Research on multi-scale segmentation parameter selection during the object-oriented remote-sensing image information extraction

By Leo Lawrence,2015-04-02 15:21
8 views 0
Research on multi-scale segmentation parameter selection during the object-oriented remote-sensing image information extraction

    Research on multi-scale segmentation parameter

    selection during the object-oriented remote-

    sensing image information extraction

     Abstract. with the appearance of high spatial resolution remote-sensing image, the object-oriented image analysis technology provides new thinking and approach. The multi-scale segmentation is one of the key technologies for the object-oriented high spatial resolution remote-sensing analysis. This paper studies the effect of parameters such as color, shape and segmentation scale on the image segmentation quality based on the heterogeneity minimization region-merging algorithm in multi-scale segmentation.

     Key words: object-oriented, multi-scale

    segmentation, segmentation parameters

     1. Introduction

     The high spatial resolution remote-sensing has abundant spatial information and clear detail texture information. Adopting traditional pixel-based spectral information extraction technology can not satisfy the demand any more, while the object-oriented image analysis technology provides new thinking and approach for high spatial resolution remote-sensing image information [1]and has been a research hot-spot in remote-sensing image processing field with the key technologies including image segmentation and

    information extraction[2], of which the image segmentation is one of the key technologies to analyze the object-oriented remote-sensing image, which will affect the information extraction and identification results. Zhang Yujin[3] classified it into

    segmentation technology based on region and segmentation technology based on border; Guan Yuanxiu[4] etc classified it into two methods including data-driven method (from bottom to top) and knowledge-driven method (from top to bottom); Zhou Chenghu[5] etc classified it into region extraction, feature spatial clustering, histogram threshold and edge detection. During multi-scale segmentation, the heterogeneity covers color factor and shape factor, but the scale is the minimum threshold of image object heterogeneity. Therefore, the parameter selection of color, shape and segmentation scale becomes particularly important during multi-scale segmentation.

     2. The region-merging algorithm based on the heterogeneity minimization

     The basic idea of heterogeneity minimization region-merging algorithm is to emerge the single neighbor pixel into lots of small image objects by heterogeneity minimization, then emerge the small image objects into larger image until the objects forming polygon objects eventually. The heterogeneity of image objects f can be represented as:


     In the formula, hcolor refers to spectrum heterogeneity factor; hshape refers to shape heterogeneity factor; wcolor refers to spectrum factor

    weight, wshape refers to shape factor weight and 0?wshape?1? 0?wcolor ?1?wshape+ wcolor =1.

    3. The multi-scale segmentation parameter selection

     The setting of various parameters is very important in the process of multi-scale image segmentation. The multi-scale segmentation parameters include varying color factor, shape factor and segmentation factorcolor. The construction of multi-scale parameters, see figure 1.

     3.1 Color factor

     Color factor reflects the spectrum information that is the main information contained in the image data[6]. The weight setting is one of the important elements to affect the segmentation quality. In the process of image segmentation, the color factor is the most important factors to generate image object with the larger weight setting.

     3.2 Shape factor

     Shape factor helps to avoid too fragmented image objects and can avoid the phenomenon of “same object” with different spectrum”, “same spectrum with different objects” and “salt & pepper noise” so as to improve the segmentation quality and information extraction precision[7]. The shape factor is comprised by compactness and smoothness, of which the compactness is used to separate the compact and non-compact target areas but the smoothness is used to describe the border smooth degree of image objects and avoid the image objects presenting indentation on the border.

     Due to the fact that the sum of color factor and shape factor is 1, the high weight setting of shape factor lowers the color factor weight, which is not beneficial to information extraction. Therefore, to maximum the weight setting of color shape should be followed in the process of image segmentation; the necessary shape factor should be used at the time of segmentation image objects[7].

     Those variable setting should be in line with actual features requirements, because if not consider the shape information but set the color factor of 1, the image object will be generated with the narrow zigzagged spectrum; if only emphasize the shape information but ignore the spectrum information, the generated image do not have formation[8]. The segmentation quality with scale of 40, compactness of 0.5, smoothness of 0.5 and the comparative color factor of 0.9 and 0.1 respectively, see figure 2.

     As showed in figure 2, setting too small color factor weight and focusing the shape factor will lead to the massive loss of features color information, which causes the less-than-ideal image quality after segmentation. Therefore, it is needed to set larger color factor weight. The effect of scale 40, color factor 0.9 and comparative compactness factor 0.9 and 0.1 on segmentation quality, see figure 3. As can be seen in figure 3, the relatively small compactness parameter will cause the over

    consideration of internal structure of image objects; the relatively large compactness parameter will ignore the contributions of image object border to image

    objects, which cause the less-than-ideal image object quality after segmentation.

     3.3 Segmentation scale

     The selection of segmentation scale is very important, which determines the size of image objects, the quality of segmentation and the precision of information extraction directly. Hence, the reasonable parameters should be set to compare the quality of segmentation image under different scales before multi-scale segmentation. This paper found out a group of best segmentation parameters through a large number of experiments. The segmentation result at the time of color factor 0.8, shape factor 0.2, smoothness 0.5, compactness 0.5 and the comparative scales 20, 40, 70, 100 respectively, see figure 4.

     As can be seen in figure 4, when the segmentation scale is small, the features in the image will be over segmented; when the larger segmentation scale is selected, the “vague” image of small features in image will not be beneficial to the small area extraction of features. Hence, the scale is determined by type characteristics of the extracting features with the small area of features selecting relatively small segmentation scale and large area of features selecting relatively large segmentation scale.

     4. Conclusion

     It can be seen that every parameter will have an influence on segmentation quality in the image segmentation. Therefore, in the practical application, the reasonable setting and selection of segmentation parameters should be in line with image

    characteristics, different features characteristics so as to make the image quality more real and objective after segmentation.

     5. Acknowledgements

     This article is completed with the fund of the Key Laboratory of Water Environment Evolution and Pollution Co- ntrol in Three Gorges Reservoir(serial number is 2012QN-07 ) .


     [1] P.F. Xiao, X.Z. Feng. High resolution remote sensing image segmentation and information extraction (Science Press, China, 2012)

     [2] Metzler V.T., Aaeh C. object-oriented image analysis by evaluating the causal object hierarchy of a partitioned recon structive sealed-space. Proceedings of ISMM 2002 Redistribution rights reserved CSIRO Publishing, ed. by Talbot H, and Beare R, 2002, 265-276.

     [3] Y.J. Zhang. Image Processing (Tsinghua University Press, China, 2006)

     [4] Guan Yuanxiu, Cheng Xiaoyang. High Resolution Satellite Image Processing Guide (Science Press, China, 2008)

     [5] Zhou Chenghu, Luo Jiancheng. High Resolution Satellite Imagery Geosciences computing (Science Press, China, 2009)

     [6] Shackelford A.K., Davis C.H. Combined Fuzzy Pixel- based and Object-based Approach for

    Classification of High- resolution Multispectral Data over Urban Areas, IEEE Transactions on Geo-Science and Remote Sensing, Vol. 41-10(2003), 2354-2363.

     [7] H.P. Huang, B.F. Wu. Landscape Multi-Scale Image Analysis Based on the Region Growing Segmentation, Progress in Geography, Vol.23-3(2004) ,p.9-15

     [8] Benz U.C., Hofmann P., Willhauek G., et al. Multi-resolution, objected-oriented fuzzy analysis of remote sensing data for GIS-ready information[J]. ISPRS Journal of Photogrammetry and Remote Sensing.2004 (58): 239-258.

Report this document

For any questions or suggestions please email