结合马氏距离的区域化模糊聚类遥感图像分割

Remote sensing image segmentation algorithm with regional fuzzy cluster and Mahalanobis distance

传统模糊聚类算法通常将像素光谱测度间的欧式距离作为相似性准则,其仅适用于同质区域内像素光谱测度呈对称分布的图像.同时,基于像素的图像分割算法极易受噪声干扰,导致该类算法难以适用于高分辨率遥感图像分割.因此,提出结合马氏距离的区域化模糊聚类遥感图像分割算法,该算法以Voronoi多边形为基本单元,令多边形内所有像素与该多边形具有同一类属性,并采用马氏距离定义非相似性测度.此外,在目标函数中引入马氏距离规则化项以控制聚类尺度,进而构建区域化高分辨率遥感图像分割模型.通过对合成及真实遥感图像分割结果的定性及定量分析,证明了提出算法具有较强的鲁棒性.

The Euclidean distance between pixels spectral measurements, usually taken as the similarity principle in traditional fuzzy cluster algorithm, is only applicable to the images in which spectral measurements of pixels in homogeneous region are symmetrically distributed. On the other hand, the image segmentation algorithms based on pixels are easily affected by noises and can not be employed in high resolution remote sensing image segmentation. So, the remote sensing image segmentation algorithm with regional fuzzy cluster and Mahalanobis dis- tance was proposed. It uses Voronoi polygons as the basic unit and all pixels in the polygon has the same cluster attributes. Then Mahalanobis distance was used to define the dissmilarity and a region-based segmentation model for high resolution remote sensing image was established by integrating Mahalanobis distance regularization term in objective function to the control cluster scale. The qualitative and quantitative analyses results of synthetic and real images segmenta- tion indicate that the proposed algorithm has strong robustness.