基于边缘的多光谱遥感图像分割方法

从Marr视觉计算理论和Tobler地学第一定律出发,提出了基于边缘的多光谱遥感图像分割方法.在基于边缘的多光谱遥感图像分割方法中,由边缘检测、边缘综合、边缘生长、区域标号等环节组成.该遥感图像分割方法在可视化开发平台Delphi中予以编程实现.将之应用于日本熊本市（Kumamoto）的Quickbird多光谱遥感图像中,并与多种遥感分割算法进行了比较：（1）从多光谱遥感图像各波段亮度信息利用的程度上看,提出的遥感图像分割方法能充分利用多波段亮度信息;（2）从遥感图像分割结果上看,由于分别对不同的波段进行边缘检测,并在此基础上进行边缘综合、边缘生长,遥感图像中的细节特征得到了充分体现,遥感图像分割效果更理想;（3）从计算复杂度和计算效率上看,基于边缘的多光谱遥感图像分割法较其他分割方法有一定的优势.

An Algorithm of Multi-spectral Remote Sensing Image Segmentation Based on Edge Information

According to the first geographic law of Tobler and the Marr's machine vision theory,an algorithm to segmenting multi-spectral remote sensing imageries has been put forward based on the edge information extracted from them.This algorithm consists of four steps listed below:(1) Detecting edge information in each band of remote sensing imageries using a improved Canny method;(2) Integrating edge information in each band of remote sensing imageries into a binary image by methods such as overlay technique in GIS technology,and then thinning edges in the binary image by techniques of mathematical morphology using a rectangle probe;(3) conjoining disconnected edges according to the characteristics of processing edge such as length,direction and so on,to close each region;(4) at last,labeling region and remove abundant edges that do not compose region.Then,the multi-spectral remote sensing imageries of Quickbird covering the Kumamoto city,Japan,have been taken as a case study for this algorithm,and the result has been compared with other segmentation algorithms such as Multi-Threshold Gray Slice Approach(MTGSA),Iterative Self-Organized Data Analysis Technology Algorithm(ISODATA) image segmentation algorithm,Watershed Segmentation Algorithm(WSA),Fractal Net Evolution Approach(FNEA) and so on.Based on the comparative analysis,conclusions could be drawn out that(1) In term of utilizing brightness information of each band,the scope that the algorithm proposed in the paper is the most comprehensive one,and MTGSA and WSA can only use single band of multi-spectral remote sensing image;(2) The result of this algorithm could be the most satisfied,as it detects edge information of each spectral band respectively,and then integrates as well as connects them together,maximally digging out the detailed features in remote sensing imageries;(3) In the aspect of computational duration,this algorithm is relatively a bit faster than others under the same environment.As the same as the other three approaches,the algorithm proposed in the paper has also confronted the common difficulty of how to confirm the coefficient in the image segmentation procedure.