Research On Hyperspectral Image Classification Algorithm Based On Spectral-Spatial Information And Neural Network

In the field of hyperspectral image classification,spectral information extraction is often used as an important source and basis for image feature classification and recognition.Due to the complex relief of the measured ground and the limited sensitivity of remote sensor,the interference of light wave phase difference will be introduced,which makes it difficult to accurately restore the real value in the spectral information,resulting in the loss of natural image details.Therefore,the traditional hyperspectral image classification technique adopts multi-dimensional and multi-parameter compensation to solve the direction problem of data vector macroscopically,but the calculation system is complex and the target pixel classification is easy to be confused.It is difficult to analyze the pixel displacement characteristics from a comprehensive three-dimensional perspective and affect the fine classification of image categories only based on spectral information.Based on spectral space information and neural network technology,this thesis studies the classification of hyperspectral images from the perspective of spatial features and proposes two algorithms:1.Hyperspectral image classification algorithm based on spatial spectrum super-pixel kernel limit learning machine.The algorithm uses the spatial spectrum information to extract the pixel feature information.Then,the geometric property relationship of the local edge pixels in the spatial position of the image is measured.The proposed method further considers the heterogeneous and homospectral status and difference density of the pixels in the adjacent area,which can perform super-pixel segmentation on the spatial structure component of the image.The effectively segmented super-pixel is regarded as a specification adaptive region.The hyperspectral images are classified by using the advantages of linear separability of kernel strategy in high-dimensional hyperplane data and less restrictions of limit learning machine optimization algorithm.2.Hyperspectral image classification algorithm is proposed based on PCA and 2D-CNN fusion under spatial spectrum information.As an important part of deep learning,neural network can re adjust the interconnection relationship between pixels through mathematical operation model without adding additional pixel classifier.At the same time,principal component analysis（PCA）algorithm based on spatial spectrum information can solve the problem of multiple collinearity of data in extracting spectral information,avoid the phenomenon of data aggregation caused by multidimensional sparsity by dimensionality reduction,and reconstruct the information features by convolution.It can not only process the information presented in the space where the spectrum is located,but also add self-learning ability and adaptive function.Through a series of experiments,it is proved that the proposed algorithm can not only improve the accuracy of the classification results,but also leave sufficient exploration space for further research.In addition,as the mainstream algorithm,neural network can be mined and used in this experiment,which lays a solid foundation for further experiments.