On the incorporation of spatial information to endmember identification algorithms without the pure pixel assumption

Spectral unmixing is a commonly used technique in hyperspectral data exploitation. It expresses the measured spectral signature of each pixel in the data as a combination of spectrally pure constituent spectra, called endmembers, and a set of correspondent fractions, or abundances that indicate the proportion of each endmember present in the mixture. Over the last years, several algorithms have been developed for automatic endmember identification by assuming the presence of at least one pure spectral signature for each distinct material. However, this assumption often does not hold in practice due to spatial resolution, mixing phenomena, and other considerations. In this paper, we investigate if spatial information can assist the endmember searching process conducted by algorithms that do not assume the presence of pure pixels in the hyperspectral data. For this purpose, we use recently developed spatial pre-processing techniques that do not require modifications in the subsequent endmember identification process, conducted in this work using minimum volume enclosing algorithms. Our experimental results, conducted using both simulated and real hyperspectral data sets, reveal that spatial information can be beneficial to guide the endmember identification process when pure pixels are not assumed to be present in the hyperspectral data.