Existing endmember extraction algorithms are mainly based on the convex simplex hypothesis. However

the cover types in certain endmembers are not single

which will affect the unmixing accuracy of mixed pixels when performing abundance inversion. In this paper

we propose to determine the nature of the hyperspectral pixel based on the high-resolution remote sensing image. First

a spectral relatively homogeneous vector diagram of blocks is superimposed on the hyperspectral image after the highresolution image segmentation. Second

spatial relations analysis is performed to find the hyperspectral pixels that are within the blocks

which is called a quasi-endmember. Finally

endmember extraction is performed to find endmembers from the quasi-endmember set. The experimental results demonstrate that our approach can reduce the root mean square error of the extraction results by about 20%.