Research And Implementation Of Multi-Scale Remote Sensing Image Segmentation

With the rapid development of information and space technology, aiding by continuous improvement of the satellite spatial resolution, high-resolution remote sensing image is widely used in various fields. Remote sensing image segmentation is a kind of technology and process to extract the target through processing and analyzing the remote sensing image. There are no reliable ways to guide the remote sensing image segmentation, because of its features, such as rich image detail, significant geometry feature and complex target characteristics. What's more, it hinders the application of remote sensing image segmentation. Scale is a fundamental factor in remote sensing. Different objects have different optimal spatial scales, and we need to choose the appropriate scale of target segmentation to get the right the surface features of target objects.Considering the feature of remote sensing image, this thesis has studied on and improved the traditional image segmentation algorithm. We made it adapt to the feature of remote sensing image, and applied it to remote sensing image segmentation. The main work of this thesis is as follows:Firstly, a survey of remote sensing image segmentation methods was presented. Through summarizing the advantages and disadvantages of different methods, the survey provided basis for selecting a suitable image segmentation method under different conditions and in different applications of image data;Secondly, we analyzed the principle of the watershed algorithm and its mathematical model. Through discussing the classical V-S immersion watershed, we pointed out its defect of over-segmentation and the improvement of existing methods. We also analyzed the theory of mean shift algorithm and its application in image segmentation. Thus, we provided the theoretical for multi-scale remote sensing image segmentation.Thirdly, aiming at to achieve the needs in multi-scale with remote sensing image segmentation, we proposed watershed-based and mean-shift-based methods to implement multi-scale remote sensing image segmentation. The former method adopts multi-scale morphological gradient operator to detect edge and obtain the fuzzy C means clustering on initial segmentation with watershed transform. With different numbers of cluster centers and different parameter setting in watershed transform, we achieved multi-scale segmentation based on panel. The latter method introduced multi-scale segmentation method to the optimization phase in mean shift algorithm, we obtained the different scales results though the multi-scale merger of segmentation. Experiments showed the two methods can obtain good results in multi-scale segmentation.Finally, a demo system of remote sensing image processing has been designed and implemented. The demo system has some common functions, such as image segmentation, edge detection, image enhancement and multi-scale remote sensing image segmentation and so on. Multi-scale remote sensing image segmentation has been applied in urban green space segmentation.