Hyperspectral Image Processing Based On Determinantal Point Processes And Singular Spectrum Analysis

Hyperspectral image is a three-dimensional image,contains rich spatial information and spectral information,can reflect the features of surface objects in more detail,and is widely used in agricultural detection,marine ecological environment monitoring and military detection and other fields.There are still some challenges in current hyperspectral image processing.High-dimensional data will bring data redundancy,greatly increasing the difficulty of data processing.Some bands will carry noise during imaging,which will seriously affect the classification accuracy of hyperspectral images.The noise of hyperspectral equipment and different atmospheric scattering conditions have led to the current widespread occurrence of noise,which seriously affect the classification accuracy.In addition,too little label data also leads to insufficient training data,which affects the accuracy of classification.In order to solve the above problems,this paper mainly studies the two main topics of hyperspectral including image band selection and hyperspectral classification,and proposes new methods.The main contents of the research are summarized as follows:(1)This thesis proposes an unsupervised hyperspectral image band selection algorithm based on maximum information entropy-minimum noise and improved determinant point process(MIMN-DPP).In order to protect the original information of hyperspectral images,a band selection criterion is designed in this thesis to select bands with high information entropy and low noise.However,on the basis of this selection criterion,finding a low-redundancy band subset is an NP-hard problem.In order to solve this problem,this thesis considers the correlation between the bands from the original hyperspectral band information and its neighborhood information,and constructs a double-graph structure to describe the relationship between them.Secondly,based on the algorithm of determinant point process,this thesis proposes an improved determinant point process algorithm as a search method for selecting a low-redundant band subset from a double-graph structure.After a lot of experiments,it is shown that the performance of the band selection algorithm proposed in this thesis is greatly improved than the current existing algorithms.(2)This thesis proposes a hyperspectral image classification algorithm based on multi-scale singular spectrum signal component adaptive signal fusion network.The main contribution points of the method are as follows: First,the two-dimensional Singular Spectrum Analysis is used to effectively remove the noise of the hyperspectral image,enhance the characteristics of the data,and effectively reduce the number of training samples.Secondly,the two-dimensional Singular Spectrum Analysis applied to each band of the hyperspectral image can enhance the spatial correlation of pixels,reduce the pixel difference of local image blocks,and make it more uniform.It can effectively reduce the difference of pixels in the field of the convolutional neural network.The impact of classification accuracy.Using neural network to adaptively fuse the singular spectrum components,the weights of different components can be assigned,and the potential information of the components can be fully utilized.At the same time,multi-scale singular spectral features and convolutional frames of different receptive fields can extract feature information in different spatial ranges,making it more generalized on data sets with different characteristics.Experiments show that the proposed method achieves better classification results than existing methods.