Research On Hyperspectral Image Spatial And Spectral Super-resolution Methods

Hyperspectral imaging technology can measure spectral signature of each pixel in the observed scene.Physical and chemical properties such as material category of objects in the scene can be determined using hyperspectral imaging technology,which is widely applied in target detection,land-cover classification,mineral exploration and so on.For the sake of imaging mechanism and optical equipment,resolution of hyperspectral image is often limited.Due to the diversity of category of land-cover and the complexity of distribution of land-cover,lots of mixed pixels exist in hyperspectral image,and it has negative effect on the following applications.The current hyperspectral image super-resolution methods treat super-resolution and following interpretation as two separate parts.Result of super-resolution is used as input of following interpretation,therefore the error generated in super-resolution can not be suppressed,and the error would be propagated to the interpretation and decrease the accuracy of interpretation.On the other hand,the traditional hyperspectral image super-resolution only focuses on the enhancement of spatial domain,but neglect the enhancement of spectral domain.The enhancement of spectral domain is significant for hyperspectral imaging,not only can reduce the number of spectral bands,thus decrease the volume of data and the burden of data transmitting,but also increase the width of spectral band and spatial quality of each image.However,currently there are few of work in spectral super-resolution enhancement.In this paper,we work on the above two problems in hyperspectral image super-resolution,the work is as follows:In Chapter 2,we analyze the problem of traditional framework of hyperspectral image processing,we also analyze the relationship between super-resolution and unmixing.A feedback framework for super-resolution and unmixing is proposed by introducing closed-loop feedback idea.In this framework,spatial super-resolution and unmixing is combined and optimized jointly,both of them act as constraint for each other.On the one hand,the residual between abundance and ground truth is used as feedback for super-resolution to reduce spectral distortion.On the other hand,spatial information is used as feedback for unmixing to improve its performance.This algorithm is under the framework of sparse representation,high resolution dictionary is trained from panchromatic image,hyperspectral image with high resolution is reconstructed with sparsity as regularization.Spectral regularizer is designed using abundance and endmembers from the result of unmixing,this regularizer connectssuper-resolution and unmixing together and improves them jointly.The experiment results on simulated data and real data demonstrate the effectiveness of the proposed method.By exploiting the unmixing as feedback for super-resolution,spectral distortion of the reconstructed image is less and spatial consistency of the reconstructed image is higher.Both of the super-resolution and unmixing result is competitive to the state-of-the-art methods.In Chapter 3,we focus on super-resolution for spectral domain neglected by the remote sensing community.A super-resolution method for spectral domain is proposed based on dictionary learning.3-D hyperspectral image is transformed into 2-D matrix,dictionary is learned from this matrix.Because both spatial and spectral information is contained in the 2-D matrix,spatial and spectral correlation is contained in the learned dictionary.A spatial fidelity term is introduced as regularizer.The experiment result on Sandiego hyperspectral image demonstrate that the proposed algorithm can enhance the spectral resolution efficiently,image with high spatial consistency can be reconstructed using the proposed method.