Research On Dimensionality Reduction Of Hyperspectral Remote Sensing Image Data Based On Block Non-negative Sparsity Reconstruction Embedding

While hyperspectral remote sensing imagery brings rich information in the era of big data,there are also outstanding problems such as excessively high band dimension and huge amount of data.As one of the key steps in hyperspectral remote sensing image processing,dimensionality reduction is to make full use of A method to solve the problem of too high dimensionality of the feature space under the premise of using its massive data.The paper improves the reconstruction error iterative algorithm of block nonnegative sparse reconstruction embedding.Taking the Pavia University data set as an example,the algorithm and the ant colony optimization algorithm,principal component analysis,nuclear principal component analysis,non-Negative sparseness keeps embedding five dimensionality reduction methods for experiments.The main research contents are as follows:(1)Explain in detail the principle and method of dimensionality reduction of hyperspectral remote sensing data.The ant colony optimization algorithm(ACO),principal component analysis(PCA),kernel principal component analysis(KPCA),and non-negative sparse preserving embedding(NSPE)mathematical models of four algorithms for dimensionality reduction of hyperspectral remote sensing images are combed and deduced and passed Python Programming realization.(2)The block non-negative sparse reconstruction and embedding(BNSRE)algorithm achieves dimensionality reduction of hyperspectral remote sensing image data through two parts: block non-negative sparse representation and low-dimensional embedding projection.In the part of block non-negative sparse representation,the block regularized orthogonal matching pursuit algorithm(BROMP)is used to replace the original block orthogonal matching pursuit algorithm to improve the iterative algorithm for solving local optimal reconstruction errors;low-dimensional embedding projection In the part,maximize the non-local non-negative sparse reconstruction weight matrix and minimize the local non-negative sparse reconstruction weight matrix,so that the sample data with low similarity is far away and the sample data with high similarity is close to obtain the global optimum Low-dimensional subspace hyperspectral remote sensing image data.The BNSRE dimensionality reduction algorithm is implemented through Python programming,and the algorithm is verified using the Pavia University data set.The results show that the block non-negative sparse reconstruction embedding algorithm(BNSRE)based on the block regularization orthogonal matching pursuit algorithm(BROMP)is a feasible unsupervised dimensionality reduction method.(3)Use the Pavia University data set to perform dimensionality reduction experiments,and verify the dimensionality reduction effect based on the k-nearest neighbor method(KNN)classifier.The experimental results show that when the number of sample dictionary blocks is equal to the number of classification categories is 9,the result of the block non-negative sparse reconstruction embedding algorithm reaches the optimal solution under this condition,the overall classification accuracy reaches 85.93%,and the Kappa coefficient also reaches Above the high consistency standard,its dimensionality reduction time is only 1.09 s slower than the fastest dimensionality reduction time,which is within the acceptable range of time cost;and when the number of blocks is 9,there are only a few odds outside the diagonal of the sparse weight matrix This is a bright spot,which shows that under this condition,the reduced dimensionality data can be guaranteed to contain more valuable information.The results of five dimensionality reduction algorithms,ACO,PCA,KPCA,NSPE,and BNSRE,are compared with the direct classification of hyperspectral data sets.The results show that when the output of the same dimensionality is 45 dimensions,the dimensionality reduction time of the BNSRE algorithm is 15.60 s,and the overall classification The accuracy is 86.68%,and the Kappa coefficient is 0.8196,which are better than other comparison algorithms.In addition,BNSRE performs better than NSPE in lowdimensional subspaces of different dimensions;the Pavia University data set is directly classified by the KNN classifier,and its overall classification The accuracy reaches88.72%,which means that after BNSRE dimensionality reduction,the hyperspectral data is reduced from a 103-dimensional feature space to a 45-dimensional feature space,and only 2.04% of the information is lost.This is a better result.