Unsupervised unmixing analysis based on multiscale representation

Abstract

Unsupervised unmixing analysis aims to extract the basic materials, the so called endmembers, and their abundances from a hyperspectral image. Unmixing is usually performed by pixels-only techniques that do not take into account the spatial information and generally require a priori estimate of the number of endmembers. Recently, several spatial-spectral unmixing techniques have been developed. However, most of these techniques depend of spatial kernels or windows to include the spatial information in the unmixing analysis. In this work, a new unmixing approach based on multiscale representation is developed. The proposed technique extracts spectral signatures and spectral endmember classes from hyperspectral imagery in an unsupervised fashion. A multiscale representation of the hyperspectral images is obtained using nonlinear diffusion. Then, spectral endmembers are automatically identified using multigrids methods to solve the diffusion partial differential equation. The multiscale representation and multigrids allows to avoid the use of spatial kernels. Once the spectral endmembers are identified, similar spectra are clustered to build spectral endmember classes thus accounting for the spectral variability of the materials along the unmixing analysis. A comparison with other unmixing methods shows that the proposed unsupersived unmixing approach outperforms traditional spectral techniques. Capabilities of the proposed approach were validated and assessed using simulated imagery and real imagery collected with the AVIRIS and AISA sensors over different landscapes.