Abstract: High-dimensional feature vector is defectivedue to partial neighborhood superposition when extracting spectral information in hyperspectral image, and the pixels in the local region of the image have the problems are easily affected by density difference and homospectral dissimilarity, a hyperspectral image classification algorithm based on spatial spectrum super-pixel fusion kernel extreme learning machine(SSKELM) is proposed.The first principal component of the spectral space is super-pixel segmented, and each super-pixel is regarded as a shape adaptive region. Using the spatial information, the kernel weight fusion within and between super pixels to obtain the pixel category label, and taking advantage of the linear separability of kernel function in high-dimensional hyperplane data and the limited conditions of limit learning machine random hidden layer output matrix and its optimization algorithm, the spatial spectrum pixel points are fused and trained to form a new matrix sample output. Compared with the real results, the overall accuracy of the test results conducted through the University of Pavia and Indian pines data sets, the overall accuracy OA value is improved by 1.76% and 2.80% respectively compared with other algorithms. It effectively verifies that the proposed method has a certain value in HSI classification.