Methodological Research Of Hyperspectral Image Superpixel Classification Based On Sparse Representation

The virtual land environment is an important part of the integrated natural environment and plays an important role in virtual testing.Hyperspectral image classification technology is an important technical foundation for constructing a virtual terrestrial environment.With the increase of the spectral dim ension of hyperspectral images(HSI),there are some problems in the field of hyperspectral image classification,such as the high dimensionality of spectral data,insufficient utilization of adjacent spatial information,and severe salt-pepper noise.It is important to study how to solve these problems in the construction of v irtual terrestrial environments.This dissertation uses image segmentation technology and sparse representation classification technology to design hyperspectral image superpixel classification method based on sparse representation.It can solve the problem of rough use of spatial information and severe salt and pepper noise and improve the classification accuracy.For the problem of superpixel segmentation,a hyperspectral superpixel segmentation algorithm based on principal component weighted false color composition and color histogram-driven(FCC-CHD)is designed.In the design,we store hyperspectral information through a principal component weighted false color image.After that,the color histogram driving function was proposed to evaluate the segmentation.Finally,the hill-climbing method is used to solve all the superpixel regions through reasonable problem transformation.And the superpixel segmentation result is obtained through this fast and accurate optimization solution method.For the sparse representation part,we designed semi-supervised K-SVD and multiscale sparse representation(SK-MSR).In this method,we convert the unsupervised K-SVD method to semi-supervised K-SVD dictionary learning methods that emphasize class features.After the fusion of the spectral-spatial information of the rectangular region and the superpixel region,a stable and accurate sparse representation of each pixel is obtained.Afterwards,this dissertation presents a complete HSI superpixel classification method based on sparse representation,this method combines FCC-CHD and SKMSR methods by firstly super-pixel segmentation and post-sparse representation.Then the final classification results are obtained through residual classification and superpixel voting.And through a comprehensive verification experiment of this method,it is verified that: through scientific design,the FCC-CHD and SK-MSR methods of this dissertation can improve the performance of various classification algorithms,and the complete hyperspectral image superpixel classification method has the best and most stable classification performance on different types of ground objects,and can meet the needs of various virtual ground environments.