Research Of Dimensionality Reduction For Hyperspectral Images Based On Supervised Learning

In recent years,with the continuous development of remote sensing technolbgy,hyperspectral image,as a multi-channel image containing a large number of ground target details,has been widely used in classification,detection and clustering.However,hyperspectral images also contain a large amount of interference information,resulting in the illusion of"rich information and lack of knowledge”,and the high dimensionality brings many inconveniences to data processing,that is,the so-called"dimension disaster"..Therefore,dimensionality reduction has become an important preprocessing step in hyperspectral image research.At present,dimension reduction methods can be divided into three categories:supervised dimension reduction algorithm,semi-supervised dimension reduction algorithm and unsupervised dimensron reduction algorithm.In the supervised learning method,low rank representation is introduced to overcome the limitation of sparse representation,and the global structure information is captured while the local information of data is acquired,so that the spectral information of hyperspectral images can be used more effectively.In the dimension reduction method based on semi-supervised learning,the unmarked sanples are concerned while the unmarked sarmples are concerned,and the class knots of unlabeled data are used.Construct information to capture local geometric information of the whole data.In this thesis,the following improvements are made in the supervisory method and semi-supervisory method respectively:1)In the aspect of supervised learning based dimensionality reductio n algo rithm,a dimensionality reduction algorithm based on low rank weighted sparse graph is proposed in this thesis.The basic idea of the algorithm is to strengthen the focal importance of some related atoms in the dictionary of sparse representation by weighted sparse representation,and to enphasize the global importance of the related atoms by introducing low rank constraints of coefficient matrix,and to achieve a balance between weighted sparse representation and low rank constraints by penalty parameters.In the implementation of the algorithm,we use weighted 1-norm to achieve weighted sparse representation,and use weighted nuclear norm to achieve low rank constraints of coefficient matrices,while using weights to enhance the importance of vecto rs which corresponding to larger singular values in low rank space.The experimental results on PaviaU and Indian Pines public datasets show that the performance and visual effects of the proposed algorithm outperforms those of the current mainstream supervised dimension reduction algorithms.2)In the aspect of semi-supervised learning based dimensionality reduction algorithm,this thesis proposes a class probability based semi-supervised dimensionality reduction algorithm.Firstly,the labeled data is used to construct the representation matrix of each class samples in low-dimensional embedding space,and then the class information of unlabeled data obtained by the reconstruction error of unlabeled data in various low-dimensional spaces.According to the class information of data,the inter-class and intra-class scatter matrices of labeled data and unlabeled data are constructed respectively,and a more discriminative scatter matrix is formed by proportional parameters.Finally,the dimensionality reduction projection matrix is obtained by generalized eigenvalue decomposition.The experimental results on Salinas and PaviaU public datasets show that the proposed algorithm outperforms some of the latest semi-supervised dimension reduction algorithms in classification effect.