Material identification on hyperspectral images using Bayesian source separation

Identification of materials in a scene observed by an imaging spectrometer is a common problem in Planetology. Usually the pixel size is larger than the typical size of material change over planet surfaces, leading to both linear and non-linear spatial mixing models. We propose here an unsupervised approach based on linear source separation to estimate the pure spectra of the components present in the observed scene and their abundances in each pixel. Previous application of this approach to Martian ices[1] have shown its relevance when the positivity of both the pure spectra and the abundances is taken into account. We propose here to apply this approach to detect Martian minerals and we show that adding the sum-to-one constraint (or additivity constraint) on the abundance vectors leads to an improvement of the estimation performances.