Research On Unmixing Technology In Hyperspectral Remote Sensing Imagery

Hyperspectral remote sensing imagery has the features of high spectral resolution, combining image with spectrum and many spectral bands, which can provide rich information of the earth surface. Therefore, hyperspectral imagery has been widely applied to land use, resource suvey, natural disaster monitoring and other fields. However, due to the spatial resolution of the imaging sensor and the variability of ground surfaces, mixed pixels are widespread in hyperspectral images, which go against pixel-level data analyisis and processing. In order to better use hyperspectral data and improve the precision of remote sensing application, it is necessary to decompose mixed pixels into a collection of endmembers and their corresponding abundance fractions.In hyperspectral images, spatial distributions of materials have different characteristics. There are usually pure pixels for prevalent materials, while less widespread materials usually exist in the form of subpixels. The difficulty brought by this kind of images is how to extract endmembers of all the materials accurately and unmix the images effectively. When interactions among different matrials in close proximity occur at a microscopic level or multiple interactions among the scatters at the different layers occur in a multilayered secene, nonlinear spectral mixing effects can not be neglected. In addition, incompleteness of spectral library, difficulties of actual measurements in remote areas and other factors make automatic extraction of endmembers be difficult. This dissertation centers on unmixing technology of hyperspectral remote sensing imagery, and carries out research on the aforementioned problems existing in hyperspectral imagery. The major works around these issues are as follows:(1) Considering the deficiency that current unmixing algorithms can not simutaneously extract endmembers of pure-pixel and subpixel materials well, a finite spectral unmixing method by combing the theory of convex geometry with nonnegative matrix factorization (NMF) is proposed. First, the endmember extraction methods based on the theory of convex geometry is used to generate a candidate pixel set of pure-pixel endmembers. Based on the different distribution characteristics of pure pixels and mixed pixels in a local region, the spatial purity indices of the candidate pixels are calculated to determine endmebers of prevalent materials, and then the NMF method is adaptively modified to be partial NMF (PNMF) algorithm. The objective function is constructed, and the iterative solution is also derived. In the end, subpixel endmembers and abundances of all the endmembers are obtained by PNMF algorithm. Experimental results demonstrate that the proposed finite spectral unmixing method can make up for the shortage of existing unmixing methods, and extract pure-pixel endmembers as well as subpixel endmembers effectively.(2) Excess number of endmembers in hyperspectral imagery makes nonlinear decomposition of mixed pixles comparatively time consuming and abudance estimation less accurate. For these problems, a technology of nonlinear decomposition of mixed pixels is developed by incorporating spatial information. First, a preliminary classification map of hyperspectral images is obtained based on abundances estimated by the unconstrained least square method. Then, according to the spatial correlation of objects between neighboring pixels, the actual endmember set for each pixel is determined by using appropriately sized local windows. Finally, the bilinear mixture model is improved, and the decomposition problem for the mixed pixel based on the improved bilinear model is transformed into a quadratic programming problem, which can be solved with an active method. Comprehensive experimental results on synthetic and real hyperspectral images demonstrate that by using the spatial neighborhood information, the computing cost of nonlinear spectral mixture analysis (SNSMA) is reduced sharply and the estimation accuracy of abundances is also improved.(3) As the nonlinear spectral mixture model is complicated and the model parameters are always difficult to be determined, using the kernel function methods, the original hyperspectral data is mapped into high dimensional feature space by nonlinear map, and mixed pixels in feature space can be described by the linear spectral mixture model. By combing NMF algorithm and the kernel learning theory, the kernel NMF (KNMF) algorithm is constructed, which can unmix the mapped data in high dimensional feature space. Meanwhile, to make the unmixing results tend to be unique, considering the simplex structure of hyperspectral data in feature space and the smoothness of abundance distributions, two auxiliary constraints, namely simplex volume constraint and abudance smoothness constraint, are introduced into KNMF. The experimental results on synthetic and real hyperspectral images demonstrate that compared with linear unmixing algorithms, the proposed constrained KNMF (CKNMF) algorithm can estimate the endmember spectra and their corresponding abundances better.