposed method, underscoring the value of combining
both RSbands and RI bands for improved results. Lastly, in
Overall Accuracy
the fifth approach, both PCA and RSbands were excluded,

leaving only RI bands . This configuration resulted in an OA
of 0.7968 and KC of 0.6955 for TA1, and 0.7993 with
0.903
KC of 0.6977 for TA2. While the spectral indices alone

performed reasonably well, the absence of PCA under-
F1-Score
scored its importance for achieving superior classifica-

tion accuracy. The ablation research clearly demonstrates
0.89
0.97
0.92
0.90
that each component of the proposed approach plays a

critical role in achieving optimal classification accuracy.
Recall
Table 9 Classification accuracy of the proposed method
The configuration incorporating all three elements con-

0.86
0.94
0.96
0.92
sistently outperformed other configurations. This proves

that each element contributes significantly to the final
Precision
classification results, highlighting the effectiveness of the

TA1
0.93
0.99
0.89
0.88
proposed architecture.
Class Label
BU
OL
W
V
Proposed Method
Classification
SN Computer Science
Fig. 9 TA1 classification for 〖RS〗_bands [a PiBC b CuBC c CuBCpca], TA1 classification for R_raster [d PiBC e CuBC f CuBCpca] g Pro-
posed method
Fig. 10 TA2 classification for 〖RS〗_bands [a PiBC b CuBC c CuBCpca] TA2 classification for R_raster [d PiBC e CuBC f CuBCpca] g Pro-
posed method
Table 10 Analysis of the Approach 1 Approach 2 Approach 3 Approach 4 Approach 5

effectiveness of the proposed
methods OA KC OA KC OA KC OA KC OA KC
TA1 0.834 0.738 0.6616 0.5070 0.6464 0.4876 0.8421 0.7225 0.7968 0.6955
TA2 0.855 0.762 0.6183 0.4855 0.6312 0.5016 0.8392 0.7396 0.7993 0.6977
SN Computer Science
Conclusion technique for mapping surface built-up areas using landsat 8

oli imagery. Remote Sens. 2019;11:17.
4. Younis MC, Keedwell E, Savic D. An investigation of pixel-
Considering the correlated influence of the neighborhood based and object-based image classification in remote sensing.
pixels on the target pixel within the patch, a significant Intern Confer Advan Sci Eng (ICOASE). 2018;2018:449–54.
increase in accuracy is observed for both RSbands and Rraster . 5. S. M.Suganthi. 2014. Pixel based approach of unsupervised
satellite image classification. Intern J Advan Res Comput Com
However, the Rraster approach yields the best results com- Eng.
pared to the RSbands indicating the importance of indices 6. Zhang Y, Liu X, Zhang Y, Ling X, Huang X. Automatic and
in decision boundary. Though PiBC can effectively group unsupervised water body extraction based on spectral-spatial
water and vegetation pixels but overestimates the built-up features using gf-1 satellite imagery. IEEE Geosci Remote Sens
Lett. 2019;16:927–31.
area and is unable to discriminate between built-up and 7. Zhang P, Ke Y, Zhang Z, Wang M, Li P, Zhang S. Urban land
open land pixels. CuBC method also follows the same use and land cover classification using novel deep learning mod-
trends as PiBC except an increase in performance met- els based on high spatial resolution satellite imagery. Sensors.
ric shows some enhancement in classification results. The 2018. https://doi.org/10.3390/s18113717.
8. Jamal Mustafa GWN. Satellite images multiple data using clus-
CuBCpca method showed notable improvement compared tering techniques. Iraqi J Inform Tech. 2018;9(1):97–118.
to PiBC and CuBC methods, effectively separating built- 9. Rai AK, Mandal N, Singh A, Singh KK. Landsat 8 oli satellite
up pixels from open land pixels. However, it struggles to image classification using convolutional neural network. Procedia
detect small built-up areas with sparse populations. In Computer Sci. 2020. https://doi.org/10.1016/j.procs.2020.03.398.
10. N.S.Mansor, H.Awang, S.T.S.Malami, A.Zolkafli, M.A.Taiye,
contrast, the proposed method successfully groups similar- andH.Maulana. 2024. Support vector machine for satellite
class pixels into four different clusters, outperforming the images classification using radial basis function kernel method.
other approaches. The proposed method is compared with Computing and Informatics, N. H. Zakaria, N. S. Mansor, H.
other approaches using a simple K-means algorithm, as Husni, and F. Mohammed, Eds. Singapore: Springer Nature
Singapore, Singapore
other clustering algorithms were deemed too complex and 11. Pathak D, Kalita S, Bhattacharya D. Hyperspectral image classi-
beyond the scope of this research. Comparing the different fication using support vector machine: a spectral spatial feature
classification methods, the proposed method consistently based approach. Evol Intel. 2022;15:1809–23.
demonstrated superior performance in both TA1 and TA2. 12. D. K. Pathak and S. K. Kalita. 2019. Spectral spatial feature
based classification of hyperspectral image using support vector
It achieved the highest precision, recall, and F1 scores for machine. 2019 6th International Conference on signal process-
all classes, indicating its effectiveness in accurately map- ing and integrated networks (SPIN). https://doi.org/10.1109/
ping water pixels, vegetation pixels, built-up pixels, and SPIN.2019.8711731
open land pixels. It effectively addresses the limitations of 13. Belgiu M, L. Dr˘agu¸t. Random forest in remote sensing: a review
of applications and future directions. ISPRS J Photogram Remote
the existing methods, providing more accurate clustering Sens. 2016. https://doi.org/10.1016/j.isprsjprs.2016.01.011.
and higher overall accuracy. Moreover, the proposed ALT 14. Jarrallah ZH, Khodher MAA. Satellite images classification
using FM techniques shows an effective way to automati- using cnn :a survey. Intern Confer Data Sci Intelligent Comp
cally map cluster pixels to the respective class label. In the (ICDSIC). 2022;2022:111–6.
15. Shu Y, Li W, Yang M, Cheng P, Han S. Patch-based change
future, we will try to consider more specific classes for detection method for sar images with label updating strategy.
classification. Remote Sens. 2021. https://doi.org/10.3390/rs13071236.
16. Rasmussen CB, Kirk K, Moeslund TB. The challenge of data
annotation in deep learning—a case study on whole plant corn
Declarations silage. Sensors. 2022. https://doi.org/10.3390/s22041596.
17. S. Gulyanon, W. Limprasert, P. Songmuang, and R. Kong-
Conflict of interest The authors declare no conflict of interest. kachandra. 2022. Data generation for satellite image classifica-
tion using self-supervised representation learning. https://doi.
org/10.48550/arXiv.2205.14418.
18. Safonova A, Ghazaryan G, Stiller S, Main-Knorn M, Nendel
C, Ryo M. Ten deep learning techniques to address small data
References problems with remote sensing. Int J Appl Earth Obs Geoinf.
2023;125: 103569.
19. Usman B. Satellite imagery land cover classification using
1. Giardino Bresciani V. Application of remote sensing in water
k-means clustering algorithm: computer vision for envi-
resource management: the case study of lake trasimeno,
ronmental information extraction. Elixir J CompSci Eng.
Italy. Water Resour Manage. 2010. https://d oi.o rg/1 0.1 007/
2013;10(1):18671–5.
s11269-010-9639-3.
20. S. Karuppanagounder and N. Kalaiselvi. Feature identification
2. Li Y, Gong X, Guo Z, Xu K, Hu D, Zhou H. An index and
in satellite images using k-means segmentation. 03 2012.
approach for water extraction using landsat-oli data. Intern J
21. Kharat V, Khatdeo S, Kothe H, Kshirsagar R, Dixit M, Balan
Remote Sens. 2016. https://doi.org/10.1080/01431161.2016.
M. Land cover clustering and classification of satellite images.
1201228.
ITM Web Confer. 2023;56:08.
3. Firozjaei MK, Sedighi A, Kiavarz M, Qureshi S, Haase D,
Alavipanah SK. Automated built-up extraction index: a new
SN Computer Science
22. S. Xiaoxia, Z. Jixian, and Z. Liu. 2005. A comparison of object- sentinel-2a msi imageries a case of addis ababa city, ethiopia.
oriented and pixel-based classification approachs using quick- Environ Chal. 2022;8: 100568.
bird imagery. 01 2005 45. N.E.Observatory. 2023 Measuringvegetation(ndvi&evi),” \
23. M. Oyekola and K. Adewuyi. 2018. Unsupervised classification {https://earthobservatory.nasa.gov/features/Measur ing Vegeta-
in land cover types using remote sensing and gis techniques. tion}, 28 may 2023.
24. Kendall MG. A new measure of rank correlation. Biometrika. 46. USGS, “Ndvi foundation - remote sensing phenology,” \{https://
1938;30(1/2):81–93. www.usgs.gov/special-topics/remote-sensing-phenology/science/
25. Berry KJ, Johnston JE, Zahran S, Mielke PW. Stuart’s tau meas- ndvi-foundation-remote-sensing-phenology}, 28 may 2023.
ure of effect size for ordinal variables: some methodological 47. Mzid N, Pignatti S, Huang W, Casa, and Raffaele. An analysis
considerations. Behav Res Methods. 2009;41:1144–8. of bare soil occurrence in arable croplands for remote sensing
26. Kendall MG. The treatment of ties in ranking problems. Biom- topsoil applications. Remt Sens. 2021. https://doi.org/10.3390/
etrika. 1945;33(3):239–51. rs13030474.
27. H. Abdi. 2006. The kendall rank correlation coefficient herv 48. G. Liu, K. J. Shih, T.-C. Wang, F. A. Reda, K. Sapra, Z. Yu, A.
abdi Tao, and B. Catanzaro. 2018. Partial convolution based padding.
28. Tempa K, Aryal K. Semi-automatic classification for rapid arXiv preprint arXiv:1811.11718.
delineation of the geohazard-prone areas musing sentinel-2 49. K. K. Perikamana, K. Balakrishnan, and P. Tripathy. 2021. A cnn
satellite imagery. SN App Sci. 2022. https://doi.org/10.1007/ based method for sub-pixel urban land cover classification using
s42452-022-05028-6. landsat-5 tm and resourcesat-1 liss-iv imagery,” arXiv preprint
29. M. R. Singh and A. Bhide. 2016. A review of image retrieval arXiv:2112.08841.
using different types of interpolation techniques. International 50. Park JH, Inamori T, Hamaguchi R, Otsuki K, Kim JE, Yamaoka
Research Journal of Engineering and Technology (IRJET), K. Rgb image prioritization using convolutional neural net-
2016. work on a microprocessor for nanosatellites. Remote Sensing.
30. “Tumakuru, karnataka, india - latitude and longitude,” \{https:// 2020;12(23):3941.
www.latlong.net/place/tumakuru-karnataka-india-11565.html}, 51. O. Thunder, “Reflect padding,” {https://m edium.c om/O rca_T
hund
accessed: [28 may 2023]. er/image-padding-techniques-reflect-padding-part-2-5a013cd965
31. “Bengaluru, karnataka, india - latitude and longitude,” \{https:// 37}, 2024, accessed on 06 june.
www.latlong.net/place/bengaluru-karnat aka-india-499.html}, 52. [S. Basu, S. Ganguly, S. Mukhopadhyay, R. DiBiano, M. Karki,
accessed: [28 may 2023]. and R. Nemani. 2015. Deepsat: A learning framework for satellite
32. Feyisa GL, Meilby H, Fensholt R, Proud SR. Automated water imagery. In Proceedings of the 23rd SIGSPATIAL International
extraction index: A new technique for surface water mapping Conference on Advances in Geographic Information Systems.New
using landsat imagery. Remote Sens Environ. 2014;140:23–35. York, NY, USA: Association for Computing Machinery
33. Yan D, Huang C, Ma N, Zhang Y. Improved landsat-based water 53. H. Ferdous, T. Siraj, S. J. Setu, M. M. Anwar, and M. A. Rahman.
and snow indices for extracting lake and snow cover/glacier in the 2021. Machine learning approach towards satellite image classi-
tibetan plateau. Water. 2020;12(5):1339. fication. In Proceedings of International Conference on Trends in
34. Wikipedia, “Google earth,” Accessed 6 january 2024, wikipedia, Computational and Cognitive Engineering. Springer, Singapore
the free encyclopedia. [Online]. https://en.wikipedia.org/wiki/ 54. Ahmed M, Seraj R, Islam SMS. The k-means algorithm: a com-
Google_Earth prehensive survey and performance evaluation. Electronics. 2020.
35. Khalid HW, Khalil RMZ, Qureshi MA. Evaluating spectral indices https://doi.org/10.3390/electronics9081295.
for water bodies extraction in western Tibetan plateau. Egyptian J 55. Napoleon D, Pavalakodi S. A new method for dimensionality
Remote Sens Space Sci. 2021;24(3):619–34. reduction using k-means clustering algorithm for high dimen-
36. Yang J, Du X. An enhanced water index in extracting water bodies sional data set. Intern J Com Applicat. 2011;13(7):41–6.
from landsat tm imagery. Ann GIS. 2017;23(3):141–8. 56. Estornell J, Martí-Gavilá JM, Sebastiá MT, Mengual J. Principal
37. Mondejar JP, Tongco AF. Near infrared band of landsat 8 as water component analysis applied to remote sensing. Model Sci Educat
index: a case study around cordova and lapu-lapu city, cebu, phil- Learn. 2013;6:83–9.
ippines. Sustainable Environment Research. 2019;29:1–15. 57. Abid N, Shahzad M, Malik MI, Schwanecke U, Ulges A, Kovács
38. Wei Q, Jingxuan L, Lin L, Xiao-Wen L. Research on automatic G, Shafait F. Ucl: Unsupervised curriculum learning for water
extraction of water bodies and wetlands on hj satellite ccd images. body classification from remote sensing imagery. Int J Appl Earth
Remote Sens Inform. 2011;4(2011):28–33. Obs Geoinf. 2021;105: 102568.
39. Hidayati IN, Suharyadi R. A comparative study of various indices 58. Li D, Wang S. Cost-effective land cover classification for remote
for extraction urban impervious surface of landsat 8 oli. Forum sensing images. J Cloud Comp. 2022. https://doi.org/10.1186/
Geografi. 2019;33(2):162–72. s13677-022-00335-0.
40. EOS, “Ndwi analysis,” \{https://e os.c om/m ake-a n-a nalys is/ 59. Adejoke AO, Badaru YU. Accuracy assessment of pixel-based
ndwi/}, 28 may 2023. image classification of kwali council area, abuja, nigeria. J Nat
41. Chen G, Jin R, Ye Z, Li Q, Gu J, Luo M, Luo Y, Christakos G, Sci Res. 2014;4:133–40.
Morris J, He J, et al. Spatiotemporal Mapping of Salt Marshes in 60. Sibaruddin HI, Shafri H, Pradhan B, Haron N. Comparison of
the Intertidal Zone of China during 1985–2019. J Remote Sens. pixel-based and object-based image classification techniques in
2022;2022:9793626. extracting information from uav imagery data. IOP Conf Ser Earth
42. Xu H. A study on information extraction of water body with the Env Sci. 2018. https://doi.org/10.1088/1755-1315/169/1/012098.
modified normalized difference water index (mndwi). J Remote 61. Lilay M, Taye G. Semantic segmentation model for land cover
Sen Beij. 2005;9(5):595. classification from satellite images in gambella national park,
43. Lu L, Guo H, Wang C, Pesaresi M, Ehrlich D. Monitoring bidec- ethiopia. SN Appl Sci. 2023;5:76.
adal development of urban agglomeration with remote sensing 62. Hu F, Xia G-S, Wang Z, Zhang L, Sun H. Unsupervised feature
images in the Jing-Jin-Tang area, China. J Appl Remote Sens. coding on local patch manifold for satellite image scene classifica-
2014;8(1):084592. tion. IEEE Geosci Remote Sens Sympos. 2014;2014:1273–6.
44. Kebede TA, Hailu BT, Suryabhagavan KV. Evaluation of spec-
tral built-up indices for impervious surface extraction using
SN Computer Science
63. Cheriyadat AM. Unsupervised feature learning for aer- based lenet-5 architecture. Infrared Phys Tech. 2022. https://doi.
ial scene classification. IEEE Trans Geosci Remote Sens. org/10.1016/j.infrared.2022.104470.
2014;52(1):439–51.
64. Ma Z, Liu Z, Zhao Y, Zhang L, Liu D, Ren T, Zhang XF, Li S. Publisher's Note Springer Nature remains neutral with regard to
An unsupervised crop classification method based on principal jurisdictional claims in published maps and institutional affiliations.
components isometric binning. Intern J Geo Inform. 2020. https://
doi.org/10.3390/ijgi9110648. Springer Nature or its licensor (e.g. a society or other partner) holds
65. Yan S, Zhao JF, Zhao JL, Li QZ. A method for image classifica- exclusive rights to this article under a publishing agreement with the
tion based on kernel pca. Intern Confer Machine Learn Cyber. author(s) or other rightsholder(s); author self-archiving of the accepted
2008. https://doi.org/10.1109/icmlc.2008.4620498. manuscript version of this article is solely governed by the terms of
66. Fırat H, Asker ME, Bayindir M, Hanbay D. Spatial-spectral clas- such publishing agreement and applicable law.
sification of hyperspectral remote sensing images using 3d cnn
SN Computer Science

s42979-024-03380-5

Uploaded by

Copyright:

Available Formats

s42979-024-03380-5

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

s42979-024-03380-5

Uploaded by

Copyright:

Available Formats

SN Computer Science (2024) 5:1073

Multispectral Satellite Data Classification Using Rank Correlation

Received: 9 June 2023 / Accepted: 1 October 2024

Introduction transforming multi-band raster imagery into a single-band

designed to categorize pixels, based on a labeled training Objectives

Fig. 1 a Karnataka District b TA1 c TA2

Fig. 2 Reference Image of [a

Table 3 TP, FN, FP and TN values of four class classification

Precision The ratio of predictions as the positive class was positive TP

Fig. 3 The architecture of the proposed methodology

Fig. 4 Spectral graph for the 6

Table 5 Spectral indices Spectral‑Spatial Feature Extraction

Fig. 5 Spectral graph for the 11

hand, pixel pairs from different classes tend to have correlation

determined by the likelihood of their values changing together

Formulation of PSI matrix − 1 to 1, as described in Equation 10 . Standardization was

Results and discussion

spectral bands) and Rraster (a combination of RSbands and RI bands,

in Table 8. In this section, we thoroughly analyze and compare

the performance of these classification approaches against the

proposed method using various evaluation metrics such as pre-

Coefficient (KC). The classification accuracy of the proposed

methods is summarized in Table 9, and the comparative perfor-

is depicted in Figs. 9 and 10 for TA1 and TA2, respectively.

formance. It achieved high precision and recall for the "Water"

up" class exhibited low precision, with values of 0.29 (RSbands,

separation. The recall and F1-scores for the "Built-up" class

ing some misclassification of pixels in both feature sets across

(RSbands, TA1), 70.72% (Rraster, TA1), 60% (RSbands, TA2), and

63.20% (Rraster, TA2). The Cohen’s Kappa Coefficient values,

Table 8 Classification accuracies of different approaches using Rraster feature set

and KC of 0.738 compared to the proposed approach that

and a KC of 0.5070 for TA1, and an OA of 61.83% with

evant features. In the third configuration, we retained both

RSbands and RI bands but excluded PCA. This resulted in an

with a KC of 0.5016 for TA2. Although this approach per-

formed slightly better than using RSbands alone, the results

PCA in enhancing classification accuracy when using

both multispectral bands and spectral indices together.

and a KC of 0.7225 for TA1, and an OA of 83.92% with

However, the performance still fell short compared to the

the fifth approach, both PCA and RSbands were excluded,

KC of 0.6977 for TA2. While the spectral indices alone

scored its importance for achieving superior classifica-

that each component of the proposed approach plays a

The configuration incorporating all three elements con-

sistently outperformed other configurations. This proves

classification results, highlighting the effectiveness of the

Table 10 Analysis of the Approach 1 Approach 2 Approach 3 Approach 4 Approach 5

Conclusion technique for mapping surface built-up areas using landsat 8

You might also like