Hyperspectral images have been widely used in target dectection terrain classification and so on owing to its rich spectral information. Classification, being the fundamental step to further explore the hyperspectral images, attracts wider concern. The spatial information describes the connections between pixels with its spatial neighbors which can help to solve the problems like metameric substance of same spectrum, metameric spectrum of same substance and insufficient labeled samples with a high dimension while the spectral information cannot handle well. The traditional preprocessing uses a structure element to obtain the spatial neighbors and assist the last classification with the extracted spatial features. It is obvious that the structure element matters, however one cannot find a suitable size to meet all demands. For dealing with this, a method combing watershed segmentation with composite-kernels support vector machine (SVM) is prposed. It is the characteristics of over segmentation that we use to get a self-adapting spatial neighbors, containing less dissimilar pixels and being more discriminant for every pixel, then we fuse the spatial features and the spectral through the composite-kernels SVM and give a reliable judgement. Experiments show that the proposed method can make a better use of the spatial imformation and achieve a high accuracy with limited training samples.