A novel mixed pixels unmixing method for multispectral images

The key to mixed pixels unmixing is to determine a set of spectral endmembers that are representative of the surface components in the area covered by multispectral images. We propose a geometric method utilizing the spectrum of two or three bands to determine the endmembers under the assumption that the endmembers are the pure surface components that lie in the extremities of the scatter plot of 2- or 3-band reflectance. If there are more bands than endmembers, no additive noise and signature spectral matrix S has full column rank, we can estimate the endmemher abundances by unconstrained least squares method. Finally, the validity and feasibility of the algorithm are demonstrated by simulation results.