Research On Nonlinear Methods For Subpixel Target Detection In Hyperspectral Imagery

Target detection for hyperspectral imagery plays an important role in the research of remote sensing theory and its application.For hyperspectral imagery often has characters as follows:high dimension, nonlinear relationship between bands,mixed pixels,same materal with different spectrum,etc,then the performance of traditional linear hyperspectral target detection algorithms are tend to be influenced.Using kernel method is an effective way to process the nonlinear information,but still faces challenges.Besides,owing to the limitation of spatial resolution for hyperspectral imagery,as well as the complexity and defference among earth materials,some interested targets could only exist as subpixels,which means that there are only mixed pixels consisted of interested targets and other materals in the imagery. Hence, how to effectively raise the ability of target detection on subpixels is a critical problem for further development of hyperspectral remote sensing imagery.In addition,an hyperspectral imagery is usually not a mere collection of independent and identically distributed pixels, as assumed in pixel-based image classification,but it constitutes a structured domain:spatially close pixels should intuitively belong to the same class, especially with the advent of VHR sensors, where robust techniques are needed to deal with the high spatial correlation between spectral responses of neighboring pixels.And it has been paid high attention to effectively use spacial information of a given hyperspectral imagery.This dissertation mainly research on how to extend algorithms based on linear spectral mixture model to nonlinear feature space by kernel trick for resolving the difficulties in nonlinear unmixing.And mainly study spatial information extraction in a given hyperspectral imagery by using extended mathematical morphological methods, as well as incorporate the spatial information with a family of composite kernels,so as to improve the results of subpixels target detection in hyperspectral imagery.The major works and contribution of this dissertation are as follows:(1) A kernel-based oblique subspace projection (KOBSP) technique is proposed for nonlinear subpixel target detection in hyperspectral imagery. As a nonlinear version of the oblique subspace projection (OBSP), the OBSP is used in a high-dimension feature space after the pixels of input space being mapped into the feature space via nonlinear mapping. Applying the kernel trick, KOBSP does not need to know the actual nonlinear mapping. Experimental results of simulated and real data demonstrate that the proposed KOBSP approach outperforms the OBSP method in target detection, and improves robustness to noise.(2) A composite kernel signature space orthogonal projection (CKSSP) technique, which combined the spectral information with spatial information, is proposed for target detection in hyperspectral imagery mixed in nonlinear. The grey mathematical morphological is extended into multivariate mathematical morphological based on marginal ordering and reduced ordering respectively. The pixel distance is used as the ordering scale function to establish the reduced ordering. The extended mathematical morphological with multi-structure elements is used to extract the spatial information of the hyperspectral images. Combining the spectral and spatial information, the composite kernel function is constructed and improved according to kernel function definition. The target is detected by the composite kernel signature space orthogonal projection.Experiment results prove that CKSSP outperforms the KSSP in target detection.(3) Usually,we need to do endmember extraction before target detection,especially when the target information is unknown,neither the background information,then the selection of endmember will be very important,which influences the results of next steps,including target detection,classification,recognition,etc.So in this paper,a kernel unsupervised orthogonal subspace projection(KOBSP) is proposed to retrieve endmember automatically.Experiment results show that KOBSP algo rithm is an efficient and precise approach to retrieve endmembers.(4) The former research work are mainly done with MATLAB,then this dissertation also develops a software system about the subpixels target detection in hyperspectral imagery using C++developing language.The system could open hyperspectral imagery stored in GeoTiff format,choose a local area on the imagery,endmember extraction and make a abundance estimation(include SSP,OBSP;and the corresponding kernel vision called KSSP,KOBSP;also the new method-CKSSP).This software mainly focus on the application of subpixels target detection.