The Research Of Multi-Scale Segmentation,Object-Oriented Extraction Of Target Information

With the rapid development of satellite remote sensing technology, facing with the mass of remote sensing image data, one of the important problems is how to extract desired information from mass high resolution remote sensing data accurately, quickly and automatically and utilize this information effectively.Object oriented processing and analysis technology is a new method and technology of the remote sensing image information extraction, which become the key research and application in glasses circles in recent years. This method can not only effectively improve the recognition and extraction accuracy of high resolution remote sensing image, but also provide reference for the medium resolution remote sensing image extraction.Due to the characteristic of hierarchy, structure and multi-scale of the different types ground objects in remote sensing images. During the process of image segmentation, if segment all ground objects in the same image with single scale, some ground objects will be “over segmentation” and“under-segmentation” phenomenon,it is difficult to meet the segmentation accuracy requirements. therefore, in order to overcome the limitation of the single scale, segmenting the various features in different scales, namely multi-scale segmentation technology, then extracting the ground object information from the image objects layer at different scales. multi-scale segmentation is a key technology in object oriented image analysis(Object Oriented Image Analysis,OBIA) technology. The accuracy of the analysis, recognition and extraction for the subsequent ground object are closely related to the accuracy of segmentation. therefore, the multi-scale segmentation technology has become a hot issue. The solution of this problem will be able to expand and improve the remote sensing image extraction methods, and improve the accuracy and efficiency of the target ground objects, so it has a very important scientific research and social application value.The existing research on multi-scale segmentation,object-oriented recognition and extraction of target information is mostly based on the commercial software, and it is still in the innovation phase of the application, with the appearance of remote sensing images, in order to expand the range of the object-oriented recognition and extraction of target information, a new method is needed to proposed, and then as a powerful supplement to the existing methods. also, the problem of the existing selection model of optimal segmentation scale are large computing, not match with the direct sense and pertinence is not strong and so on, research on the special data and application, intuitive and efficient selection model of the optimal segmentation scale is always become a big problem in multi-scale segmentation technology.Therefore, on the basis of the deep analysis of the oriented object multi-scale segmentation technique, this paper makes an initial attempt to solve the multi-scale segmentation technology and optimal segmentation scale selection method, the research contents and innovations are mainly in the following aspects:(1)The improvement of optimal segmentation scale selection method. In order to ensure the minimum internal variance of the target object and the maximum Moran index of among image objects, which means the divisibility among the objects is best. On the basic of the previous studies, a new method for select optimal segmentation scale is proposed. First, construct the segmentation quality evaluation function f(v, I) for p using the combination of the F(v) and F(I), which obtained by the normalization of v and Moran index. Then calculated the maximum value of f(v, I) which meet the p values in the range of [0,1] and p=0.5. Finally, draw function curves with segmentation scale change, when the discrepancy of the two functions value is minimum, the corresponding segmentation scale that is the optimal.(2)Study on the snow cover information extraction of the medium resolution remote sensing image. This paper using the object-oriented multi-scale segmentation technology, a multi-scale merge segmentation snow cover extraction model is proposed after merging image features, which can overcome the low extraction accuracy of the traditional pixel-based methods. The main ideas of the snow extraction model is: firstly, choose the image band IRS2, PC1 and CCD2 after pretreatment and combine them together, multi-scale segmentation of the combination of multi band image, building the snow extraction rule set of the optimal segmentation scale, thus extracting the snow cover information. After comparing and analyzing the results of NDSI method and the method of this paper, the results of the visual interpretation are used to evaluate the results of snow cover extraction quantitatively. The results show that the extraction based on object-oriented multi-scale segmentation method has a high reliability and accuracy.(3)Study on the landslide information extraction of high resolution image. There are some problems for the medium resolution images landslide extraction methods, such as it can't use landslide geological properties effectively and identify smaller landslide, many patches of extraction exist in map spot. This paper researches the geographical features and image features of landslide, based on multi-spectral high resolution image, a new multi-scale segmentation model is proposed. This model first select the optimal segmentation scale on the basis of the improved segmentation quality evaluation function, then construct landslides extraction rules sets and extract landslides information. Through the validation and analysis of the accuracy of landslide extraction information, the results showed that: the accuracy of landslide extraction is 75.86%. That means higher reliability and accuracy using this model. Besides, it also can provide reference for the methods and technology of the landslides extraction based on high resolution remote sensing image.