Journal of Geographic Information System, 2010, 2, 113-119
doi:10.4236/jgis.2010.22017 Published Online April 2010 (http://www.SciRP.org/journal/jgis)
Double Polarization SAR Image Classification Based on
Object-Oriented Technology
Xiuguo Liu1,2, Yongsheng Li2, Wei Gao2, Lin Xiao2
1Faculty of Information Engineering, China University of Geosciences, Wuhan, China
2Wuhan Zondy Cyber T
＆ S. Co., Ltd, Wuhan, China
Abstract
This paper proposed to use double polarization synthetic aperture radar (SAR) image to classify surface fea-
ture, based on DEM. It takes fully use of the polarization information and external information. This paper
utilizes ENVISAT ASAR APP double-polarization data of Poyang lake area in Jiangxi Province. Compared
with traditional pixel-based classification, this paper fully uses object features (color, shape, hierarchy) and
accessorial DEM information. The classification accuracy improves from the original 73.7% to 91.84%. The
result shows that object-oriented classification technology is suitable for double polarization SAR’s high
precision classification.
Keywords: Synthetic Aperture Radar, Image Classification, Object-Oriented, Pixel-Based, DEM
1. Introduction
SAR has great significance because of its imaging capa-
bility in all day and all weather which could make up for
the weakness of optical remote sensing in the application
of land use and dynamic monitoring in cloudy and rainy
areas. Electromagnetic waves are sensitive to the shape,
size, texture, surface roughness, the complex permittivity
of ground objects. Compared with single-Polarization, dual
polarization is more sensitive to the properties of the
ground objects and it’s more suitab le for ground objects’
recognition and land use classification.
Traditional information extraction technology with sin-
gle a pixel as a unit emphasizes too much on part (color
and texture of a single pixel), while ignores the geomet-
ric structure of whole map spot nearby.
Pixel-oriented solution model analyze pixel separately
which has low interpretation accuracy.
Pixel-based classification methods, such as maximum
likelihood and ISODATA, only use the backscatter coef-
ficient of SAR images, and the rich texture, roughness
and other information in images are not applied. The
result of classification is not satisfactory. The smallest
processing unit of object-oriented classification method
in information extraction is no longer the pixel, but the
object with more semantic information of adjacent pixels.
It classifies the remote sensing images in a higher level
in order to reduce the semantic information loss rate as in
the traditional pix el-based classification; so that the clas-
sification results semantic information will be richer [1].
A lot of scholars from home and abroad have studied
the land use and land-cover classification using optical
images based on object-oriented technology. The results
show the accuracy of classification improves significant ly
compared with the traditional classification methods [2,3].
There are also many scholars who have studied the object-
oriented classification using SAR images. [4,5]. However,
the theoretical foundation, model building and other as-
pects in this field are less advanced than optical images
obviously. This paper tries t o conduct the double-pol arized
SAR image classification based on object-oriented tech-
nology, and confirms that the method is also suitable for
the high-precision classification of dual-polarized SAR
images.
2. Object-Oriented
The object-oriented classification idea is to simulate hu-
man cognitive processes. The human brains always put
things in a specific environment when analyzing and iden-
tifying the things. Environment is vital important for
identifying th ings. Human’ s pe rception o f ex tern al th ing s
is unified, including the color, external contour, spacing
and other properties of things. Similarly, in making vis-
ual interpretation of remote sensing images, besides the
differences between the colors, it can identify the ground
objects by texture, shape, adjacency relationships etc. The
object-oriented classification method extracts the homo-

X. G. LIU ET AL. 114
geneous regions before classifying. The specific two-step
process is multi-scale segmentation and classification.
2.1. Multi-Scale Segmentation of Image
The multi-scale Segmentation of SAR image aims to cut
the image into multiple small areas. Each small area has
the same attributes and these regions can be taken as the
ground objects in reality. Image seg mentation shows that
each pixel is an object at the beginning. The process is to
combine the similar and nearby objects into a large new
object. Compared with the pixels, image object has mult i-
characters, such as color, size, shape, uniformity and so
on. Generally, SAR images have macroscopic and mi-
croscopic characteristics. If cut the region with single-
scale, it will result in lots of broken areas. It is not con-
ducive to information interpretation. The information of
different ground objects need to be analyzed in different
space sub-scales.
In the multi-scale segmentation process, each object
layer has a scale. The remote sensing images can be de-
scribed by a variety of the same phase and appropriate
scales that caused by several object layers, rather than a
single scale. The larger the segmentation scale, the larger
the region area in the generated object layers and vice
versa. It will form a multi-scale layers diagram of objects
afte r se ver al times’ seg mentation. The d iagram is showed
as Figure 1.
By using multi-scale segmentation technique, form
image objects at different scales, at the same time gener-
ate the adjacent relation and level inheritance relation
which have impacts on the objects. Different scales of i m-
age segmentation reduce data unit number to be addr es s ed
in the classification, and speed up the classification.
Choose appropriate le vel t o ext ract information between
multiple layers scales after the multi-scale segmentation.
The larger spatial scale surface features can be extracted
by the large split-scale layer, such as rivers, forests, etc.
Similarly, proposing to select the smaller-scale layers for
small split-scale or areas with complex feature.
2.2. Object-Oriented Classification
Object-oriented classification have two mainly methods,
including nearest neighbor distance method and member
Figure 1. Hierarchy network graph of image object.
function method. The nearest distance classification is
similar to supervised classification, which needs to select
the sample. The member function method is based on the
object characteristics with fuzzy logic, which is an un-
certain method of the analysis of things. The polygon ob-
jects after being split do not rigidly belong to a certain
ground object type, but describing the similar degree of
object types and a certain ground object. Fuzzy logic is a
kind of mathematical methods to quantify uncertainty.
Using the member function method requires only a small
amount of feature information to identify the types of
ground objects. The member function method is a regu-
late classification method by imaging the object spec-
trum, shape, texture features and other information.
At the same time, the object-oriented classification can
also use other ancillary information, such as DEM, sta-
tistical information of ground object area, as well as the
known maps of ground obj ect classification. Using DEM
can separate mountains, hills, and plains based on the
terrain trend. Using the statistical information of ground
object area can make supplementary classification accord-
ing to the area.
3. Experiment Result
3.1. Experiment Area Survey
This paper selects a nearby typical area from the Poyang
Lake region, Jiangxi Yongxiu County as an experiment
area. This region is located in northern part of Jiangxi
Province, Poyang Lake , west (long itud e 115°45’-115 °55’,
north latitude 28°59’-29°10’). It is subtropical monsoon
climat e zo ne , f ou r di sti nct i ve s e aso ns a n d s uf fic ie nt sunlight,
abundant rainfall, and is very suitable for the development
of grain, cotton, oil and aquacu lture. Land-u se types ma i nly
include hills, rivers, towns, grasslands, farmlands, as well
as floodplain and other unus ed la nd. (a s sho wn in Figure
2(a))
3.2. Data Processing
Choose SAR data in the Poyang Lake region, Jiangxi
Yongxiu County ENVISAT APP dual-polarized data which
acquired on July 28, 2004 as the experimental data in this
paper. Its polarization is HH, VV and resolution is 12.5
meters (Figure 2(b)). Before the classification of SAR
images, it needs to do pre-processing, including image
calibration, geometric correction and filtering. In order to
increase the amount of information in dual-polarized im-
ages, it increases a band HH-VV based on the original
two polarizations. The results showed that the differences
between surface features are more clearly.
At the same time, this paper uses DEM as an auxiliary
parameter to extract the hilly regions. DEM layer is set

X. G. LIU ET AL. 115
(a) (b)
Figure 2. The local of experiment area (a) optical image, (b) ENVISAT false colour synthetical image red: HH Green: VV
Blue: HH-VV.
as Layer 4 and the elevation of hills in the region is
above 50 meters. It is shown in Figure 3. remote sensing image processing platform. Non-supervised classification conduct “blind” classifi-
cation under the condition that people do not have any
prior knowledge before the classification process, accord-
ing to the statistical characteristics and natural clustering
features of the images. Its classification is only distinguish-
ing different categories, but does not determine the types
of property. Supervised classification is the method of pat-
tern recognition, by selecting the characteristic parameters
to find the characteristic parameters as the decisi o n- ma ki ng
rules before the choice of representative or typical features
in the training area and establishing discriminate function
classification of individual images.
3.3. The Traditional Classification Method
The traditional method is based on the pixel. It only takes
advantage of the pixel values of the backward scattering.
Specific classification methods include unsupervised clas -
sification and supervised classification. The traditional me -
thod completes classification based on ZONDY_SAR radar
Non-supervised classification adopts ISODATA m etho d.
As shown in Figure 4. Unsupervised classification only
recognizes 4 features, while the hills and bare land have
not been correctly id entified. Sup ervised classif ication s am-
pling method is BP neural network. As shown in Figure 5.
Supervised classification can identify most of the su rface
features, but the results have a serious error, that is hills
are assigned to farmland category.
Because the overall scattering characteristics in the
image of the fields and hills are similar, the supervised
classification methods cannot separate the two kinds of
surface features. The external solution is to use DEM,
which will be specifically introduced in the following
description. At the same time, supervised classification
and unsupervised classification has serious “salt and pep-
per” noise to some extent, which is the inherent limita-
tion of traditional classification methods. Figure 3. Auxiliary DEM image.

X. G. LIU ET AL. 116
Figure 4. Unsupervised classification result (ISODATA).
Figure 5. Supervised classification result (BP neural net-
work).
3.4. Object-Oriented Classification
The basic entity of object-oriented classification of is a
meaningful image objects, rather than single pixel. Image
object not only contains scattering information, but also
contextual information, object texture information and som e
auxiliary information. The following experiment is com-
pleted based on the Definiens 7.0 object-oriented in for ma -
tion extraction system software.
3.4.1. Multi-Scale Segmenta t i o n
The scales of multi-scale segmentation of images are 30
and 60. There is a big difference between objects that
obtained by different scales. There are a larger number of
objects in images when the scale is 30 (Figure 6(a)).
Small area such as residents can be separated from sur-
face features such as energy and the surrounding grass-
lands, farmlands, etc, while large-scale surface features
such as water bodies in the segmentation results have been
divided into a number of small objects. It is not con d u civ e
to water extraction.
Large area water can be extracted when the scale is 60
(Figure 6(b)). It can be expressed with a small amount
of polygonal objects and the effect is good. But small
object like residents is merged into the category of grass
which results in a mixed object. This is not beneficial to
the extraction of the residents. Therefore, these features
need to use smaller-scale segmentation.
3.4.2. Cl assification
According to the characteristics of experiment area fea-
tures and the relationship between information and ob-
jects, this paper uses hierarchical classification structure.
It is built on two levels. The divided scale of the first
layer Level1 is 30. It is used for the unused land such as
residents and exposed areas. The divided scale of the se c-
ond layer Level2 is 60, and is used for the hills, water,
farmland, grasslands and other regions. The specific proc-
esses are shown in Figure 7.
The first is the extraction of water. The water color in
the image is a little bit dark. The water body is extracted
by setting the rules. Due to the fact that there are a small
number of residents in some parts of hills, residents are
extracted first. Then hills in the remaining land are ex-
tracted second. This step needs the help of an external
DEM. Lastly, extract residential areas, grasslands and
farmlands in the remaining plain areas. Specific rules are
as follows,
Water: Mean Layer1 < –17
Residents: Standard deviation Layer1 > = 41.5
Hill: Mean Layer4 > 49.5 (DEM)
Farmland: Mean of Inner border Layer2 > 170
Grassland: Ratio Layer 1> = 0.57.
After the above steps, most of the features have been
extracted. The features without division should be cate-
gorized into unused land. Through the cell processing an d
cartographic generalization, the classification results are
shown in Figure 8.
As can be seen from Figure 8, the object-oriented clas-
sification eliminates the “salt and pepper” noise which ex -
ists in the traditional classificatio n and smooth th e results
in detail. The category of features is also more abundant.
With the help of DEM, object-oriented classification can
accurately extract the hills, which cannot be achieved by
traditional classification methods.

X. G. LIU ET AL.117
3.5. Accuracy Evaluation
Choose the results of supervised classification and object-
oriented classification to do accuracy evaluation compari-
son. The two methods are both based on sam ples.
(a)
(b)
Figure 6. Result of Different-Scale Segmentation. (a) scale
30; (b) scale 60.
Figure 7. Processing flow chart of object-oriented classifi-
cation.
Figure 8. Result of object-oriented classification.
Supervised classification is based on a sample of pixel,
while when it comes to object-oriented methods; the sam-
ple is based on image objects shown in Table 1 and Ta-
ble 2.
As shown in the Table 1 and Table 2 which record the
classification accuracy of evaluation results, the overall
classification accuracy and Kappa coefficient of the tra-
ditional supervised classification is 73.7% and 0.654.
While the overall object-oriented classification accuracy
and Kappa coefficient is 91.84% and 0.895, respectively.
This shows that the object-oriented technology improves
the accuracy of SAR image classification significantly.
The traditional method is very easy to mislead hill to
farmland because of the similar scattering characteristics.
By using the external DEM, the object-oriented approach

X. G. LIU ET AL. 118
Table 1. Precision of evaluate supervised classification.
Water Residents Hill GrassPlotCropland Unused land Samples sum User precision %
Water 6465 0 11 0 0 550 7026 92.02
Residents 0 1243 0 54 0 0 1297 95.84
Hill 0 0 1375 0 0 0 1375 100.00
Grassplot 0 0 922 14989 0 251 15911 94.19
Cropland 0 119 7286 1614 5389 0 14659 36.76
Unused land 37 0 0 0 22 935 994 94.16
Samples sum 6502 1362 9594 16657 5411 1736 41262
procretor % 99.43 91.26 14.33 89.99 99.59 53.86
Precision sum = 73 .7% kappa = 0.654
Table 2. Precision of object-oriented classification.
Water Residents Hill GrasslandCropland Unused land Samples sum User precision %
Water 81 0 0 0 0 2 83 97.59
Residents 0 25 1 4 0 0 30 83.33
Hill 0 1 17 0 2 0 20 85.00
Grassland 0 1 1 59 3 0 64 92.29
Cropland 0 0 0 3 28 1 15 87.50
Unused land 0 0 0 0 1 15 16 93.75
Samples sum 81 27 19 66 34 18 245
procretor % 100 92.59 89.47 89.39 82.35 83.33
Precision sum = 91 .84% kappa = 0.895
separate farmland and hills through the elevation infor-
mation. This increases the classification accuracy of the
hills greatly, and this method cannot be achieved by tra-
ditional classification methods. As is shown from the ac-
curacy of evaluation results, the producer accuracy of the
hills in supervised classification is only 14.33%. The ac-
curacy of supervised classification results is not high
mainly due to the inaccuracy of this indicator.
Although the classification accuracy of some indica-
tors of the object-oriented technology is lower than the
supervised classification, this does not affect the overall
classification accuracy.
4. Result and Argumentation
This paper uses external DEM, and classifies dual-polari-
zation SAR images by using object-oriented technology.
The classification accuracy is 91.84%, while the tradi-
tional pixel-based classification accuracy is only 73.7%.
1) The traditional classification can only make use of
image pixel gray values, rather than other ancillary in-
formation, such as external DEM elevation in this pap er,
except its own image pixel gray value. The object-o ri en ted
classification not only takes the multiple features into ac-
count, but also can judge by making use of external aux il-
iary information. The results showed that object-oriented
classification method can estimate the type of ground ob-
ject effectively.
2) The results of object-oriented classification eliminate
the inherent “salt and pepper” noise of pixel-based classi-
fication. What’s more, the classification of object-oriented
technology not only make use of the scattering informa-
tion of image, but also takes advantage of the shape, tex-
ture information, and external auxiliary info r ma ti o n of th e
object. The experimental results adequately depict that the
object-oriented technology is also suitable for high-preci-
sion SAR image classification.
With th e developmen t of SAR techno logy, full-po la ri za -
tion and high-resolution SAR images can provide more

X. G. LIU ET AL.119
detailed features. Future research direction is applying
object-oriented technology in the full-polarization SAR
or high-resolution SAR image classification.
5. Acknowledgements
Thanks for providing experimental data by Envisat satel-
lite radar data-sharing project of Center for Earth Obser-
vation and Digital Earth Chinese Academy of Sciences
to this paper.
6. References
[1] L. P. You, “Object Oriented Classification Method from
High Resolution Remote Sensing Imagery,” Master The-
sis, Fujian Normal University, Fuzhou, 2007.
[2] C. B. Ursula, H. Peter, W. Gregor, et al., “Multi-Resolution,
Object-Oriented Fuzzy Analysis of Remote Sensing Data
for GIS-Ready Information,” ISPRS Journal of Photo-
grammetry & Remote Sensing, Vol. 58, No. 3-4, 2004, pp.
239-258.
[3] G. Z. Shen and J. J. Liao, “An Object Oriented Methodol -
ogy for Automatic Analysis of Inundate Extent Using
Multi-Polarized SAR Image,” Remote Sensing Technol-
ogy and Application, Vol. 22, No. 1, 2007, pp. 79-82.
[4] T. Michael, E. Thomas and D. Stefan, “Object-Oriented
Detection of Settlement Areas from TerraSAR-X DATA,
Initial Reports,” Remote Sensing-New Challenges of High
Resolution, 2008, pp. 242-248.
[5] F. L. Chen, C. Wang and H. Zhang, “The Analysis of S in-
gle Polarization Synthetic Aperture Radar Images for the
Application of Land-Use and Land-Cover Change,” Re-
mote Sensing Technology and Application, Vol. 23, No. 3,
2008, pp. 289-293.
[6] J. Qian, Q. M. Zhou and Q. Hou, “Comparison of Pixel-
Based and Object-Oriented Classification Methods for
Extracting Built-up Areas IN AridZone,” ISPRS Work-
shop on Updating Geo-Spatial Databases with Imagery
& The 5th ISPRS Workshop on DMGISs, 2007, pp.
163-171.
[7] L. Stanistaw, “Object-Oriented Classification of Landsat
ETM+ Satellite Image,” Journal of Water and Land De-
velopment, Vol. 20, No. 10, 2006, pp. 91-106.
[8] Q. L. Tan, Z. G. Liu and W. Shen, “An Algorithm for
Object-Oriented Multi-Scale Remote Sensing Image
Segmentation,” Journal of Beijing Jiaotong University,
Vol. 31, No. 4, 2007, pp. 111-114.
[9] F. L. Du, Q. G. Tian and X. Q. Xia, “Object-Oriented
Image Classification Analysis and Evaluation,” Remote
Sensing Technology and Application, Vol. 19, No. 1,
2004, pp. 20-23.
[10] C. Y. Zhou, “Object Oriented Information Extraction
Technology from High Resolution Remote Sensing Im-
agery,” Master Thesis, Shan Dong University of Science
and Technology, Tsingtao, 2006.
[11] H. M. Zhang, Z. Z. Bian, et al., “A Novel Multi-Resolu-
tion Fuzzy Segmentation Method on MR Image,” Journal
of Computer Science and Technology, Vol. 18, No. 5,
2003, pp. 659-666.