The Study On Hyperspectral Remote Sensing Image Fusion Based On Wavelet Transform

Hyperspectral remote sensing image fusion is one of the key technologies of hyperspectral remote sensing image processing,whose purpose is to get an image with high clarity and recognizability,by removing redundancy information and merging complementary information.The fused image can be better applied to ground mapping,geological exploration and agricultural survey,etc.As hyperspectral images have the characteristics of mass data and high redundancy,it is difficult to analyze and process hyperspectral images.On the basis of studying the selection method of hyperspectral band,A band selection method based on band index and spectral angle separability is proposed in this thesis,whose aim is to select the best subset of bands to represent the whole hyperspectral data.Firstly,dividing the hyperspectral data space into several subspaces according to interband correlation.Then,the band index is used to select the high quality bands in each subspace.The spectral angle matching(SAM)method is used to select the band combination with the better separability between ground object categories at last.The experimental results show that the band combination performance index selected by this method is superior to other common methods.When using spatial domain fusion methods to fuse hyperspectral multi-band images,the fused images have the problems of low definition and inconspicuous details.Based on the research of image fusion method which is based on wavelet transform,in this thesis,the fusion rules of wavelet transform are improved as follows:(1)The traditional Spatial Frequency(SF)method only considers the vertical and horizontal directions,and can not adequately describe the peripheral pixel information of the central pixel.The SF is improved according to the gradient values in four directions in this thesis.The spatial frequency of each central pixel in the low frequency component image is calculated by block method,then the SF image is obtained.Due to the clear blocks in the image correspond to the larger pixel values of the SF image,therefore,the rule of choosing larger SF is used to fuse the low frequency components.(2)The central pixel with larger energy in the region represents the details of the image.The fusion rule of choosing bigger regional energy directly calculates theregional energy of the central pixel according to values of each pixel in the region.While the regional characteristics of neighborhood pixels are not taken into account.In this thesis,convolution weighted method is used to calculate the region energy of the central pixel.Then,calculating the energy matching degree of the corresponding region.Finally,the fused high frequency wavelet coefficients are determined by matching threshold and region energy.The experimental results show that,compared with the weighted average method,the principal component analysis method and other fusion rules wavelet transform method,the objective assessments,such as information entropy,average gradient and spatial frequency of the proposed method are better.