Remote Sensing Satellite Image Database System Allowing Image Portion Retrievals Utilizing Principal Component Which Consists Spectral and Spatial Features Extracted from Imagery Data

Remote Sensing satellite image database which allows image portion retrievals utilizing principal component which consists of spectral and spatial features extracted from the imagery data is proposed. Through the experiments with actual remote sensing satellite images, it is found that the proposed image retrieval does work so well.


INTRODUCTION
In accordance with expansion in the multimedia technologies and the Internet, CBIR has been an active research topic since the first 1990's.The concept of content based retrieval (CBR) in image start from the first 1980s and Serious applications started in the first 1990s [1].Retrieval from databases with a large number of images has attracted considerable attention from the computer vision and pattern recognition society.
Brahmi et al. mentioned the two drawbacks in the keyword annotation image retrieval.First, images are not always annotated and the manual annotation expensive also time consuming.
Second, human annotation is not objective the same image may be annotated differently by different observers [2].Unlike the traditional approach that using the keyword annotation as a method to search images, CBIR system performs retrieval based on the similarity feature vector of color, texture, shape and other image content.Comparing to the traditional systems, the CBIR systems perform retrieval more objectiveness [3].
Global features related to color or texture are commonly used to describe the image content in image retrieval.The problem using global features is this method cannot capture all parts of the image having different characteristics [4].In order to capture specific parts of the image the local feature is used.The proposed method uses 2D Discrete wavelet transform with Haar base function, combined the two high sub-band frequency to make significant points and edge, choose any part of image by threshold the high coefficient value.
Image feature based image portion retrieval method is proposed in this paper.Principal component which consists of spectral and spatial features extracted from the image are used as features.
The following section describes the proposed image database system followed by some experiments.Then conclusion is described together with some discussions.

II. PROPOSED SATELLITE IMAGE DATABASE SYSTEM
A. Image Database Database system has to have the following three functions,

1) Image coding and archiving with attributes information 2) Database operations including image retrievals and updating database (3)Database management including maintenance
Attributes can be used for meta search while meta search allows image retrievals.The meta database is information of image information.Therefore, it is useful for keyword search.For image search, image coding is important.If the image coding is performed perfectly, then image retrievals can be done easily.On the other hands, there is image portion retrievals which allows image portion search in the image in the image database.In the image portion retrievals, some measures for representation of image features are required.One of the simple features is similarity between the template image in the image database and the current query image portions.Other than these, principal component is used as feature.The extracted spectral and spatial features can be projected in principal component space as a vector.The vector can be interpreted as a feature of the image portion in concern and can be used as measure of the similarity between template and the query image portion.

B. Image Retrieval
Two image retrieval methods are supported by the proposed image retrieval system.One is keyword based search utilizing the following keywords, www.ijarai.thesai.org

1) Image file name 2) Satellite name 3) Sensor name 4) Band number 5) Area name 6) Observation date 7) Observation time 8) Area feature 9) Cloud content
Another image retrieval method is image feature based method.There are utilizing features for image retrievals.

C. Image Feature
Average, variance, skewness, kurtosis, energy, entropy, the number of edges, minimum/maximum pixel values, contrast, digital count at most frequent pixel values are selected as features.After calculation of features, principal component analysis is applied to the features.The first three principal components are extracted then the mapped features are divided into groups a shown in Figure 1.The first three principal components are extracted then the mapped features are divided into groups Thus the images in the image database are projected in the principal component space.Then the group number is used for one of attribute information.

D. Principal Component Analysis
Let x i , i=1, 2, ..,p be features extracted from the image assuming correlation among the features.Then new variable, z is defined as follows, Where a i is determined as maximizing variance of z under the condition of . Such this z is called as the first principal component.Also such these coefficients a i are called as the first principal component vector.Rewriting the first principal component as z 1 , and the first principal component vector a 1 , then the arbitrary order, j of principal component z j and principal component vector a j can be written as follows,

E. The Number of Edges
The following Sobel operator derived edges are extracted from the images for one of the features for image retrievals.

G. Principal Component for Image Retrievals
PCA component derived from satellite imagery data has different variance.In order to normalize the variance, the following equation is used, (6) Where x ij denotes two dimensional imagery of pixel data while s i denotes variance of the pixel data of band i. Correlation coefficient between band i and j is expressed as follows, (7) Using the correlation coefficients, eigen value and eigen vector of correlation coefficient is estimated utilizing Jacob method.Let X, Λ be eigen vector and eigen value of A. Non singular matrix M is utilized, then, (8) Thus it is found that eigen value and eigen vector is not going to be changed with the aforementioned operation with the non singular matrix, M. The followings are candidates of M, Thus eigen values are determined followed by the corresponding eigen vector.

H. Implementation of Image Retrievals
Web design for image retrievals is conducted.Figure 3 shows the registration form.The aforementioned key words for key word image retrievals can be input to the database.www.ijarai.thesai.orgFig. 3.
Registration for which allows input keywords through dialog boxes The registered image with attribution data can be confirmed as shown in Figure 4. Figure 5 shows top page of the image retrieval system.If user input query information through the web page which is shown in Figure 6, then search results can be obtained from the image retrieval system as shown in Figure 7. Keyword search can be done through the web page of Figure 6.Image feature based search is also done through the web page of Figure 8. Then the search results can be obtained as shown in Figure 9. Also the attribution information is obtained as shown in Figure 10.www.ijarai.thesai.org  12 students are nominated for performance tests of hit ratio and the required number of steps for image retrievals.

B. B. Image Retrieved Results
The examples of the first principal component images (the largest eigen vector which corresponds to the largest eigen value) are shown in Figure 11.Hit ratio and the required number of steps for search are totally depending on the complexity of the query image.The highest hit ratio is shown for the query image "F" while the lowest hit ratio is shown for the query image of "E".Therefore it is said that image feature rich image of "F" is easy to retrieve while image feature poor image of "E" is difficult for search.On the other hands, the number of steps required for image retrievals for the query image "C" is greatest and that for the query image "E" is smallest.This implies that it takes long time to search the query image "C", at the same time, not so large hit ratio cannot be achieved for such query image.Meanwhile, user tried to search images for the query image "E", and then user quit to search because it is difficult to search.Therefore, not only hit ratio but also the number of steps required for search is poor.
IV. CONCLUSION Remote Sensing satellite image database which allows image portion retrievals utilizing principal component which consists of spectral and spatial features extracted from the imagery data is proposed.Through the experiments with actual remote sensing satellite images, it is found that the proposed image retrieval does work so well.
Experimental results show that hit ratio depends on the image features which are containing in the satellite image in concern.The required number of steps does not depend on the image features which are contained in the satellite image in concern.Because user forgets about the image retrievals when they feel that it is difficult to search due to the fact that satellite image in concern does not have any specific features.

Fig. 1 .
Fig. 1.The first three principal components are extracted then the mapped features are divided into groups

( 4 )
Where f x , f y denotes pixel values of edge enhanced images in the horizontal and vertical directions.With appropriated threshold, the horizontal and vertical edges are extracted from the enhanced images.Figure2shows one of examples of edge detected images.
Fig. 2.Edge detections as basic orthogonal matrices.Using the basic orthogonal matrix, power of A can be written as follows, the maximum absolute value of the non diagonal elements, a ij(s) is found, then(12)It is possible to find the θ as to equation (12) is satisfied.Then,(13)Thus the eigen values are obtained,(14)If we choose Mi as follows,Then,

Fig. 4 .
Fig. 4. Confirmation of the registered image and attribution data

Fig. 5 .
Fig. 5. page of the image retrieval system

Fig. 11 .
Fig. 11.Examples of the first principal component images (the largest eigen vector which corresponds to the largest eigen value) Image feature based image retrieval is conducted.If user clicks one of the displayed image of Figure11, then the retrieved image candidates are displayed in accordance with the distance between the query image and the candidate images as shown in Figure12.Figure12shows the example images extracted from the first group of the first principal component images.It is found that image complexity can be expressed with the first principal component images.

Fig. 12 .
Fig. 12. Example images extracted from the first group of the first principal component images.

Fig. 13 .
Fig. 13.Query images which are selected from the 288 of sample Landsat TM images in the image database.

TABLE I .
HIT RATIO AND THE REQUIRED NUMBER OF STEPS FOR SEARCH WHICH ARE EVALUATED WITH LAND SAT SATELLITE TM IMAGERY DATA BY 12 STUDENTS.