Spectral partitioning and fusion techniques for hyperspectral data classification and unmixing

Hyperspectral images are characterized by their large contiguous set of wavelengths. Therefore, it is possible to benefit from this `hyper' spectral information in order to reduce the classification and unmixing errors. For this reason, we propose new classification and unmixing techniques that take into account the correlation between successive spectral bands, by dividing the spectrum into non-overlapping subsets of correlated bands. Afterwards, classification and unmixing are performed on each subset separately, such as to yield several labels per pixel in the classification case, or abundances in the unmixing case. Then, several fusion techniques are proposed to obtain the final decision. Results show that spectral partitioning and appropriate fusion allow a significant gain in performance compared to previous classification and unmixing techniques.