Abstract: A novel remote-sensing image classification algorithm based on semi-supervised Kernel-Fuzzy C-Means and local spatial information is given to resolve the problems of the traditional Fuzzy C-Means algorithm which would produce “local optimization” and is sensitive to noise in some extent. The algorithm uses the reflectivity of the surface to classify based on the eigenspace; it considers the spatial relationship of pixels to the corresponding objective functions which contains neighbor information, modifies the value of fuzzy membership degrees based on the weighted spatial distance during the clustering iterations; introducing the Kernel-theory, Kernel-Fuzzy C-Means is put forward which resolves the problem of the nonlinear when it is used in remote sensing imagery, replacing the original Euclidean distance with kernel-induced distance to optimize the character of the image sample. Using semi-supervised classification technology, the algorithm makes use of the few labeled data to improve the classification precision. Experimental results show that the novel algorithm can get a better classification result and get a high precision with small samples and enhance their insensitiveness to noise. The iteration time is reduced and can enhance efficiency.

Key words: remote sensing, semi-supervised image classification, spatial relationship, Kernel Fuzzy C-Means（KFCM）

摘要： 针对传统的模糊C-均值在遥感图像分类时容易产生局部最优现象以及对噪声过于敏感等问题，提出了一种基于半监督、核函数及空间邻域信息的模糊C-均值遥感图像分类算法。该算法基于遥感图像的光谱特征空间，根据地物的地表反射率大小进行聚类；在聚类迭代过程中，考虑到像素单元空间邻域的相关性，根据空间邻域信息加权调整像素点的隶属度大小；引入了核理论，解决遥感图像分类的非线性问题，使用内核诱导距离取代原模糊C-均值中的欧氏距离，优化图像样本特征；算法还使用了半监督分类技术，充分利用少量的已知标记信息，达到提高分类精度的目的。实验结果表明，该算法能有效提高分类精度，有效抑制噪声干扰，减少了迭代次数和时间。

关键词: 遥感, 半监督图像分类, 空间邻域, 核模糊C均值算法