Sss

Unsupervised Land Cover Classification of SAR Images by Contour Tracing V.V.Chamundeeswari, D.Singh, K.Singh Department of Electronics & Computer Engineering Indian Institute of Technology Roorkee Roorkee, India e-mail: [email protected], [email protected], [email protected] requires training samples and a priori information about the region like digital elevation map etc. Unsupervised classification of SAR images requires no a priori information about the image. For segmentation and labeling of classes, the algorithm extracts information from the image itself. In this paper, we present an algorithm for unsupervised classification of SAR images by block based segmentation and contour tracing. We segment the SAR image into monotone, texture areas and edges. The monotone and textural regions are also differentiated according to intensity and textural patterns. MRFs (Markov Random Field) have been used to model image textural features [1]. But the main drawback of MRF algorithms is that the fine structures like 1-3 pixel wide line segments may disappear partially or entirely and region borders are not precisely located. Statistical Image model may not be accurate enough and classes are overlapped in the feature space. This results in some confusion in data classification in homogenous areas as well as at region borders or close to fine structures. Other popular methods for the analysis of image texture are Gray level cooccurence matrix (GLCM) [2], Gabor wavelets [3], tree structured wavelets [4], wavelet packets [5] etc. Randen et al [6] compares performance of general texture analysis schemes. GLCM features were found to be more sensitive to texture boundaries compared to MRF. GLCM provided better classification accuracy for optical images [7]. Niedermeier [8] has used wavelet decomposition and contour tracing algorithm for coastline extraction. Active contour tracing algorithm with filling/removing loops was employed to identify coastlines. Niedermeier has developed contour tracing only to separate sea from land area. Identifying contours in a mixed area with urban, water and vegetation require complex analysis of spatial data.

Abstract— The potentiality of Synthetic Aperture Radar (SAR) Images for land cover mapping is an important area of research. For Single band, single polarized SAR Image, information is available in the form of Intensity and texture only. Land cover classification of SAR Images requires exploitation of spatial relationship of pixels also, in addition to pixel level segmentation. SAR image can be segmented successfully if the regions with homogeneous intensity and texture areas can be identified and grouped together. So far, contour tracing has been used only in demarcating sea and land. Identifying contours in a domesticated area with a mixture of water, urban and vegetation areas require complex analysis of spatial distribution of pixels. In this paper, we have presented an unsupervised classification algorithm using Maximum a posteriori (MAP) segmentation for SAR images in which SAR image is classified into monotone, texture and edge regions. Monotone and textured regions are labeled as land cover types like water, urban and vegetation areas using K-means classification. SAR Image of the region with latitude varying from 77.86º to 77.91º and longitude varying between 29.89 º and 29.85 º of Haridwar region, India is considered for segmentation. We have compared the segmented image obtained by this methodology with the topographic map of the corresponding region. The water, urban and vegetation areas are clearly recognized with proposed classification approach which represents a very good agreement with the original topographic sheet . Keywords- Unsupervised, SAR segmentation, Contour tracing, MAPestimation, texture segmentation, block based SAR segmentation.

I.

INTRODUCTION

Land cover mapping using SAR images is an important area of research. Since SAR sensors provide all-time and allweather surveying, potentiality of using SAR images in various applications like land cover classification, object detection is to be explored. Classification of land cover into classes like water, urban and vegetation helps in planning and management of urban regions, such as sustainable development and smart growth. Segmentation is a basic technique of digital image processing with an ultimate goal of improving an image for subsequent analysis and scene description. An image can be segmented in to classes based on gray levels, textures, edges etc. A single band and single polarized SAR image contain information only in the form of intensity and texture. Segmentation of single band, single polarized SAR image is approached by two classical techniques: Supervised and Unsupervised. Supervised classification of SAR Images

1-4244-1212-9/07/$25.00 ©2007 IEEE.

In this paper, we approach SAR image segmentation problem as block based segmentation. SAR image is divided into blocks. Then, each block is analyzed for its homogeneity of gray levels, textural patterns and edges. Blocks are grouped and labeled according to their characteristics. Edge blocks are checked for its continuity and contour is traced. Edge connectivity helps in removal of noise and improves classification accuracy. Region labeling is done to connect adjacent homogenous blocks. In this process, any isolated homogenous or edge blocks with small regions are labeled as undecided. These undecided blocks are assigned to the neighbouring homogenous regions in the final step. K-means classification is performed on the block based segmented image and water, urban and vegetation areas are identified.

547

Authorized licensed use limited to: BANARAS HINDU UNIVERSITY. Downloaded on October 18, 2008 at 02:39 from IEEE Xplore. Restrictions apply.

The proposed method is described in Section II, including Block based MAP(Maximum a posteriori) and texture differentiation by wavelets, Section III presents the experimental results on a SAR image. Finally, Section IV provides conclusion of the paper.

x* = argmax P( X = x Y = y)

= argmax[ln P(Y = y X = x) + ln P(X = x)]   G  Rg = argmax ∑ ∑− lnσ~tg2  + ∑Vc ( x) + constant   c∈C  t∈ΩA g=1 2    

II. THE PROPOSED METHOD

---(1)

A.

Block based Segmentation The overall classification methodology is depicted in Fig. 1. For the analysis of block based image segmentation, the input

where

image of size N1 × N 2 is divided into non-overlapping small square blocks with a s × s dimension. Then, there are n1 × n2 blocks in the image, where n1 = N1 / s

X = {X t = xt t ∈ Ω A , xt ∈ {0,1,2,..K }}

Ω A = {(i, j ) i = 1,2..n1, j = 1,2...n2 }

,

σ

Input Image

where

Division of image to equal sized image blocks Calculation of mean Intensity and Variance for each image block

characteristics defined by the block label k. The block labels k corresponds to the various edges as follows: For vertical edge, k=0, horizontal edge, k=1, diagonal edges in +45 and -45 degrees correspond to k=2 and k=3 respectively. Y is defined as the random field for the set of all observed gray-levels in Ω . Then, the observed image data can be written as Y=y and

} where

y ijB

is the closed

c, and C is the set of all cliques in Ω A associated with the neighborhood system [11]. Concepts of clique potentials and computing clique potential using Gibbs distribution is explained in [11].

then it implies that the block at t ∈ Ω A has the block

{

σ~tg2

form expressed ML estimate of tg . G represents the total number of distinctive areas in the block. constant represents constant value which is not affecting maximization. For all edge blocks, there will be two distinct areas within the blocks separated by the edges. Vc(x) is the clique potential for a clique

denotes the set of all block indices. When the random variable Xt has a value xt =k,

y = yij (i, j ) ∈ Ω A

denotes the total number of pixels in the gth

area of the block separated by an edge and

and n2 = N 2 / s . Each of the blocks has to be identified as monotone, texture and edge blocks. As the first step, mean and variance are computed for every block for identifying monotone blocks. If variance is less than 2% of its mean intensity, then the image is classified as monotone blocks. In the second step, blocks with horizontal, vertical and diagonal edges are recognized and grouped as edge blocks. Let X denote 2-D random field representing an image configuration composing of all image characteristics (ie. Monotone, texture or one of various edge blocks). Then

Rg

Check for Var<2% of mean intensity

Edge block labeling by MAP

is the set of all gray levels

Blocks labelled as Monotone

horizontal, vertical and diagonal edge blocks are labelled

Extraction of Textural feature vectors

in the image block located at (i, j ) ∈ Ω A .

For identifying and labeling edge blocks, maximum a posteriori distribution is used. Aim is to find the block label configuration x* that maximizes the a posteriori probability P(X/Y). Then, according to Bayesian formulation, the optimal block label configuration is obtained by equivalently maximizing the following criterion for all possible x [9-10]. For any edge block, two distinctive areas represented as g=1 and g=2 is present. For any edge block, the mean intensity level in g=1 is different from that of g=2. G is taken as the total distinctive areas in the region. For G=1, those blocks are grouped as monotone blocks.

Comparing textural feature vectors

Blocks labelled as Texture

Region labeling K-means Classification

Classified Image

Figure 1. Classification Methodology Once monotone and edge blocks are identified, all the remaining blocks can be considered as texture blocks. For all these texture blocks, feature vectors are extracted to differentiate various texture patterns. Wavelet transform is used

1-4244-1212-9/07/$25.00 ©2007 IEEE.

548

Authorized licensed use limited to: BANARAS HINDU UNIVERSITY. Downloaded on October 18, 2008 at 02:39 from IEEE Xplore. Restrictions apply.

The geo-referenced and speckle suppressed SAR image is of size 302 x 302. Then, this image is divided into image blocks of dimension 2 x 2 leading to 151 x 151 blocks. Then, mean intensity and variance are computed for individual blocks. The threshold for variance with mean intensity is checked and monotone blocks are identified. Then, using MAP based segmentation, edge blocks of labels horizontal, vertical and diagonal edges are identified. Daubechies wavelet transform is applied on the image and textural features are captured. Wavelet coefficients are compared to identify different texture patterns. All the image blocks are thus labeled as one of monotone, texture and edge blocks where every monotone and texture pattern is labeled. Region labeling is performed to connect adjacent blocks with similar intensity and texture patterns. Eight neighborhood connectivity is used for checking the adjacency. All neighboring regions with similar intensity and texture properties are merged and thus provide an unsupervised segmented image with three types of regions, namely monotone, texture and edge regions. Now, K-means classification is used to classify the segmented image into water, urban and vegetation areas. For computing classification accuracy, ground truth points representing water, urban and vegetation areas are taken for reference from topographic sheet. The table 1 below shows the confusion matrix generated for the classified image by our methodology.

to represent the textural features. Daubechies (dB4) wavelet transform is applied on every image block. dB4 wavelet decomposition results in four components, high pass, low pass, a horizontal and a vertical component. For every image block, textural feature vector is represented as

{Tijl i = 1,2..n1 , j = 1,2..n2 , l = 1,2,3,4}

where (i,j) represents the block index and l represents the wavelet component of the image block. B.

Region Labeling Once the image blocks are identified as monotone, texture and edge blocks, and the blocks adjacent to each other and have the same intensity and texture are connected and a common region label is provided. For each block, eight neighborhood connectivity is checked for its adjacency with neighboring blocks. For connecting monotone blocks, mean and variance form the feature vector and textural blocks, textural feature vector comprising of its db4 wavelet components is used for comparing the adjacent texture blocks. Only similar block types are compared for adjacency and merged together to form the same region. Let the feature vector be represented as Fij where i,j are block indices. The two neighboring blocks (i,j) and (i+1,j) belong to the same region if the following condition is satisfied.

Fij − Fi +1, j

2


Overall Accuracy= 95.3971% ---(2) TABLE I.

For every region thus connected, label or region number is given so that each region by the end of this process represents the distinct monotone and texture regions. Edge blocks are not considered in this process.

Sl.No

Ground Truth (Percent) Vegetation

Water

95.94

0.00

1.65

2

Urban

0.00

92.23

0.00

3

Vegetation

4.06

7.77

98.35

Total

100.00

100.00

100.00

TABLE II. Sl.No

The ERS-2 SAR-C band image acquired on July, 2001 was taken as input image. Since, SAR images are acquired in the microwave region of electromagnetic spectrum, visual identification of ground control points is very difficult. Thus, ERS-2 SAR image is geo-referenced to geographical coordinates using eight ground control points (four at the corners, one at center of the image and three from topographical map). A first order polynomial transformation function and the nearest neighbor re-sampling technique have been used to perform geo referencing. Subset of SAR image with latitudes and longitudes ranging from (77.86,29.89) to (77.91,29.85) is chosen for implementation of methodology. Adaptive Lee filter is used for speckle suppression. Lee filter is able to smooth away noise in flat regions, but leave fine details unchanged [12].

Urban

1

RESULTS AND DISCUSSION

1-4244-1212-9/07/$25.00 ©2007 IEEE.

Class

Water

C. K-means classification All image blocks present in the entire image are identified either as monotone, texture or edge blocks and blocks with similar characteristics adjacent to each other are merged and regions are labeled. K-means classification is applied on the segmented image to identify water, urban, vegetation and other unclassified areas. III.

CONFUSION MATRIX

COMMISSION &OMISSION ERROR Class

Commission Error (percent)

Omission Error (percent)

1

Water

1.78

4.06

2

Urban

0.00

7.77

3

Vegetation

11.17

1.65

TABLE III. Sl.No

PRODUCER AND USER ACCURACY Class

Producer’s Accuracy (percent)

User’s Accuracy (percent)

1

Water

95.94

98.22

2

Urban

92.23

100.00

3

Vegetation

98.35

88.83

549

Authorized licensed use limited to: BANARAS HINDU UNIVERSITY. Downloaded on October 18, 2008 at 02:39 from IEEE Xplore. Restrictions apply.

measures between class discrimination and results when one class on the ground is misidentified as other class by the observing classifier. Hence, 1.78% of water pixels are improperly labeled as water pixels and 4.06 % of water pixels are misidentified as other land type. The total number of correct pixels in a category is divided by the total number of pixels of that category as derived from the reference data (i.e. Column total) is termed as “Producer’s accuracy” because the producer of the classification is interested in how well a certain area is classified. On the other hand, if the total number of correct pixels in a category is divided by the total number of pixels that were classified in that category, this measure is called “User’s accuracy”. The producer of the map can claim that 95.94% of the time an area is identified as water was identified as such, a user of this map will find that 98.22% of the time will an area he visits that the map says is water will actually be water. Thus, classification of single band, single polarized SAR image is performed by block based segmentation and contour tracing. The results are compared with topographic sheet.

Figure 2. Raw SAR Image of the study area

Water

Urban

Vegetation

REFERENCES [1]

H.Deng and David A. Clausi, ”Unsupervised Image segmentation using a simple MRF model T with a new implementation scheme,” Pattern Recognition, vol. 37, pp.2323-2335, 2004. [2] R.M.Haralick, ”Statistical and structural approaches to texture,” IEEE Proc. , vol.67,no.5,pp.786-804, May 1979. [3] A.K.Jain and F.Farrokhnia, ”Unsupervised texture segmentation using Gabor filters,” Pattern Recognition,vol.24,no.12, pp.1167-1186, 1991. [4] T. Chang and C. C.Jay Kuo, “Texture analysis and classification with tree structured wavelet transform,” IEEE Trans. Image Processing, vol. 2, no.4, pp. 429-441, Oct 1993. [5] C. M. Pun and M. C. Lee, “Log-polar wavelet energy signatures for rotation and scale invariant texture classification,” IEEE Trans. Pattern Anal. Mach. Intell., vol.21, no. 4, pp. 291-310, Apr 1999. [6] T. Randen and J.H.Husey, “Filtering for texture classification: A comparative study,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 21, no. 4, pp. 291-310, Apr, 1999. [7] J.R.Carr and F.P.Miranda, “The semivariogram in comparison to the cooccurrence matrix for classification of image texture,” IEEE Trans. Geosci. Remote Sens., vol. 36, no.6, pp. 1945-1952, Nov. 1998. [8] A.Niedermeier, E.Romaneesen, and S.Lehner, ”Detection of coastline SAR images using wavelet methods,” IEEE Trans. Geosci. Remote Sensing, vol. 38, pp. 2270-2281, Sept.2000. [9] H. Derin and H..Elliott, “Modeling and segmentation of noisy and textured images using Gibbs random fields,” IEEE Trans. Pattern Anal. Mach. Intell., vol.9, no. 1, pp.39-55,1987. [10] C.S.Won and H.Derin, “Unsupervised segmentation of noisy and textured images using Markov Random fields,” CVGIP: Graph. Models Image Process. Vol.54, no.4, pp. 308-328, 1992. [11] C.S.Won, ”A block based MAP segmentation for Image compressions,” IEEE Trans. Circuits and Systems for Video Technology, vol.8, no.5,Sept. 1998. [12] J.S.Lee,”Digital image enhancement and noise filtering by use of local statistics,” IEEE Trans. Pattern Anal. Machine Intell. Vol.42, no.7,pp.165-168, 1980.

Figure 3. Topographic sheet of the study area Fig. 2 shows the raw SAR image of the study area taken as input. Fig. 3 shows the topographic sheet of the same area from which ground truth points representing water, urban and vegetation areas are taken for reference. Raw SAR image is georeferenced and speckle suppressed. Then, classification methodology explained in section II is applied and the Fig. 4 shows the segmented image with water, urban and vegetation areas. Water

Vegetation

Urban

Figure 4. Segmented SAR Image with Water, urban and Vegetation areas. Table I shows the confusion matrix generated for the classified image by comparing with ground truth points from topographic sheet. It gives an overall accuracy of 95.3971%. Table II lists the commission and omission error. An error of commission is a measure of the ability to discriminate within a class and occurs when the classifier incorrectly commits pixels of the class being sought to other classes. In this example, the commission error for water stems from improperly calling other classes water, so that three pixels labeled as water are really a composite of other classes. An error of omission

1-4244-1212-9/07/$25.00 ©2007 IEEE.

550

Authorized licensed use limited to: BANARAS HINDU UNIVERSITY. Downloaded on October 18, 2008 at 02:39 from IEEE Xplore. Restrictions apply.