Professional Documents
Culture Documents
Serbia
Abstract—this paper presents methods for land cover change detection analysis. Land use - refers to the
change detection, using object based image analysis. activity, economic purpose, intended use and/or
Considering that huge area in Serbia is covered with forest, management strategy placed on the land cover type(s) by
this study is of great value for tracking changes over time. human agents or land managers. Changes in intent or
The research involved collecting data. The data used in this management practice likewise constitute land use change.
study are optical remote sensing images taken by Landsat When used together the phrase Land Use/Land Cover
TM satellite and Landsat ETM+ satellite. These images generally refers to the categorization or classification of
were processed in order to classify four main land cover human activities and natural elements on the landscape
classes, namely: forest, agricultural, urban and water areas. within a specific time frame based on the established
The images are analyzed in terms of change detection. The scientific and statistical methods of analysis of
results show the type and the place of land-use changes appropriate source materials. Land use and land cover
occurred in these main four classes during those years. change has become a central component in current
Change analyses have been executed based on object image strategies for managing natural resources and monitoring
analysis, and results are given in raster and vector format. environmental changes. Every parcel on the Earth surface
is unique in the term of what exists on that parcel [2, 3].
I. INTRODUCTION
The Earth surface is constantly and rapidly changing in
Remote sensing is defined as the acquisition of data many ways. In most parts of the world, the process of soil
using a remotely located sensing device, and the surface changing is very dynamic. With the exception of
extraction of information from the data [1]. In the narrow seasonal rainfall effects, rapid changes of temperature,
sense, remote sensing includes analysis and interpretation rainfall and drought, man has affected the vegetation by
of various images which represent the Earth surface. changing natural vegetation areas to arable land and vice
These images can be taken from the ground, air or space. versa. On the other hand, there are other dynamic aspects
Electromagnetic energy sensors acquire data in the such as highly volatile pressure of mountain ranges,
way various earth surface features emit and reflect forest fires, floods etc.
electromagnetic energy, and these data are analyzed to Change detection can be defined as the process of
provide information about the resources under identifying differences in the state of an object or
investigation. phenomenon by observing it at different times [4].
Shortly, image represents electromagnetic energy Change detection is an important application of remote
which is registered with a sensor. A first step in the sensing technology. It is a process of determining the
process of studying images is the process of separating changes of specific features (object or phenomena of
areas with different properties and grouping ones with interest) within a certain time interval. Change detection is
similar properties. Realizing the differences in these very important part of the monitoring and control of
properties and the extraction of certain areas according to natural resources and urban development due to
these differences represent analysis of the image. providing the spatial distribution of features and
Identified differences in these properties emphasize the qualitative and quantitative information of feature
meaning in terms of science discipline for whose changes. It involves the type, distribution and quantity of
purposes these methods are used. Interpretation of images changes, that is the ground surface types, boundary
represents an explanation of the differences in the changes and trends before and after the changes.
properties and the determination of its significance. The main aim of this paper is the analysis and the
Therefore, interpretation represents the most important interpretation of satellite images by using object-based
part in the process of remote sensing. image analysis methods, with the goal of identifying
When talking about terms such as detection on the changes in the field in terms of thematic transition from
Earth surface or change detection, it is important to one class (land cover) into another.
explain the meaning of terms land cover and land use.
The terms: land use and land cover are often used II. OBJECT BASED IMAGE ANALYSIS
interchangeably, but each term has its own unique
meaning. Land cover - refers to the characteristics and A. Image segmentation
surface cover of the Earth surface, as represented by land Segmentation refers to the process of partitioning a
cover category: vegetation, forest, water, bare earth, digital image into non intersectable regions, in the way
impervious surfaces and other physical features of the that each region is homogeneous and that union of two
land. Identification of land cover establishes the baseline adjecnt regions is also homegeneous [5]. Segmentation is
information for activities such as thematic mapping and a way of partitioning raster images into segments based on
pixel values and locations. Pixels which are spatially classification techniques provides the effective
connected and have similar values are grouped in a single identification of deforestation area in the zone of the
segment. Process of image classification based on the Brazilian Amazon. Also in the paper "Image segmentation
creation of the regions, ie. segmentation, avoids problems for humid tropical forest classification in Landsat TM
of mixed pixels and provides more valuable parameters data” [11] the authors used Landsat TM image and applied
for the classification (texture, shape, pixel environement, a mixture of edge detection and segmentation methods
distance and orentation) then the usual pixel value (light) using region growing/merging algorithm. Pixel-by-pixel
[6]. classification which was applied after segmentation
Object extraction can be accessed by using classical or produced acceptable results.
fuzzy mathematical methods. First group of these In general, object-oriented classification process may
techniques is based on the value of the histogram be divided in two main working steps; first one would be
threshold, edge detection, relaxation, semantics and segmentation and second one would be classification of
syntax. Fuzzy mathematical methods are also based on the segments based on the previously acquired knowledge.
edge detection methods, thresholding and relaxation. Congalton and Macleod [12] defined a list of four
The histogram threshold method separates image in phenomena that are important for change detection when
subregions, based on the threshold value for each the status of natural resources is monitored:
subregion. If the image is shown in two levels, then the 1. detection of the occurred changes;
image separation is done in two regions (black objects and 2. identification of the nature of changes;
white background). If image histogram shows different
jump values (picks), it means that image is created from 3. measurement of changed area;
different regions, and then the different threshold value is 4. creation of a spatial pattern of occurred change.
necessary for each one of the regions [7 and 8]. Relaxation The basis for the usage of remote sensing data for
is an iterative approach to the segmentation in which each change detection purposes was the fact that changes in the
pixel is classified in parallel process. Decisions that are land cover cause the changes in the values of radiation
made in the neighboring points in current iteration are (reflection) that can be remotely sensed. Change detection
used later and combined for decision making in the techniques - with the help of satellite imagery have
following iteration process. Neural network methods are become numerous, which caused the versatility in digital
also used for image segmentation [5]. Segmentation data processing and computer skills development.
process of images with different values of gray color is Numerous variations of digital change detection
based on the edge detection of different regions and on the techniques have been developed over the past twenty
location of sudden, rapid and important changes of years. Altogether eleven different change detection
intensity. algorithms has been affirmed so far [4, 13-20]. This
Few methodological approaches are used for image includes:
edge detection. Some of them are explained in detail in 1) Mono-temporal change delineation;
work of Pal and Sankar [5].
2) Delta or post classification comparisons;
Fuzzy methods are based on the fuzzy sets which use an
imprecise knowledge to define a certain event. This 3) Multidimensional temporal feature space
approach allows the quantification of unreliable facts. analysis;
Such uncertainties arise from a lack of information, such 4) Composite analysis;
as incomplete, inaccurate, wrongly defined or 5) Image differencing;
contradictory statements [5]. 6) Multitemporal linear data transformation;
Most of above mentioned methods have been 7) Change vector analysis;
developed for the main usage in the field of medicine and
telecommunications, for the use of panchromatic images 8) Image regression;
[5]. The application of segmentation process in the field of 9) Multitemporal biomass index;
remote sensing was presented recently; first commercial 10) Background subtraction;
software was introduced in 2000 [9]. 11) Image rationing.
B. Related work In some situations, land use and land cover changes can
There are advantages in the classification process based cause greater damage than benefit to the environment in
on the object-oriented approach to the standard approach the economic and social terms. Therefore, these data are
which is based on the pixel value. Apart from the spectral primarily needed for planners, so they could react timely
information, objects contain additional attributes (e.g. and prevent some new, negative consequences caused by
shape, texture, etc) which can be used for the purpose of the changes. These data are used as an input in modeling
classification. In addition, image segmentation produces and predicting future changes.
homogeneous objects in the image, thus avoiding the Tuncay Kuleli in his work [21] used Landsat satellite
effect of „snow” (salt-and-pepper effect). images to understand the changes on the coast of
Object classification approach has been used in solving Mediterranean part of Turkey. Two methods were used,
some specific problems in the forestry. Authors of the supervised classification of land-cover and comparison of
paper “Using shade fraction image segmentation to the results of two classification processes that were
evaluate deforestation in Landsat Thematic Mapper performed in the images of the same area in two time
images of the Amazon Region” [10] showed that the periods. The second method is named an unsupervised
segmentation algorithm using the region growing/merging classification with the implementation of ISODATA
algorithms in combination with unsupervised algorithm, where six land cover and land use types were
detected: agricultural land, moisture soil, sand, vegetation,
– 232 –
SISY 2010 • 2010 IEEE 8th International Symposium on Intelligent Systems and Informatics • September 10-11, 2010, Subotica, Serbia
shallow and deep water. Comparing the results of basic characteristics of Landsat TM and ETM+ Band
classification showed that there has been a rapid change in characteristic.
structure of the land cover and land use in a time period of
eight years. A transformation of sand and wet areas into TABLE I. LANDSAT TM ETM+ BAND CHARACTERISTICS
the agricultural land was identified, showing an alarming
Type of Date of
issue in such a short period of eight years. This change image acquisition
Some characteristic
indicates the disappearance of wet lands and the
disappearance of wetland landscapes, which can lead to Seven spectral bands, including a thermal
band :
the unforeseeable consequences for the ecosystem of the • Band 1 Visible (0.45 – 0.52µm) 30m
observed area. • Band 2 Visible (0.52 – 0.60µm) 30m
22-08-1986 • Band 3 Visible (0.63 – 0.69µm) 30m
In the paper "Change detection of vegetation cover Landsat
09-08-1987 • Band 4 NIR (0.76 – 0.90µm) 30m
using multi-temporal remote sensing data and GIS TM
17-06-1991 • Band 5 NIR (1.55 – 1.75µm) 30m
techniques" from 2008, the authors Adia and Rabi [22] • Band 6 Thermal (10.40 – 12.50µm) 120m
also used Landsat satellite images in change detection • Band 7 Mid IR (2.08 – 2.35µm) 30m
Ground Sampling Interval (pixel size): 30m
process. There were several objectives in this study: the reflective 120m thermal
identification of vegetation cover and its spatial
distribution, the analysis of spatial-temporal changes in Eight spectral bands, including a pan and
thermal band:
vegetation cover, the calculation and the analysis of • Band 1 Visible (0.45 – 0.52µm) 30m
normalized difference vegetation index NDVI, description • Band 2 Visible (0.52 – 0.60µm) 30m
of a vegetation reflection and finally the production of • Band 3 Visible (0.63 – 0.69µm) 30m
NDVI maps. • Band 4 NIR (0.76 – 0.90µm) 30m
• Band 5 NIR (1.55 – 1.75µm) 30m
Landsat 20-08-2000
Three primary methods of data analysis are applied: ETM + 23-06-2002 • Band 6 Thermal (10.40 – 12.50µm) 60 m
Low Gain / High Gain
1) supervised classification (maximum likelihood • Band 7 Mid IR (2.08 – 2.35µm) 30m
algorithm), where images from 1986 and 2001 were • Band 8 Panchromatic (0.52 - 0.9µm) 15m
classified in order to create maps with thick bushes, less Ground Sampling Interval (pixel size): 30m
dense bushes, agricultural crop areas, grasslands, artificial reflective 60m thermal
Added the Band 6 Low and High gain 60
objects, areas with sparse vegetation and water surface meter thermal bands
areas;
2) calculation of the NDVI for both images in order
to detect the areas with the decrease of vegetation cover With its 30-meter spatial resolution, Landsat TM and
and ETM + had primacy in the study of land use, mapping and
monitoring and change detection for years. Another
3) overlapping of the results. reason for choosing this satellite platform was a selection
In addition to the fact that the results of the study of information, whose access is often limited. The choice
showed a decrease or increase in the areas of mentioned is often a compromise between the target period of data
covers, as well as changes in the NDVI index, this study acquisition, the acquisition interval between two or more
has shown that the techniques of classification and images and the existence of data. Imagery cost is also a
overlapping produce satisfactory results for this type of limiting factor in the selection of data that should not be
studies. Spatial and spectral resolution of Landsat images excluded.
also fully met the needs. In order to get applicable and measurable results of this
In all these mentioned papers, change detection is based study, a reference data was needed. Referring to the fact
primarily on the classification and the analysis of images that the main aim was to detect changes which occurred in
using standard image analysis techniques. All results were a period of time, the reference data had to be related to
obtained using the pixel values as a basis for analysis and two time-points of the same area. As the first reference
classification. Object-based image analysis represents a value, the area of the National Park Fruska Gora of
step forward in the meaning of change detection process, 25475.1413 hectares was chosen. In the imagery of this
especially for change detection in forest areas. area, classification and change detection processes were
performed.
III. THE STUDY AREA AND DATA
In order to begin the process of classification and
Given the fact that Srem represents the part of Serbia change detection, it was necessary to prepare some basic
where agricultural production and forestry constitutes the input data. The data, besides the already mentioned
most important part of economy, this paper might be of Landsat satellite images, which were used in this analysis,
major importance for monitoring changes in the field of are:
land coverage [23]. For analyzing methodology and 1) topographic map scale of 1:100 000 (TK100)
possibilities region of Srem is chosen. The study site lies Novi Sad tags 378, which content of the map corresponds
approximately between 44°35’ and 45°16’ N latitude and to state (1985-1987), showen in Fig. 1;
between 17°60’ and 20°80'E longitude. Remote sensing
data used in this paper are Landsat TM images from 1986, 2) data obtained from the general basis for forest
1987 and 1991 and Landsat ETM+ from 2000 and 2002, management for the National Park Fruska Gora, for the
whose characteristics are specified in Table I. These period (2002-2011).
images are used for classification of four basic types of Mentioned topographic map - TK100 matches the
land cover: water and water bodies, forests, settlements situation for the period 1985-1987. These topographic
and agriculture areas, for the purpose of detection of maps were used for vectorization of the total forest area in
changes for the above mentioned types and the above the park Fruska Gora. Vectorization of the forest area was
mentioned time period. The following table includes the performed by using MicroStation V8 2004 CAD tool.
– 233 –
D. Jovanovic et al. • Object-based Image Analysis in Forestry Change Detection
– 234 –
SISY 2010 • 2010 IEEE 8th International Symposium on Intelligent Systems and Informatics • September 10-11, 2010, Subotica, Serbia
– 235 –
D. Jovanovic et al. • Object-based Image Analysis in Forestry Change Detection
TABLE IV. POSTCLASSIFICATION RESULTS [2] Meyer, W.B. "Past and Present Land-use and Land-cover in the
U.S.A. Consequences", 1995.
Postclassification
[3] Change detection in land use and land cover using remote sensing
Water Artificial Agricultural data and gis. http://www.gisdevelopment.net/thesis/ visited in
Forest area
bodies areas area December 2007.
1986-08- 23031,417 46,855 [4] Singh, A. "Digital Change Detection Techniques Using Remotely
3,692 ha 2393,169 ha
22 ha ha Sensed Data", International Journal of Remote Sensing. Vol. 10,
2002-06- 24417,517 No. 6, p. 989-1003, 1989.
5,949 ha 59,07 ha 992,606 ha
23 ha [5] Pal, N.R., Pal, Sankar K. "A review on image segmentation
Area 12,215 techniques. Pattern Recognition" 26:9, 1277-1294, 1993.
1386,1 ha 2,257 ha 1400,563 ha
difference ha
[6] Blaschke, T., Strobl, J. "What’s wrong with pixels?: some recent
developments interfacing remote sensing and GIS" GIS-
TABLE V. Zeitschrift für Geoinformationssysteme 6, 12-17, 2001.
Reference values Postclassification OBIA [7] Fu, K.S., Mui, J.K. "A survey on image segmentation. Pattern
Recognition" 13, 3-16, 1981.
1986-08-22 22682,4915 ha 23031,417 ha
[8] Ryherd, S., Woodcock, C. "Combining spectral and textural data
2002-06-23 24.287,1604 ha 24417,517 ha in the segmentation of remotely sensed images" Photogrammetric
Area difference 1604,6689 ha 1386,1 ha Engineering and Remote Sensing 62:2, 181-194, 1996.
[9] Schiewe, J., Tufte, L., Ehlers, M. "Potential and problems of
CONCLUSION multi-scale segmentation methods in remote sensing" GIS-
Zeitschrift für Geoinformationssysteme 6, 34-39, 2001.
In the past years, the observation of the Earth became
very important issue and important way of collecting a [10] Shimabukuro Y.E., Batista G.T., Mello E.M.K., Moreira J.C.,
Duarte V. "Using shade fraction image segmentation to evaluate
large amount of data. The data collected by remote deforestation in Landsat TM images of the Amazon Region"
sensing methods increase the value of information which International Journal of Remote Sensing 19:3, 535-541, 1998.
can be seen from the raw, unprocessed data. Detection of [11] Hill, R.A. "Image segmentation for humid tropical forest
changes in environment represents one of the very classification in Landsat TM data" International Journal of Remote
important tasks which did not have an important role in Sensing 20:5, 1039-1044, 1999.
decision making which relates on the quality of life or [12] Macleod & Congalton. "A Quantitative Comparison of Change
environment, until recently. Detection Algorithms for Monitoring Eelgrass from Remotely
The most of methods in change detection use different Sensed Data", Photogrammetric Engineering & Remote Sensing.
Vol. 64. No. 3. p. 207 - 216, 1998.
algorithms for processing of satellite and other images
[13] Coppin, P. & Bauer, M. "Digital Change Detection in Forest
from different time periods. The basis for using remotely Ecosystems with Remote Sensing Imagery", Remote Sensing
sensed data for change detection rely on the fact that the Reviews. Vol. 13. p. 207-234, 1996.
changes in land cover cause the changes in the reflection [14] Maillard, P., Clausi, D.A., "Comparing classification metrics for
which can be remotely sensed by various sensors. labeling segmented remote sensing images", Computer and Robot
During the preparation of this paper, the usage of Vision, 2005. Proceedings, Page(s): 421 – 428.
different techniques of change detection has been [15] Balasubramanian, R. Rajan, S. Doraiswami, R. Stevenson, H., "A
analyzed. Standard techniques of image classification are reliable composite classification strategy", Electrical and
those which have been used until recently, for processing Computer Engineering, page(s): 914-917 vol.2, 1998.
of the satellite and other types of images so that they can [16] John C. Russ, The Image Processing Handbook, Springer-Verlag;
provide adequate results of change detection. The change October 1998.
detection approach for all these techniques was common [17] Spitzer, H.; Franck, R.; Kollewe, M.; Rega, N.; Rothkirch, A.;
Wiemker, R., "Change detection with 1 m resolution satellite and
and it was based on the pixel of image, namely its value. aerial images", Geosciences and Remote Sensing Symposium,
Object orientated approach imposed in past few years as a 2001, Page(s):2256 - 2258 vol.5
solution and a new procedure of classification and image [18] L. Bruzzone and S. B. Serpico, "An iterative technique for the
analyses, in respect to the approach based on the pixel detection of land-cover transitions in multi-temporal remote-
value. Beside the spectral information, objects contain sensing images", IEEE Trans. Geosci. Remote Sensing, vol. 35,
additional attributes (for example shape, texture etc.) pp. 858-867, July 1997.
which are used in the purpose of classification. [19] Radke R.J., Andra S., Al-Kofani O. and Roysan B, "Image change
In the case of direct change detection, specifically in detection algorithms: A systematic survey", IEEE Trans. Image
Processing, 14 (3):291-307, 2005.
analyzed case, this method did not show satisfactory
[20] Shashi Shekar, Hui Xiong (Eds.), Encyclopedia of GIS, © 2008
results. One of the possible reasons for that was coarse SpringerScience, ISBN: 978-0-387-35973-1
spatial resolution of Landsat images which have been
[21] Change detection and assessment using multi temporal satellite
analyzed. The other object method, which implied the image for North-East Mediterranean Coast, Tuncay Kuleli,
classification of land covers with the object approach, and University of Cukurova Faculty of Fisheries Adana, Turkey.
the postclassification later on, resulted as a more accurate [22] Adia S.O, "Change detection of vegetation cover using multi-
method, whether in results of classification, whether in temporal remote sensing data and GIS techniques", Department of
values of detected changes in forest areas. Meteorology, Rabiu A. B Space Physics Lab, Physics Department
Federal University of Technology, Akure Ondo State, India.
Object oriented method of classification resulted as
[23] Dušan Jovanović, "Monitoring Land-Use change area of
very accurate in the usage for the classification of land Vojvodina, Serbia using LANDSAT ETM+ AND TM DATA",
cover. The method of postclassification over the results of 4th International Conference, Bulgaria, InterGeo East 2007.
object classification resulted as a very accurate method of [24] Dušan Jovanović, "A Comparison of object-oriented classification
change detection of forest areas. and standard techniques of image analyses in change detection of
forest areas", msc thesis, Faculty of Technical Sciences, Serbia,
REFERENCES 2010.
[1] Keith R. McCloy "Resource Management Information Systems:
Remote Sensing, GIS and Modelling", Taylor & Francis, 2006.
– 236 –