Researches On Pan-Sharpening With Sparse Representation And Kernel Learning

With the development of the astronautical technology,the number of remote sensing satellites are increasing,meanwhile the spatial resolution,temporal resolution and the diversity of remote sensing data are improved significantly.The earth observation remote sensing satellite system containing resource,meteorology,ocean,environment and national defense has been established in our country,which can provide information service and decision support for modern agriculture,disaster reduction,resource,public security and other important areas.At the same time,the type of remote sensing image sensors becomes ever more diverse.Besides,the multi-satellite network mode will produce a large amount of multi-sensor and multi-temporal remote sensing data.In order to effectively utilize the massive amounts of data,all types of image fusion methods are proposed.Multi-source remote sensing data fusion can integrate the remote sensing data from different sources by signal analysis and image processing techniques to improve the precision and reliability of information and enhance the spatial and spectral characteristics.Therefore,the data fusion technique plays an important role in remote sensing application.Due to the fact that multispectral images contain rich spectral information,which is beneficial to recognition and interpretation of objects,but the spatial information of multispectral images is poor.Inversely,panchromatic image performs well in spatial domain,which can provide enough information about spatial details,but the spectral information is limited.Pansharpening aims at producing a fused image which has both high spectral and spatial resolution and can offer more comprehensive information for subsequent identification and classification,and it has become an important branch of data fusion filed.Sparse analysis technology,as a new image representation technique,can give the more concise representations of the signal,based on which sparse representation of complex remote sensing data can be obtained by solving a sparse optimization problem.In this thesis,the fusion methods of multispectral and panchromatic images are studied and a series of pansharpening methods based on sparse representation and kernel learning are proposed to enhance the spatial resolution of fused image and preserve spectral information.The main contributions of this thesis are listed as follows:1.A pan-sharpening method based on sparse non-negative matrix factorization is proposed.Non-negative matrix factorization is a very efficient method for decomposing multivariate data into strictly positive activations and basis vectors.And the non-negative basis vectors can represent the local features of the data.In order to reduce the spectral and spatial distortion,a novel method based on sparse non-negative matrix factorization is proposed for the fusion of multispectral and panchromatic images.Firstly,the high spatial resolution and low spatial resolution dictionaries are learned from panchromatic images.Then we construct a sparse non-negative matrix factorization model of the multispectral image.Thus,the coefficient matrix with spectral information can be obtained.The high-resolution multispectral image is produced by the multiplying high-resolution dictionary and the coefficient matrix.By introducing the sparse regularization,the instability of the standard non-negative matrix factorization is conquered and the fused image can preserve the high spectral and spatial information.Some experiments are made on QuickBird and GeoEye satellite datasets,and experimental results show that our proposed method can reduce distortions in both the spectral and spatial domains,and outperform some related pan-sharpening approaches in visual results and numerical guidelines.2.A pan-sharpening method based on SVT and Semi-NMF is proposed.Semi-NMF extends the range of application of non-negative matrix factorization,and it relaxes the constraints of the non-negative matrix factorization.Semi-NMF do not require the data matrix to be non-negative,but the coefficient matrix is restricted to be non-negative.When an image is transformed by SVT,we can obtain the component which contain positive or negative value.For low-frequency components,the rule of fusion is Semi-NMF method and the only characteristic base is the fused low-frequency component.For support values,the fusion rules are region energy rules.Some experiments are made on QuickBird and GeoEye satellite datasets.And the experimental results show that our proposed method can reduce distortions in both the spectral and spatial domains.3.A novel method based on deep support value learning networks is proposed for fusion of remote sensing images.In order to avoid the loss of information,we abandon the downsampling of feature mapping layer of traditional convolution neural network.The deep support value learning networks(DSVL Nets)contains five hidden layers,where each layer consists of convolution layer and linear layer.All convolution layers and the fifth linear layer are regarded as the outputs of DSVL Nets.The convolution layers images are fused by abs-maximum model.The linear layer images are sparsely represented on overcomplete dictionary,and then the coefficients are fused by abs-maximum model.The fused convolution layers images and the linear layer image are reconstructed,and one can obtain the fused result image.Some experiments are taken on several QuickBird and GeoEye satellite datasets.Compared with PCA,AWLP,PN-TSSC and SVT,the experimental results show that the proposed method outperforms some related pan-sharpening approaches in both visual results and numerical guidelines,and reduces the distortion in both the spectral and spatial domain.4.A refined pan-sharpening method with geometric multiscale analysis and hierarchical sparse auto-encoder is proposed.First,a Geometric Multiscale Analysis(GMA)tool,NonSubsampled Contourlet Transform(NSCT),is used to capture directional details of panchromatic image at multiple scales.Then at each scale,hierarchical sparse auto-encoder(HSAE)is developed to gradually filter out the refined spatial details,via sparsely coding under spatial self-dictionaries.The refined details are then injected into multispectral images to alleviate spectral distortion.By exploring the spatial structure in images and refining the spatial details injection via HSAE,our proposed method can reduce distortions to present fidelity colors and sharp appearance.Some experiments are taken on several datasets collected by QuickBird,GeoEye and IKONOS satellites,and the experimental results show that our proposed method can reduce distortion in both the spectral and spatial domains,and outperform among some related methods in terms of both visual results and numerical guidelines.5.A pan-sharpening method based on the storage of overcomplete dictionary by multi-direction tree is proposed.Remote sensing images have a variety of terrain types and complex spatial structure information.In order to discover these information,each image is divided into smooth patches,erratic patches and directive patches.For smooth patches and erratic patches,the overcomplete DCT dictionary and KSVD dictionary are applied,respectively.While directive patches are sparsely represented under multi-direction tree-ridgelet dictionaries.The KSVD dictionary well represent the erratic patches,and the directive patches have good approximation performance on the tree-ridgelet dictionaries.Some experiments are taken on several datasets collected by QuickBird,GeoEye and IKONOS satellites,and the experimental results show that proposed method can reduce distortions in both the spectral and spatial domains.6.A pan-sharpening method using geometric steerable kernel learning is proposed.According to the observation model,the relationships among low-resolution multispectral,panchromatic and high spatial resolution multispectral images are formulated.The structure similarity prior within a local region in the panchromatic image is employed to regularize the solution space to obtain a more accurate solution.Then,the prior is embedded into the low-resolution multispectral image to enhance the spatial resolution.In order to capture better geometrical structure information,such as orientation information and geometric textures,the steerable kernel is used to calculate the similarity coefficients in a local window.Some experiments are considered on QuickBird and IKONOS datasets and the results show that the proposed method can improve the visual effect and the quantitative values.