Research On Feature Extraction Of Hyperspectral Image Based On Sparse And Manifold

With the development of remote sensing technology and imaging spectrometer,hyperspectral image is widely used in many fields.However,due to the number of band hyperspectral image has numerous,large amount of data,with the development of imaging spectrometer data obtained to improve resolution.so that the original data in the use of the traditional algorithm for processing,often encounter a long calculation time,the traditional algorithm applicability is limited.At present,many algorithms have been validated and applied a strong.For the primitive reduction of high-dimensional data is one of the methods to solve the problems.In addition,the hyperspectral data acquired by imaging spectrometers in time,inevitably will obtain a non-linear information,linear feature extraction algorithm of traditional obviously cannot obtain these nonlinear factor,and the kernel based algorithms because of its physical meaning is not clear and the computation complexity and other issues.In recent years,manifold learning and sparse representation theory is a hot research topic,this paper first analysis the advantages and disadvantages of the traditional algorithm,proposed several improved algorithms and the problem mentioned above is studied in depth.The research works of this paper are as follows:The first feature of hyperspectral image extraction algorithms,a detailed summary of the Traditional theory of algorithms,the calculation process,experiments were performed using the features of real data,verify the effectiveness of the algorithm.Then aiming at the semi supervised feature extraction algorithm in the existing problems,and proposes a new semi supervised local sparse embedding feature extraction algorithm.To overcome the classical sparse representation for sparse coefficient algorithm complexity and can keep the local geometric information of the data in the solution of sparse graph fast at the same time.Aiming at the high spectrum image semi supervised drop open select neighbor parameters based on manifold learning dimension algorithm,and local access to information label data is not an effective use of traditional algorithm,put forward a kind of no need to consider the parameters of semi supervised local sparse embedding(SELSE)algorithm for feature extraction of hyperspectral image.Algorithm based on sparse representation theory,by solving the optimization problems of sparse coefficient map construction norm,and maximizing the between class information using the tag data limited.Through the experimental analysis calculation concludes the indicators,time and classification precision comparing the proposed algorithm with traditional semi supervised algorithm between: the algorithm obtains better classification accuracy.Finally,considering the semi supervised algorithm need to know in advance the original data set of a priori knowledge,and unsupervised algorithm need not,however,currently based on manifold learning unsupervised algorithm can only separate description of local or global geometric structure,this paper proposes an unsupervised feature extraction algorithm based on global and local manifold structure-(Global and Local Manifold Structure Feature Extraction Algorithm,GLMS).By constructing graph embedding model while keeping the geometric structure of different types,the global and local.Experiments prove that by simultaneously considering the geometric structure of the data between the local and global,manifold learning algorithm is better than other traditional in the index classification performance.