Hyperspectral Unmixing Based On Homogeneous Region Analysis

Hyperspectral imagery is acquired by high-spectral-resolution imaging sensors, containing hundreds of contiguous narrow spectral band images, with lots of valuable spectral, spatial and radiation information. However, due to the distribution complexity of the materials and the low spatial resolution of the sensors, there are lots of mixed pixels in hyperspectral imagery, which is the biggest obstacle of quantitative analysis and application of hyperspectral data. Hence, hyperspectrral unmixing is one of the key issues of interpretation and application of hyperspectral data.Hyperspeetral unmixing, which decomposes a mixed pixel into a collection of constituent spectra, or endmember, and their corresponding fractional abundance, is a crucial preprocessing step for hyperspectral imagery applications. Many hyperspectral unmixing methods have been proposed in recent years, with mainly two classes:1) the two-stage methods, which extract or separate endmembers firstly, then, estimate the abundance based on the extracted endmembers. Most of tradintional methods are in this way.2) The single stage methods, which obtain the endmembers and their abundances simultaneously, such as the blind source separation methods. However, most approaches have been designed from a spectroscopic viewpoint and thus, tend to neglect the existing spatial correlation between pixels. This paper argues that the hypespectral unmxing problem should be performed taking advantage of the spatial information in the hyperspectral imagery, obtaining the endmember and their abundances using both spatial and spectral information in a combined manner. Then, this paper presents a new hyperspectral unmxing technique framework based on homogeneous region analysis, described as following:(1) A new spectral similarity measure, i.e. the Spectral Pan-similarity Measure (SPM) is presented, based on geometric distance, correlation coefficient and relative entropy. SPM objectively quantifies differences between spectra in three spectral features, the vector magnitude, spectral curve shape and spectral information content. The experimental results demonstrate that the new spectral similarity measure is more effective in spectral discriminatory power and spectral identification uncertainty, and is used in the various stages of the hyperspectral unmixing for spectral similarity analysis.(2) A novel measure, termed as Homogeneous Index (HI), is defined for quantitative analysis of the spectral similarity between the pixel and its surrounding pixels, with the mathematical model based on the SPM. In order to integrate the imagery spatial information in the hyperspectral unmixng, the image is divided into homogeneous region and transition region with different HI. The spectrum of a pixel is similar to the spectra of its neighboring pixels in homogeneous region, and distinct from the spectra of its adjacent pixels in transition region. Intuitively, the transition between different land-cover classes would likely contain some mixed pixels. The spatial and spectral information of the homogeneous region will be used to extract endmember spectra and define new constraints in hyperspectral unmixing problem.(3) A new endmember extraction algorithm is proposed using both of spatial and spectral information based on the homogeneous region, termed as Homogeneous Region based Endmember Extraction (HREE). HREE method exploits three facts:1) the endmembers are the vertices of a simplex,2) all the pixels of the hyperspectral imagery are convex combinations of endmember spectra, and3) all the pixels of the homogeneous region are convex combinations of endmember spectra. Candidate endmember spectra indexes are got through OSP method in the feature space of the homogeneous region. Then the candidate endmember spectra are refined with spatial context and spectral information. The experiments show that the HREE outperforms the other popular endmember extraction algorithms.(4) Give and discuss two propositions, i.e. the Abundance Sparseness proposition and the Abundance Smoothness proposition, based on the spatial and spectral feature of the homogeneous region. Then a new nonnegative matrix factorization method is proposed with abundance smoothness constraint, termed as Constrained Nonnegative Factorization (CNMF). The objective function of CNMF is given, and the mathematical convergence of the iteration rules proven. Experiments results based on both synthetic and real hyperspectral images demonstrate CNMF is an effective unsupervised technique for hyperspectral unmixing.(5) The abundance retrieval problem is studied in the case of known endmember spectra. A new algorithm is presented to determine optimal per-homogeneous region endmember sets from the image endmember set through iterative implementation of spectral mixture analysis. Then, a new abundance retrieval technique is given based on the neighbor information, i.e. the optimal transition region endmember sets is determined by the connected homogeneous region, and the optimal pixel endmember sets by the region belonging to. The experiments show that the abundance retrieval accuracy is improved.