Abstract: A hierarchical segmentation method is proposed for multispectral imagery by combining extended watershed transform and dynamics of contours. First, the gradient magnitude of an image is computed, and then local irrelevant minima are eliminated. After computing dynamics of contours, hierarchical (multi-level) segmentation results are produced by thresholding. The proposed method is evaluated with a high spatial resolution image of the urban area, and compared with the existing image segmentation method. The results show that the method can get hierarchical segmentation results, and segmente most of image objects. The segmentation results give the hierarchical representation of land cover types, which is helpful for further object based analysis.

Key words: multispectral imagery, hierarchical segmentation, watershed transform, dynamics of contours

摘要： 综合扩展的分水岭变换和分水线动态范围，提出了适用于高分辨率多光谱图像的多尺度分割方法。该方法首先计算多光谱图像的梯度，接着去除梯度图像上不相关的局部最小区域，然后计算分水岭的动态范围，通过设置不同的阈值，得到多层次（多尺度）分割结果。采用城市地区的高分辨率数据来评价所提出的方法，并与现有的图像分割方法进行比较。实验结果表明，该方法可生成多层次的分割结果，并较准确地分割出地物特征。分割结果反映了地物的多层次信息，为后续的面向对象分析提供了有意义的参考。

关键词: 多光谱图像, 多尺度分割, 分水岭变换, 分水线动态范围