Sparse Matrix Factorization Based Remote Sensing Image Fusion

At present,remote sensing satellites for earth observation have been intensively launched in our country.Various high resolution remote sensing images are being widely used on many major strategies,such as modern agriculture,disaster prevention and reduction,survey of resource,environmental protection and national security.However,due to the limitation between the hardware technology of satellite sensor and the cost of launching,there exists an inherent tradeoff among spatial resolution,time resolution,spectral resolution and scanning width.Therefore,it is difficult for remote sensing satellites to simultaneously obtain high spatial resolution and high spectral resolution images.For example,panchromatic image only contains one band,but its spatial resolution is high.The spatial resolution of multispectral image is low,but it is made up of several bands.Hyperspectral image with lower spatial resolution consists of hundreds of bands,which can provide abundant spectral signatures.Therefore,in order to obtain high resolution remote sensing images with multiple attribute,researchers adopt image fusion technique to take full advantage of multisource remote sensing images to achieve more accurate interpretation for the observed scenes.Remote sensing images can be regarded as high dimension matrix and multisource remote sensing image fusion can be formulated as a matrix recovery problem.For the problem of spectral distortion and spatial information loss in existing methods,this dissertation exploits the sparsity in spatial and spectral domains to explore the remote sensing image fusion technology based sparse matrix factorization.The fusion of panchromatic and multispectral images and the fusion of multispectral and hyperspectral images are thoroughly researched by non-negative factorization,convolution sparse coding,low-rank factorization.The works in this dissertation are summarized as follows:(1)By exploring the sparsity in spatial and spectral degradation model of high resolution images and combining with the non-negativity of pixel values in remote sensing images,a multispectral and panchromatic image fusion method based on coupled sparse non-negative matrix factorization is proposed.Firstly,non-negative factorization is implemented on low spatial resolution multispectral and panchromatic images and a joint sparse representation of multispectral and panchromatic images is developed by the observation model of source images and fused image.Then,the model is optimized by alternate iteration algorithm.Finally,the fused image can be produced by multiplying the high resolution dictionary from panchromatic image with the coefficient matrix from multispectral image.The results on Quick Bird and Geoeye-1 satellite data show that the proposed method can efficiently preserve the spectral information in fused images.The spatial structures in these images can be revealed more efficiently by the learned high resolution dictionary.(2)By introducing the structural property in bands of multispectral image,a multispectral and panchromatic image fusion method based on convolution structure sparse coding is proposed.In this method,low spatial resolution multispectral and panchromatic images are regarded as the results of the fused image after spatial and spectral degradation,respectively.Then,the proposed method combines convolution sparse coding model with the degradation relationship.In order to utilize the correlation in bands of multispectral image,structure sparse constraint is employed to establish convolution structure sparse coding,which can capture the correlation prior.Finally,the algorithm based on alternating direction method of multipliers is derived to estimate the feature maps to construct the fused image.Besides,a high/low resolution filter joint learning framework based on block coordinate descent is described to ensure correspondence and consistency.The performance of the proposed method is verified on Quick Bird and Geoeye-1 satellite data and the results imply that the method can improve the spatial and spectral quality of fused images.(3)In order to produce more accurate spatial and spectral injection components,a multispectral and panchromatic image fusion method based on low-rank matrix factorization and spatial-spectral offsets is proposed.In this method,low-rank matrix factorization is employed to model the relationship between high/low spatial resolution multispectral images.Then,two offsets are defined.One offset represents the spatial differences between high/low spatial resolution multispectral images.The other is the spectral differences among them.In order to reduce the spatial and spectral distortions in the fused image,spatial equalization and spectral proportion regularizations are designed to learn the spatial and spectral offsets from low spatial resolution multispectral and panchromatic images.Then,the low-rank fusion framework is established after combining the spectral degradation of panchromatic image.Finally,the fused image can be obtained by augmented Lagrange multiplier algorithm.The performance of the proposed method is also verified on Quick Bird and Geoeye-1 satellite data and the method can reduce the spatial and spectral distortions in fused images.(4)By exploiting the group correlation in the bands of hyperspectral images,a multispectral and hyperspectral image fusion method based on low-rank factorization and group spectral embedding is proposed.This method combines the spectral degradation relationship with the group sparse prior in difference image to build the low-rank fusion model.Besides,in order to make full use of the manifold structure in low spatial resolution hyperspectral image,group spectral embedding regularizer is constructed.The geometric relationship and similarity in bands of low spatial resolution hyperspectral image can be shared with high spatial resolution hyperspectral image by the regularizer,which ensures the similarity between adjacent bands of hyperspectral image.The results on two different kinds of hyperspectral data show that the fusion results of the method can produce more consistent spectral curve with the reference image.(5)For the whole preservation of spatial and spectral structures in hyperspectral images,a multispectral and hyperspectral image fusion method based on spatial and spectral graph regularized low rank tensor decomposition is proposed.This method adopts low rank tensor factorization to model the source images and fused image.In order to further improve the quality of the fused image,spatial graph and spectral graph are constructed from multispectral and low spatial resolution hyperspectral image,respectively.The spatial graph and spectral graph can ensure that the prior information in source images can be further inherited into fused image.Due to the consideration of spatial consistency and spectral structure,the proposed method can produce satisfactory fused results.The performance is verified on two kinds of hyperspectral data and the proposed method can more effectively preserve the whole structure of hyperspectral images.