Abstract: Aiming at the huge data amount and pixel complex ownership of remote sensing images,a prototypes-extraction spectral clustering algorithm for remote sensing image segmentation was proposed.Firstly,the generalized fuzzy c-means algorithm was adopted to perform an over-segmentation of the image,and the obtained clustering prototypes were regarded as the representative points of segmentation regions to reduce the data amount of original image.Secondly,the similarity matrix between the representative points was constructed,and then the spectral graph partitioning method was utilized to cluster the representative points.Eventually,based on the clustering result of representative points,the image pixels were reclassified to obtain the final image segmentation results.There are three parameters in the prototypes-extraction spectral clustering algorithm.In order to overcome the parameter sensitivity and inherent randomness of this method,an ensemble strategy was further introduced into the method and its ensemble algorithm is presented.The segmentation experiments on artificial texture and remote sensing images show that this proposed ensemble method behaves well in segmentation performance.