Proceedings of the 5th International Conference on Smart Systems and Inventive Technology (ICSSIT 2023)

IEEE Xplore Part Number: CFP23P17-ART; ISBN: 978-1-6654-7467-2

Detection of Land Cover Changes using

Satellite Image Classification Technique
2023 5th International Conference on Smart Systems and Inventive Technology (ICSSIT) | 978-1-6654-7467-2/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICSSIT55814.2023.10060992

Puligadda Pavan B Shanmukha Sai Varma Koyagura Asish

B. Tech, Information Technology B. Tech, Information Technology B. Tech, Information Technology
V R Siddhartha Engineering V R Siddhartha Engineering V R Siddhartha Engineering
College College College
Vijayawada, India Vijayawada, India Vijayawada, India
[email protected] [email protected] [email protected]
m

M anne Suneetha
Ph. D, Information Technology
V R Siddhartha Engineering
College
Vi ja ya wa da , Indi a
[email protected]
.in

Abstract— The land cover of an area refers to its terraforming of land in various forms lead to
surface coverage, such as grasslands, housing areas, various changes in land cover. As a result, it is
or water bodies. Changes take place in the structure necessary to keep track of the various changes
of a land cover over time due to various natural
occurring in the land in order to take appropriate
environmental variables and human actions. In
addition, it affects the life cycle of that region and the measures. The current research uses NAIP satellite
plans of human authorities for that piece of land. In images of land cover to detect changes in land
order to detect the changes in a land cover quickly cover over two timestamps. A scheme for the
and accurately, a model that detects the changes over detection of changes in land is designed by
two timestamps are required. Using satellite NAIP studying the land patterns in the selected area. This
images, this paper uses two-time stamps to detect scheme employs the object-based image
land cover changes over a region of land. classification technique to classify various types of
land cover and determine the differences between
Keywords — Remote S ensing, Change Detection, the land covers in the two timestamps by
Image Classification, Image S egmentation, Object- comparing their classified images.
Based Image Classification, Land cover.
The images used are from the Natural
Agriculture Imagery Program (NAIP) due to their
high resolution of the remotely sensed images of
I. INTRODUCTION land cover, which gives us a more accurate
assessment of the data. The NAIP images perform
well with the object-based image classification
Land cover is the surface cover of the land technique, allowing us to maintain a precise scale
upon which the living beings survive, such as and low error value for detecting changes in land
grassland, agricultural fields, houses, and water cover. Landcover changes detection is hugely
bodies. This land cover undergoes changes with the helpful in land surveying and also in disaster
passage of time due to various reasons. The management and recovery. It provides us with
constant changes in land cover affect the Earth’s thematic maps to understand the surface of earth
geographical and weather patterns by changing easily.
territorial and worldwide environmental patterns
and affecting people’s lives in that area. For II. LITERATURE REVIEW
example, the urbanization of land leads to changes
in the grasslands, forests, and urban areas of a land
cover over a time period. Also, the land use and Land Cover Change detection is defined as the

978-1-6654-7467-2/23/$31.00 ©2023 IEEE 1499

Authorized licensed use limited to: Zhejiang University. Downloaded on July 17,2023 at 12:18:41 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the 5th International Conference on Smart Systems and Inventive Technology (ICSSIT 2023)
IEEE Xplore Part Number: CFP23P17-ART; ISBN: 978-1-6654-7467-2

process by which the differences in the state of a method for classification. The bands of multi-
land cover are identified at different times or over spectrum are classified into various objects, and an
different time periods [1]. The techniques for analysis that could be obtained from the bands of
detecting these changes in the land are based on multi-spectrum constitutes the n-dimensional
satellite images of the land cover. These methods feature space for the classification of data.
are based on pixel-based change detection (referred Following that, data inputs for object-based land
to as PBCD) and object-based change detection cover classification are investigated and discussed.
(referred to as OBCD). The pixel-based techniques The detection of changes from testing data for two
are used for low- and medium-resolution images, time series of land cover is presented. Therefore,
and Principal Component Analysis (PCA) is used assessments to help improve the accuracy of the
for change detection [2]. It is preferable to consider result of the land cover classification are presented.
both texture and intensity differences when The utility of this method of classification is that it
detecting changes rather than relying solely on a performs detection using objects in the classified
single feature [7]. The introduction of high- data. This classification method doesn’t require any
resolution remotely sensed spatial imagery paves input or user-defined parameters for detection, such
the way to apply change detection with image as thresholds. It operates mechanically.
analysis using object-based techniques (that is,
OBCD). Object-based image analysis outperformed
traditional pixel-based analysis for detecting
changes in a landscape. To represent objects in
OBCD, images in the remote-sensed data are
segmented into "segments" based on their
attributes, either spatially or spectrally. Typically,
objects defined in the land cover can either (1) be
compared based upon the features of the objects in
the land cover [3] or (2) be compared after an
object-based classification [4].
Landcover changes detection has various
advantages that range from land surveys done
through Government officials to disaster
management and recovery due to natural
calamities. That is why a good changes detection in
landcover is essential in the present and in
upcoming years. There’s also a way of improving
the already existing methods to achieve good
changes detection in landcover. An Ensemble
Learning approach can be used to combine the
results of various classifiers and yield the most
promising result. This paper uses Random Forest
classifier for object-based classification of images.
Random Forest is less sensitive to training input
and unlike decision trees it has a higher success
rate.

III. DESIGN METHODOLOGY

In the proposed method, the dataset

(multispectral images) is extracted from the USGS
website. A sample NAIP image of a place between
two different timestamps is took. The segmented
images are visualized using the QGIS software tool
and training data is generated manually in this
QGIS tool. Overall and class-wise accuracies are
found based on the classification. This method
using remote-sensed data is very useful compared Figure-1: Design M ethodology
to manually gathering data for changes detection.
This method breaks down groups of data
As a result, we can now proceed with our
containing pixels that represent the objects of land
research with improved knowledge and a solid
cover in QGIS software. The procedure is
constructed upon a supervised, most likely accurate

978-1-6654-7467-2/23/$31.00 ©2023 IEEE 1500

Authorized licensed use limited to: Zhejiang University. Downloaded on July 17,2023 at 12:18:41 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the 5th International Conference on Smart Systems and Inventive Technology (ICSSIT 2023)
IEEE Xplore Part Number: CFP23P17-ART; ISBN: 978-1-6654-7467-2

foundation in land cover object-based C. United States Geographical Survey (USGS)

classification.
The United States Geological Survey
IV. DATASET & TO O LS abbreviated USGS and formerly simply known as
the Geological Survey, is a scientific agency that
A. Dataset studies landscapes. Scientists of USGS research
about the land, its resources present naturally in it,
and the natural threats and disasters that causes
The dataset is extracted from the USGS menaces. They excel in various research studies
website. The images used for land cover like geography, hydrology, biology etc. This is a
classification are NAIP images. NAIP images are research organization with no supervision. The
provided in USGS open-source website under the present motto of the USGS, "science for changing
category of aerial imagery. USGS provides us with world".
many satellite images for free, but we prefer NAIP
images for this paper because of their higher
resolution. Sample images taken are shown below. D. Google Colab

Google Colab for is a powerful platform

developed and provided by Google Research. It
allows anybody to write and execute arbitrary
python code through their web browser, and is
well suited to machine learning, data analysis and
education. In this paper it is used for object-based
image classification and also to plot the pie charts
of percentage landcover in the images. The results
can be downloaded straight for this.

Figure-2: Sample Image from 2015 V.RESULTS

A. Segmentation

The multispectral images are segmented with

SLIC algorithm into 70,000 segments each. With
segmentation done the individual objects become
easier to classify.

Figure-3: Sample Image from 2018

B. Quantum Geographic Information System

(QGIS)

QGIS is a desktop platform that provides Figure-4: 2015 Segmented image

data on the various features of land. It is open-
source and free software. It is a tool that supports
various operations for research on land like
modification, observation, etc. This tool aids in the
analysis of land cover. It is designed as a
geographical information system that provides
various operations on land cover classification
along with creating and exploring geographical
maps of land cover. It illustrates various labels and
map overviews needed for the classification.
Figure-5: 2018 Segmented image

978-1-6654-7467-2/23/$31.00 ©2023 IEEE 1501

Authorized licensed use limited to: Zhejiang University. Downloaded on July 17,2023 at 12:18:41 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the 5th International Conference on Smart Systems and Inventive Technology (ICSSIT 2023)
IEEE Xplore Part Number: CFP23P17-ART; ISBN: 978-1-6654-7467-2

Figures 4,5 show the segmented images from D. Land cover Classification
2015 and 2018. The SLIC algorithm used for this
process segments the images based on colour and This is one of the most important parts in the
spatial distance. Hence, a high-resolution image analysis. First identify and label the training
gives us better results as segmentation is performed objects. In this process each spectral band within
more accurately. the segment is associated with a label of land cover
type. Now, that everything is set and ready a
B. Spectral Properties of Image Segments classifier is trained to use it to predict across all
segments in the image. This paper uses random
After the segmentation of image, each segment forests, a popular classification algorithm to
of the image must be quantitatively described with classify the image.
its respective spectral properties. This part of the
workflow calculates descriptive statistics like
mean, max, min, and variance for each band, given
the number of pixels. The random forests algorithm
uses these values to classify the segments into land
cover types.

C. Truth (Training and Test Data)

The classification process in the paper needs Figure-7: Classified image from 2015
some truth data describing various land cover
types. The points below are generated using QGIS
software to show six different land cover classes in
a way that even a common person can unders tand
them. The six landcover classes that were
considered with unique IDs are as following:

1. Water

2. Urban
Figure-8: Classified image from 2018

3. Bare
Figures 7 and 8 are the classified images
output by the Random Forest classifier, for which
4. Road training and testing data are taken as shown in
Figure 6. The random forest algorithm is a very
5. Meadow (Grassland) comprehensive supervised machine learning
technique that is widely used for various types of
6. Trees classifications. Hence, we can get a very precise
classification of data and a good, trained model
using this technique.

E. Accuracy Assessment of Classification with

Confusion Matrix

The assessment of accuracy is an

important aspect of any classification. In this paper,
the test data set is loaded and converted to a raster
format, so it is compatible with the generated
predictions. The predicted values are then simply
queried from the locations where test data exists.
Finally, a confusion matrix is generated from all
Figure-6: Truth Data
the corresponding values. The classification must
be done with the utmost care, because if it’s not
Figure-6 depicts the training data generated right, then the whole point of classification is
manually in QGIS software; each colored dot worthless. Here, as an accuracy assessment, a
represents a unique class. confusion matrix is generated. Load the test data

978-1-6654-7467-2/23/$31.00 ©2023 IEEE 1502

Authorized licensed use limited to: Zhejiang University. Downloaded on July 17,2023 at 12:18:41 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the 5th International Conference on Smart Systems and Inventive Technology (ICSSIT 2023)
IEEE Xplore Part Number: CFP23P17-ART; ISBN: 978-1-6654-7467-2

set that was created in QGIS (as a shape file) and water, 0.75 for urban, 0.86 for bare, 0.77 for road,
convert it to raster format so it is compatible with 0.90 for meadow, and 0.92 for trees.
the generated predictions. Finally, generate the
confusion matrix from the corresponding values. F. Change Detection

Table 1: Tabulated Confusion M atrix 2015 Image OpenCV is used to detect changes in images
using object-based image classification techniques .
In this research, we perform color detection to
Land 1 2 3 4 5 6 Correct differentiate the different object colors in the
Cover sampled classified image of the land covers in the two
ID images. Previously classified images from 2015
1 6 0 0 0 0 0 6 and 2018 are used to perform change detection. A
2 0 27 0 0 0 0 27 simple pie chart visualizing the percentage of each
3 0 0 25 3 0 0 25 landcover is enough to detect the percentage
4 0 0 2 31 1 0 31 change in each landcover. To plot these pie charts,
5 0 0 0 0 23 1 23 the images must be in array format. The pie charts
6 0 1 0 0 0 19 19 contain the hex color representation of classified
images (RGB). This paper used K-means clustering
to arrange colors in different clusters.
[ ]
 (1)

[ ]  (2)

[ ]  (3)

We calculated the overall accuracy for the

2015 classified image from Equation (1) as
94.22%. Precision of water is 1, precision of urban
is 1, precision of barren is 0.89, precision of road is
0.91, precision of meadow is 0.95, and precision of
trees is 0.95, according to Equation (2). From Figure-9: Landcover Pie chart for 2015
Equation (3), recall values are calculated as 1 for
water, 0.96 for urban, 0.92 for bare, 0.91 for road,
0.95 for meadow, and 0.95 for trees.

Table 2: Tabulated Confusion M atrix 2018 Image

Land 1 2 3 4 5 6 Correct
Cover sample
ID
1 5 0 0 0 0 0 5
2 0 41 2 1 0 2 41
3 0 2 52 4 0 0 52
4 0 8 3 17 0 0 17
5 0 2 3 0 10 0 10
6 0 1 0 0 1 26 26
Figure-10: Landcover Pie chart for 2018

We calculated the overall accuracy for the

2015 classified image from Equation (1) as
84.35%. From Equation (2), precision of water is 1,
precision of urban is 0.89, precision of barren is
0.89, precision of road is 0.6, precision of meadow
is 0.67, and precision of trees is 0.92. From
Equation (3), recall values are calculated as 1 for

978-1-6654-7467-2/23/$31.00 ©2023 IEEE 1503

Authorized licensed use limited to: Zhejiang University. Downloaded on July 17,2023 at 12:18:41 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the 5th International Conference on Smart Systems and Inventive Technology (ICSSIT 2023)
IEEE Xplore Part Number: CFP23P17-ART; ISBN: 978-1-6654-7467-2

Finally, this paper provides with change detection

maps which show strong, obvious, subtle a changes
and unchanged spots.

In the present research, when the results of the

images of both the time stamps and the sample
images are compared, it can be stated that object-
based classification of the image was useful in
representing the percentage change in each
landcover over time. Present paper uses Random
Figures 9 and 10 show the percentage of each Forest for object-based image classification. From
type of land cover type in the classified images of observations, the proposed technique is very useful
time stamps. Blue is water, red is urban, black is in the detection of changes in large pieces of land
roads, green is trees, yellow is meadow, and orange cover. Due to classification we’re provided with
is bare land. By understanding the differences in thematic maps, which can be used as inputs to
the two pie charts, the establishment of the detect percentage of each landcover present at a
landcover change chart for the time stamps is as particular timestamp. Finally, a tabular form of
shown below. percentage change occurred in every landcover
classified, is obtained.
Table 3: Changes Table for the Two Timestamps
In the present paper changes in each landcover
Land can be known easily. Whereas the paper using
Cover/ 2015 2018 % Change in Land Rotation Forest and Coarse-to-Fine uncertainty
Year Cover analyses couldn’t provide the changes in each
landcover, it only provides us with changes map.
Water 49 51 2 By comparing both works, the current paper has an
advantage due to image classification. Because of
Forest 12 9 3 image classification, changes in each landcover
were easily identified in the present paper.
Meadow 14 5 9
VII. CONCLUSION & FUTURE WORK
Bare Land 19 27 8
The introduction of object-based classification
Roads 2 1 1 into the classification of remote-sensing data of
land cover images resulted in greater changes in
Urban 4 7 3 detection methods than previous methods. While
Area identifying changes, NAIP images belonging to the
2015 and 2018 years from Bear Lake have been
used. Only because of object-based classification
As shown in Table 3, an object-based we’re provided with good thematic maps, which
classification technique can provide a clear change are easily understood by people to distinguish
detection result in an area by quantifying the between different landcovers. The pie charts
changes. provide with percentage of each landcover that’s
present in the NAIP images. A manual table to
label the percentage changes in each landcover
VI. RESULT ANALYSIS
from the classified images is tabulated.

In the paper “A Novel Change Detection In the present research the method was able to
Approach for Multi-Temporal High-Resolution classify the objects, meaning landcovers in the
Remote Sensing Images Based on Rotation Forest
images quite well. Many advancements in the
and Coarse-to-Fine Uncertainty Analyses” [2], a present discussed field of study are going on. One
change detection method for high resolution remote promising advancement can be achieved by using
sensing images had been discussed. It is based on
ensemble learning models. Object-based landcover
Rotation Forest and coarse-to-fine uncertainty classification in this research is only suitable for
analyses. This paper suggests that deciding the detecting large plots of land. For small plots of
optimal segmentation scale is the key for change
land, the number of objects and texture features is
detection information. Here multiple classifiers are ridiculously high and highly complex for
used, and the results obtained are integrated using convenience, during which accuracy is also put into
Majority Voting approach to provide results.
question. Hence, a better technique should still be

978-1-6654-7467-2/23/$31.00 ©2023 IEEE 1504

Authorized licensed use limited to: Zhejiang University. Downloaded on July 17,2023 at 12:18:41 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the 5th International Conference on Smart Systems and Inventive Technology (ICSSIT 2023)
IEEE Xplore Part Number: CFP23P17-ART; ISBN: 978-1-6654-7467-2

improved or invented on the basis of these differences”, 2nd International Asia Conference on Informatics
in Control, Automation and Robotics, April 2010.
techniques for better evaluation and research. This
is the focus of our work in the future. [8] Mengyuan Liu, Yujiang Liu, "A novel high resolution
remote sensing image change detection algorithm based on
image fusion and fuzzy clustering models", 2016 International
VIII. REFERENCES Conference on Inventive Computation T echnologies (ICICT ),
vol.3, pp.1-6, 2016.
[9] Sigit Adinugroho, Dorothée Vallot, Pontus Westrin, Robin
[1] A. Singh; “Review article digital change detection Strand, "Calving events detection and quantification from time-
techniques using remotely-sensed data”, International Journal of lapse images in T unabreen glacier", 2015 International
Remote Sensing, vol. 10, issue 6, pp. 989-1003, June 1989. Conference on Information & Communication Technology and
[2] Wenqing Feng, Haigang Sui; Jihui T u; Weiming Huang Systems (ICT S), pp.61-66, 2015.
&Kaimin Sun (2018); “A Novel Change Detection Approach for [10] Wen-jie WANG, Zhong-ming ZHAO and Hai-qing ZHU,
Multi-T emporal High-Resolution Remote Sensing Images "Object-oriented Change Detection Method Based on Multi-
Based on Rotation Forest and Coarse-to-Fine Uncertainty scale and Multi-Feature Fusion", Urban Remote Sensing Event,
Analyses”, International Journal of Remote Sensing, vol. 39, pp. 1-5, May 2009.
issue 22, pp. 7998-8021, November 2018.
[11] D. Lu, P. Mausel, E. Brondizio and E. Moran, "Change
[3] Antoine Lefebvre; Thomas Corpetti; Laurence Hubert-Moy; detection techniques", International Journal of Remote Sensing,
“ Object-oriented approach and texture analysis for change pp. 2365-2401, June 2004.
detection in very high-resolution images”, IEEE International
Geoscience and Remote Sensing Symposium, vol. 4, February [12] Shikha Sharda, Mohit Srivastava, "Object-Based Glacier
2009. Surface Change Detection using Multispectral Satellite Data",
2020 IEEE 7th Uttar Pradesh Section International Conference
[4] T im De Chant; Maggi Kelly; “Individual object change on Electrical, Electronics and Computer Engineering (UPCON),
detection for monitoring the impact of a forest pathogen on a pp.1-5, 2020.
Hardwood forest”, Photogrammetric Engineering and Remote
Sensing vol. 75, issue 8, pp. 1005-1013, August 2009. [13] B.A. Robson, D. Hölbling, C. Nuth, T . Strozzi and S.O.
Dahl, "Decadal Scale Changes in Glacier Area in the Hohe
[5] M. Hussain, D. Chen, A. Cheng, H. Wei and D. Stanley, T auern National Park (Austria) Determined by Object -Based
“ Change detection from remotely sensed images: From pixel- Image Analysis", Remote Sensing, vol. 8, pp. 67, January 2016.
based to object -based approaches”, ISPRS Journal of
Photogrammetry and Remote Sensing, Volume 80, p. 91-106, [14] K.V. Mitkari, M.K. Arora and R.K. T iwari, "Extraction of
June 2013. Glacial Lakes in Gangotri Glacier using Object Based Image
Analysis", IEEE Journal of Selected T opics in Applied Earth
[6] Dyah R. Panuju; David J. Paull; Amy L. Griffin: “Change Observations and Remote Sensing, vol. 10, no. 12, pp. 5275 -
Detection T echniques Based on Multispectral Images for 5283, August 2017.
Investigating Land Cover Dynamics”, Remote Sensing, vol. 12,
issue 11, p. 1781, June 2020. [15] K. Kumar, R.K. Dumka, M.S. Miral, G.S. Satyal and M.
Pant, "Estimation of retreat rate of Gangotri Glacier using rapid
[7] Lingcao Huang; Guo Zhang; Yang Li; “An object -based static and kinematic GPS survey", Current Science, vol. 94, pp.
change detection approach by integrating intensity and texture 258-262, January2008.

978-1-6654-7467-2/23/$31.00 ©2023 IEEE 1505

Authorized licensed use limited to: Zhejiang University. Downloaded on July 17,2023 at 12:18:41 UTC from IEEE Xplore. Restrictions apply.