Spectral Unmixing-oriented Fast Processing Techniques For Hyperspectral Image

With the improvement of the spatial and spectral resolution, the hyperspectral sensors provide plenty of valuable high accuracy data after contiguous and detailed observations to the same scene, which bring to more chances for the processing and application for hyperspectral data. However, the huge hyperspectral data and increasing complexity of processing algorithms also bring challenge to practical application. Nowadays, most of algorithms for hyperspectral images need to be processed under the platform which can provide fast data analysis and processing capabilities and modular business processing modes. And unmixing as one of important branches of hyperspectral image processing is also need efficient para llel algorithms and special high performance hardware to accelerate implementation. Consequently, an advanced fine-grained multithreads parallel technology based on CPU/GPU coordination is introduced to accelerate hyperspectral interpretation processing. Focus on fast processing of hyperspectral unmixing, this thesis researches on some related techniques, which can mine the potentials of hyperspectral data and extend their application fields. Therefore, the research of spectral unmixing-oriented fast processing techniques for hyperspectral image is of theoretical significance and research value.Firstly, this dissertation starts from parallel analysis and fine-grained parallelism design of hyperspectral unmixing algorithms. The previous studies focus on storage structure of hyperspectral data, tasks dependencies and data dependencies in the typical unmixing algorithm. On the basis of this, the thesis carries out related research work. And then, further researches aim at parallel algorithm design and implementation based on CPU/GPU coordination. To effectively evaluate algorithm acceleration effect and hardware performance, several groups of experiments are done to explore the main influence factor of hardware prosperities, algorithm itself prosperities, data size and access mode. In order to make use of all kinds of optimization strategies, the PCAM(Partitioning-Communication-AgglomerationMapping) theory is introduced to design the assessment model. Moreover, some other influence factors are also considered overall such as storage structure of hyperspectral data, data transfer mode(synchronous/asynchronous), fine thread allocation and computational complexity of algorithm. Based on these, a performance analytical model of CPU/GPU cooperative work is propose d in which hyperspectral image, algorithm and hardware are input parameters of wavelet neural network. By this model, the optimizational program achieves global speedup accurately in the pattern of CPU/GPU cooperative computing.And then, fast implementations of preprocessing and similarity measure algorithms are mainly researched for hyperspectral image. The former is applied in the preprocessing phase, while the latter is adopted in the assessment of the accuracy of spectral unmixing. In the aspect of dimension reduction of hyperspectral image, this thesis mainly focus on fast processing of dimensionality reduction based on principal component analysis and acceleration of CPU/GPU coordination. In order to reduce computational time of PCA, an improved algor ithm which is used to calculate Eigen values and Eigen vectors is adopted based on QR iterator and NIPALS methods. When the number of components is little, the NIPALS method has higher efficiency. But the components lack of orthogonality using NIPALS method. In order to solve the problem, a new improved NIPALS-PCA method is proposed upon Gram-Schmidt process. Using this method, the extracted components maintain orthogonality without increasing computation complexity evidently. Another work concentrates upon fast processing of spectral similarity measure algorithm. The previous studies focus on parallel design and implementation of Euclidean distance, spectral angel mapping, spectral information divergence and combination between them. In order to increase the computational accuracy and speed of the SAM-SID algorithm, a novel method is proposed based on kernel transformation. Furthermore, GPU is also used to speed up the improved algorithm without reducing the matching accuracy. Above works establishes the bas is of fast processing of hyperspectral unmixing chain.On the basis of the above researches, parallel design and validation are carried out around hyperspectral unmixing based on GPU acceleration. In the aspect of parallel design of hyperspectral unmixing algorithms, a complete parallel implementation scheme is proposed to accelerate hyperspectral processing based on Berkeley parallel design pattern combined with the character of hyperspectral data storage, processing algorithms themselves in accordance with pixels and hardware architectures of SIMT. Furthermore, this thesis focuses on parallel design about linear spectral unmixing which include endmember extraction algorithms based upon space geometrical structure and abundance estimation algorithms based on fully constrained least squares. An automatic endmember extraction method has been put forward based on the geometrical properties of the simplex in high-dimensional feature space which also be used in N-FINDR and SGA methods. In new method,the number of endmembers is initialized combined VD and Hy Sime methods. At the same time, in order to accelerate processing speed, some optimization strategies are presented under the computational accuracy requ irements. In order to reduce the algorithm complexity, a fast implementation method of MSVA is proposed based on partitioned determinant operations to get further speedup. Above works can realize automatic and fast spectral unmixing.At last, the algorithms of spectral unmixing chains are implemented under the platform of CPU/GPU coordination. Furthermore,an evaluating method in which computational accuracy, acceleration effect and hardware performance are considered simultaneously is established for the effect of parallel implementation of hyperspectral unmixing algorithms. And then, the proposed method not only can be used to test and verify the rationality and effectiveness of the fruit in this thesis but also can be used to guide to achieve further speedup.