Automatic endmember bundle unmixing methodology for lunar regional area mineral mapping

The mineral distribution on lunar surface can contribute to studying lunar evolution, while abundance quantification is still challenging. Unmixing on spectral reflectance data is an effective way for mineral resource explanation, especially in hardly accessible area. In regional area unmixing, some existing unmixing models mainly rely on spectral libraries, which limits the scene adapability in the absence of some prior information. Meanwhile, spectral variation is a common phenomenon but often neglected, which may lead to subsequent abundance inversion errors. In this paper, we address a novel automatic image-based endmember bundle unmixing model, which is called AEBU, to solve these problems. Differently from many unmixing algorithms using a single spectrum to represent a type of mineral, we accommodate spectral variation and construct a set of spectra, i.e., endmember bundle, to represent each material, which will allow for comprehensive endmember expression. The endmember bundles are extracted from the imagery and regarded as a spectra catalog for abundance inversion to avoid the dependence on spectral library. The proposed AEBU model contains two major steps: image-based endmember bundle construction and abundance inversion. To construct endmember bundles effectively, we use pixel-wise sparse representation to extract image pixels as endmember candidates, and then analyze the shape feature of candidate spectra to separate endmember bundles. In abundance inversion, we consider the extracted endmember bundles as existing spectra library and propose a block sparse representation-based algorithm to automatically select reasonable endmembers for per-pixel unmixing. The performance of AEBU is compared with the state-of-the-art bundle unmixing algorithms on simulated lunar data. The experimental results demonstrate excellent performance of the proposed AEBU. Finally, we map the mineral distribution on lunar regional areas by AEBU using interference imaging spectrometer (IIM) data collected by ChangE-1 and moon mineralogy mapper (M3) data collected by Chandrayaan-1, and unmix the Cuprite data to show more application of AEBU.