A SHAPE SIMILARITY BASED CHANGE DETECTION APPROACH OF MULTI-RESOLUTION REMOTE SENSING IMAGES

While a lot of approaches have been reported for change detection using high resolution images, few approaches which are used for different scale images have been developed. This paper proposes a change detection approach which can be used to detect the change in images with different scales. In this study, image objects have been extracted from images obtained at different time. Then, descriptors to describe the shape features are constructed. After that, the similarity between corresponding features is defined to measure whether the observed objects changed or not. To test the proposed approach, a test for building change detection is designed. The accuracy of our result is 96.1%, which is distinctly higher than that of conventional method 81.9%. The primary results show that the proposed approach is feasible and valid.


INTRODUCTION
Change detection plays an important role in the information extraction, and it is widely used in many aspects, such as disaster assessment, deforestation, urban growth and so on. A great amount of research about change detection has been done by many experts and scholars (Singh, 1998, Coppin et al., 2004, Radke et al., 2005, Mao et al., 2011. However, with the development of satellite and sensor technologies, the resolution of image changes quite significantly. Images at the finest scales are not only inclined to highlight geometrical details, but are more easily to be affected by noise; while other images at coarser scales exhibit less precise details but a stronger immunity to noise (Moser et al., 2011). Therefore, how to detect the change among multi-resolution images is a hot topic, on which there have been several heated discussions going on. The O.Hall raised a multi-resolution framework for landscape analysis (Hay et al., 2001). It is noted that landscape analysis performed at a single scale is insufficient to understand multiresolution patterns and processes (Coburn and Roberts, 2004). Methods of multi-resolution change detection can be categorized into pixel-level based and object-level based, according to the nature of the data processing granularity. The former analyzes and detects changes at pixel level (Coburn and Roberts, 2004, Bruzzone and Carlin, 2006, Celik, 2009 and the latter finds out changes at the object level (Hay et al., 2001, Lang andBlaschke, 2003). And the similarity measure for change detection was used for other kind of data source,such as SAR image. (Inglada and Mercier, 2007).
Although many methods have been raised for change detection, there are few approaches of change detection based on the shape similarity of multi-resolution images. This paper proposes a change detection approach, called FVA (feature vector analysis), for multiresolution images, which is based on calculating the shape similarity of corresponding features. The FVA consists of four steps: (1) Image registration: selecting two images with different resolutions, and performing the position registration based on the same geographic coordinate system; (2) Image segmentation: applying segmentation algorithm to images to extract image blobs (objects); (3) Image classification: classification and extracting silhouette of target objects from each image; (4) Shape similarity calculation: calculating shape similarity of the corresponding features. The appropriate threshold is determined to separate changed area from unchanged region.
The paper is divided into four sections. In section I, related research and a brief introduction of the FVA algorithm are given. The methodology is explained in section II. The result of experiment is discussed in section III. Finally, section IV draws the conclusion.

METHOD
Unlike the traditional method of change detection, the basic idea of FVA is to detect changes by calculating shape similarity of corresponding features. Since shapes of corresponding objects in multi-resolution images are similar and the shape of an object can be described, it is useful for detecting changes in multi-resolution images. In the paper, the focus is on how to describe the shape and calculate the shape similarity between the features of each candidate objects in multi-resolution images. The flowchart of FVA is shown in Figure1 and its pivotal steps are detailed in the following section.

Reprocessing and classification of images
The approach of FVA is similar to other kinds of change detection techniques. It also requires some basic data pre-processing. The important pre-processing steps for FVA are as follows: (a) radiometric and coordinate normalization; (b) position registration based on the same geographic coordinate system; (c) multiresolution segmentation; (d) classification based on the result of segmentation. The feature vectors of classification result are exported into a shapefile. In this paper, these steps are accomplished by ENVI and eCongnition software.
The quality of registration is critical for change detection. Misregistration will lead the corruption of change detection result which is difficult to be mitigated via post-processing intervention. For the multi-resolution image, it is impossible to register Figure 1. The flowchart of FVA approach the paired images pixel by pixel accurately. The FVA avoids strict requirement for pixel-registration through position registration by selecting ground control points manually, which is based on the same geographic coordinate system. Radiometric normalization can reduce the effect of atmospheric conditions, solar illumination, sensor calibration, phonologic variability, and other conditions. The segmentation procedure starts with single image objects of one pixel and repeatedly merges them in several loops in pairs to larger units as long as an upper threshold of homogeneity is not exceeded locally. The homogeneity criterion is defined as a combination of spectral homogeneity and shape homogeneity. The silhouettes of features are extracted from each image through classification algorithm based on the result of segmentation. The features of objects do not overlap with each other, and they are organized into one layer. Finally, the feature vectors of classification are exported into a shapefile together.

Features' shape matching algorithm
After the previous pre-processing step, the extracted features form the canditate object feature data set (T1-feature data set and T2-feature data set in Figure1). Features' shape matching is another crital step for the proposed approach,the mis-matching error will lead to the corruption of change detection result, so correct matching is critical for the following calculation. The algorithm of feature matching contains three steps: where P is the canditate object feature data set, p1,p2,p3,· · · ,pi are elements of the feature data set. The candidate object feature data set P is traversed one by one for matching corresponding features .
step2: Get pi from the set P , select the features which are contained/intersectant with pi, the alternative feature data set Q for matching is composed of the selected features: , qj is the candidate feature for matching.
step3: The distance between the gravity center of pi and qj is assumed as di. The minimum distance dn is chosen from the D,and The minnimum distance dn = (pm, qn), pm and qn are the matched corresponding features.
The feature pm matches with qn, which means they are corresponding features at the same geographical location in each image. As the statement in section II, the silhouette of an object has different representation in different spatial resolution, and the algorithm matches the polygons which are the nearest gravity center among the polygons feature dataset. Obviously, the algorithm matches features based on the geographical location, if there have no changes, the algorithm matches the true corresponding features, and on the contrary, the matching result is false.

Calculation of shape similarity and change detection
The shape of an object is an important visual feature for describing image content (Loncaric, 1998, Zhang and. The techniques about representation and description of shape can be generally classified into two classes: contour-based and regionbased method (Zhang and Lu, 2004). In this paper, the contour shape descriptors are used (Peura and Iivarinen, 1997) and they are developed for calculating the SSM (shape similarity measurement). The notations used here are listed in Table 1. The bounding rectangle refers to the convexity of polygon (Peura and Iivarinen, 1997), the height (H) is defined as the longer edge of a rectangle than the width (W ). HausDroff distance is a classical correspondence-based shape matching method (Peura and Iivarinen, 1997). It has often been used to measure similarity between shapes (Chetverikov andKhenokh, 1999, Huttenlocher andRucklidge, 1991). In the proposed approach of FVA, the pm and qn are matched features who are selected from the feature dataset to calculate their SSM value by the above descriptors, according the following rules: The above four parameters are used for calculating the value of SSM (pm, qn), which is defined in Equation (1) SSM (pm, qn) = K1 · SSM (S) + K2 · SSM (P )+ Where K1, K2, K3 and K4 are weight of each parameter factor, and ∑ 4 i=1 Ki = 1, SSM (pm, qn) ∈ [0, 1]. When change occurs, the value of SSM (pm, qn) is lower; on the contrary, the value of SSM (pm, qn) is higher. The weights K1, K2, K3 andK4 of each factor reflects the subjective impact on the result. Therefore, they can be determined by the owner of an application. To generate change map by FVA, it's necessary to process the superimposed feature polygon according to the value of SSM (pm, qn) as the following rules: (a): if SSM (pm, qn) > T , the unchanged region is pm ∪ qn.
Where T is the threshold of SSM, which is used to determine if an object is changed. The changed region and unchanged region are generated by the rule (a) and rule (b). The processed superimposed feature polygon is used as mask for classifying the classification result of post image, then the change detection thematic map of FVA is generated.

RESULTS
To verify the proposed approach of FVA is feasible and effective, two images are used to detect changes of building in the experimental district.   In the experiment, the approach of FVA integrates the spatial feature in multi-resolution images, including similarity of bounding rectangle, the area of features, perimeter of features, the Haus-Droff Distance of features. There is no mis-matching error for the experimental district, and the error on the pixel level is lead by the difference of the corresponding features' shape in multiresolution images and the deviation of the relax geographic registration.

DISCUSSION AND CONCLUSIONS
China has been experiencing national geographical state monitoring, which needs to integrate multi-resolution image data to detect change. As one of the important means for earth surface monitoring, remote sensing technology provides necessary data  Table 4. Accuracy of the FVA change detection for timely detection change at regional and global scales. With the improvement of remote sensing image resolution, traditional change detection methods cannot cope with multi-resolution and multi-sensor source data. This study has proposed an FVA approach aimed at effectively detecting change of multi-resolution and multi-sensor source data. The basic idea of FVA method is to utilize the shape similarity of corresponding features in multiscale space.
The performance of FVA was evaluated by using data from two remote sensing platforms, QuickBird satellite (0.61m) and aerial ADS camera (0.5m). Comparing the true change which is obtained by analyzing the contemporaneous cadastral map with each image, the result consistently demonstrated the improvement of the FVA. Quantitative assessment also demonstrated that the FVA achieved higher accuracy than the conventional approach with the overall accuracy increased from 81.9% to 96.1%, and the Kappa coefficient increased from 0.93 to 0.98. The omission error of the approach of FVA is decreased from 20.8% to 5.7%, comparing with the method of image difference.
This research has shown that the shape similarity of corresponding features in multi-resolution image is highly useful for change detection. The change is detected by utilizing shape similarity, which can avoid strict pixel-registration. Although an object in the multi-resolution images may have different shapes, when these shapes are superimposed, the gravity centers of them are spatially close to each other and thus these shapes are corresponding features and considered similar. The result of experiment showed that the approach of FVA can detect changed regions of multi-resolution images effectively.
However, more complex and larger spatial area has not yet been used for testing effectiveness of the proposed FVA approach. The FVA is invalid for detecting the change in texture other than in shape, for example: an area changes entirely from grassplot to concrete floor. Therefore, questions that remained to be addressed in the further study are: (1) To consider the topical texture and the partial spectrum in shape similarity calculation; (2) To design more effective and robust algorithms for matching corresponding features; (3) Larger images and other kinds of change detection are needed to verify the performance of the FVA approach.

ACKNOWLEDGMENT
This work is partly supported by project 2012BAJ15B04.