一种新的高光谱遥感图像超像素分割方法

A New Superpixel Segmentation Method for Hyperspectral Remote Sensing Images

为了解决简单线性迭代聚类算法在高光谱遥感图像超像素分割任务中分割精度较低的问题,提出一种基于多级线性迭代聚类结合改进标签传播算法(LPA)的新的无监督高光谱遥感图像超像素分割方法。首先,扩充简单线性迭代聚类(SLIC)的适用范围至多通道对高光谱图像进行超像素初分割;然后,对色彩标准差较大的超像素进行多级迭代细致分割,引入基于局部二进制模式的高光谱遥感图像纹理特征提取方法计算高光谱图像纹理特征并融合多段光谱特征计算超像素间相似度以构建带权图网络;最后,改进LPA社区发现方法进行超像素合并,将改进的标签传播算法运用于超像素合并可以得到更加稳定准确的超像素合并效果,提高超像素分割精度。将该方法与多种方法进行比较,结果表明,该方法对高光谱遥感图像的超像素分割结果更准确,超像素边缘更贴合真实地物边界,能有效改善高光谱遥感图像超像素分割中精度较低的问题。

In order to solve the problem of low segmentation accuracy of simple linear iterative clustering algorithm in hyperspectral remote sensing image superpixel segmentation tasks,a new unsupervised hyperspectral remote sensing image superpixel segmentation method based on multi-level linear iterative clustering combined with improved label propagation algorithm(LPA)is proposed.Firstly,we expand the applicability of Simple Linear Iterative Clustering(SLIC)to perform superpixel initial segmentation on hyperspectral images through multiple channels,and then perform multi-level iterative and detailed segmentation on superpixels with large color standard deviations.A texture feature extraction method based on local binary mode for hyperspectral remote sensing images is introduced to calculate the texture features of hyperspectral images and fuse multiple spectral features to calculate the similarity between superpixels to construct a weighted graph network,Finally,the LPA community discovery method is improved for superpixel merging,and the improved label propagation algorithm is applied to superpixel merging to obtain a more stable and accurate superpixel merging effect,improving the accuracy of superpixel segmentation.Compared with various methods,the proposed method has more accurate superpixel segmentation results for hyperspectral remote sensing images,and the superpixel edges are more closely aligned with the real boundary of land objects.It can effectively improve the problem of low accuracy in superpixel segmentation of hyperspectral remote sensing images.