| Three-dimensional hyperspectral image contains both the spatial information and a wealth of spectral reflectance information, with wide band coverage, rapid non-destructive, full spectral information content and other characteristics. Hyperspectral image has abundant spectral information and can provide effective classification and identification for image features based on image analysis and pattern recognition algorithm. Therefore, hyperspectral imaging technology has a very wide range of applications in agriculture and the military. Due to the complex diversity of nature and the limited resolution of image sensors, one pixel may have mixed spectra of different materials. Such pixels are called mixed pixels. The mixed pixels greatly reduce the accuracy of the classification of surface features, because we cannot divide the mixed pixels into ascertain category. To solve the problems, the pixels are decomposed into meta-level sub-pixel. This process is referred to unmixing. The process of unmixing decomposes the mixed pixels into different basic composition of the material (endmember), and the proportion (abundance) of each basic material by geometry, statistics, modeling and other methods.Before mixed pixel unmixing, we need to create spectral mixed models. The basic models are linear mixed model and nonlinear mixed model. Linear mixed model assumes that the pixels are a linear combination of endmember and abundance, ignores the magnitude of the multiple scattering between different types of materials. Compared with the nonlinear mixed model, linear mixed model has the advantages of simple modeling, definite physical meanings and good scientific theory. Therefore, the commonly used unmixing algorithms are generally based on linear mixed model.This paper focuses on the study of algorithms for hyperspectral image linear unmixing. This paper begins with an overview of the three different types of unmixing technology based on the linear mixed model. In geometry unmixing technology, unmixing contains endmember extraction and abundance inversion. Endmember extraction uses the geometric orientation of hyperspectral data in multidimensional spaces, selection of endmember from the vertices of a simplex. Unsupervised unmixing technology uses hyperspectral data itself abundant spectral information and spatial information to unmixing in the absence of other external conditions. Sparse unmxing technology based on spectral library uses the spectral library combined with the sparse constraint to extract the endmember and abundance.On the basis of three types unmxing technology, this paper presents a fuzzy C-means unmixing (FCMU) method and endmember extraction method based on blocked VCA(vertex component analysis).(1) FCMU introduces cluster theory and fuzzy theory into hyperspectral unmixing model. Cluster theory makes endmembers which are obtained by unmixing have physical meaning and high accuracy. The results of fuzzy theory meet non-negativity constraint and sum to 1. In the case of the number of endmember is known, FCMU using objective function and iterative algorithms simultaneously obtain endmember and relative abundance. Experimental results show that endmember accuracy has been greatly improved by FCMU.(2) This paper presents endmember extraction method based on blocked VCA to solve the problem of missing endmember and low extraction accuracy caused by VCA. The basic idea of this method is that, complex environment hyperspectral image is segmented into a plurality of relatively simple image portions by certain unsupervised classification method, and then use VCA in each regional block. This method reduces the effect of noise on the extent of the global image of the algorithm and avoids missing the main endmember. In the experimental part, use two types of hyperspectral images with different complexity. Experimental results show that, the endmember extraction method based on blocked VCA method has greatly improved the accuracy of the endmember and the precision of inversion abundance. |
PDF Full Text Request