Research On Nonlinear Hyperspectral Imagery Unmixing

Hyperspectral imagery (HSI) gets the sacrificial signatures and is obtained by imaging spectrometers, which can capture the spatial information and near-continuous spectra at the same time, so hyperspectral imagery has high spectral resolution and the high recognition and classification capability of surface object. At the same time, the limit of spatial resolution,mixed pixels are widely existent in hyperspectral imagery, the mixed pixel is that one pixel may contain several disparate surface objects' spectral. Hyperspectral imagery applications are more and more widely, and the applications call for advanced techniques for hyperspectral data processing. Mixed pixels unmxing became a difficult and hot research. Spetral unmixng is the procedure by which the measured spectral of a mixed pixel decomposed into extracting several objects' endmember spectral or endmember extraction and their abundance. For avoiding blind unmixing, all classes included in mixed pixels are should to be known. This is the work of hyperspectral classification that prepare supervised information for unmixing procedure. Endmember extraction is that extracting the pure pixels information contained in the hyperspectral imagery, also prepare for the abundance unmixing procedure. Due to the research of hyperspectral classification and of endmember extraction is more extensive, this paper focuses on the abundance unmixing procedure.There is liner spectral mixing model and nonlinear mixing model, Linear mixing model(LMM) simplify the formation of mixed pixel. LMM assume that there are several pure materials spectral in the mixed pixel, and the mixed pixel spectral is linear combination of the pure materials spectral.Because of its easy to understand and its clearly physical meaning,LMM has a wide range of applications. However, in many cases, such as Hyperspectral image contain gravel, minerals, vegetation and water, etc. In the process of the formation of the mixed pixel, ignore nonlinear factors that may affect the accuracy of unmixing.1. Focused on the nonlinear mixed model based on polynomial, according to its shortcomings, putting forward a new nonlinear mixed model. The basic idea of nonlinear mixed model based on polynomial is the basis of the linear mixed model, adding nonlinear factors to simulate the formation of the mixed pixel. Due to the interaction of reflection and refraction between different surface objects, incident photons with other photons reflected twice or more, leads to the nonlinear mixed of a pixel. The present nonlinear models considers second-order interactions between different endmembers, but these models use Hadamard(term-by-term) product simulate second-order interactions. When the value of spectral pixel is too small or too big, the simulated second-order endmember also too small or too big compared with the true value of second-order interactions, unmixing accuracy will be affected.The new unmixed model proposed is on this basis,improving the second-order reflection,using two endmembers' spectral dot product then dividing their module as the representation of second-order spectral. AVIRIS imagery and simulation data was used for experiment with the proposed model,which proved that the new nonlinear unmixed is effective.2. A nonlinear algorithm based on LMS (Least Mean Square) Volterra serial is used for unmixing mixed pixel using the new nonlinear mixed model, the LMS criterion is applied to the filter in the unmixing iterative process. Since the LMS criterion don't need the square or differential in the iterative process, the algorithm has high computing speed and accuracy. In this paper,the Genetic Algorithm(GA) is used for nonlinear hyperspectral unmixing,GA is easy to use, simulated biological evolutionary theory, has high unmixing precision. But it's computing speed is too slow,so we only discuss the application of GA on nonlinear hyperspectral unmixing. In the experiment, by comparing the existing nonlinear algorithms and prove the effectiveness of LMS Volterra serial and GA nonlinear unmixing algorithm.3. Generally,the unmixing accuracy of nonlinear model is high compared with linear model, but not all cases are suitable for nonlinear unmixing model, such as the imagery of dense canopy. In the unmixing model,the Gaussian distribution characteristics of noise makes the observation pixel Gaussian distribution in the unknown parameters,unknown parameter estimates are observing pixel signal maximum likelihood estimation. Nonlinearity coefficient is assumed normal distribution,depending on the nonlinear coefficient in different models,set binary hypothesis. Using the generalized likelihood mixed pixel detection to determine what belongs to the unmixing model. This method will prove the unmixing accurary that verified in experiments using simulation data and AVIRIS imagery.