Endmember extraction is the key step of the mixed pixel decomposition in hyperspectral remote sensing images. Traditional endmember extraction algorithms ignore the spatial correlation and nonlinear structure of hyperspectral images, which restricts the accuracy of endmember extraction. Aiming at the spatial relationship and nonlinear structure of hyperspectral images, a nonlinear endmember extraction algorithm based on homogeneous region segmentation is proposed. The hyperspectral image was divided into several homogeneous regions using superpixel segmentation method, and the manifold learning method was used to ensure nonlinear structure of hyperspectral images, extracting preferred endmembers within homogeneous regions. Simulation data and real hyperspectral image experiments show that the algorithm in this paper can guarantee the nonlinear structure of hyperspectral data, and the endmember extraction results are better than other traditional linear endmember extraction methods. In the case of low signal-to-noise ratio, it can maintain a good endmember extraction results.