Remote Sensing Image Classification And Change Detection Based On Structured Representation Learning

The measurement of similarity is the basis of classification and clustering,which has been widely researched in the fields of machine learning and pattern recognition.The result of similarity measurement determines the performances of algorithms to a large extent,therefore,how to effectively and exactly measure the similarity among samples is critical to the effectiveness of algorithms.In this thesis,we start from the basic problem of the similarity measurement of samples,incorporate the prior information of structural characteristic of samples,combine the spatial information of neighboring pixels in images,to construct several models and apply them to the real applications of hyperspectral image classification and SAR image change detection,in which the problems of incomplete capture of data structures in hyperspectral images and the great effect of speckle noise on the result of SAR image change detection are mainly solved.Specially,the main works of this thesis can be summarized as the following five aspects:?1?A novel method which combines local collaborative representation with adaptive dictionary selection is proposed for hyperspectral image classification.For the degraded representation of redundant and irrelative pixels when all the labeled samples are used as a dictionary in collaborative representation based classification,we propose to select the K most similar pixels to each test pixel from the dictionary to construct a new one for classification,to solve the problem of inexact representation of relationship among the test pixel and atoms in dictionary.By considering the local consistency of hyperspectral images,we first average the neighboring pixels to incorporate the spatial information before the selection of atoms.The experimental results show that the proposed method outperform support vector machines and some spectral-spatial representation based classification methods.?2?We propose a novel locality-constrained collaborative subspace clustering method for band selection of hyperspectral images.It captures the structure among spectral bands from global and local perspectives,simultaneously,under a framework of subspace clustering.Each band can be seemed as a vertex in a graph,where the similarities among bands are evaluated by the representation coefficients.Two main contributions of the proposed method can be concluded as follows:1)A collaborative subspace representation method is used to capture the global structure among all the spectral bands;and 2)A locality based graph formulation is adopted to preserve the similarity consistency of local spectral bands between the spectral space and representation space.The closest bands to the cluster centers are finally chosen as the selected bands.The effectiveness of the proposed method is demonstrated via the classification results of hyperspectral images.?3?We propose a local constraint low-rank representation method for graph construction,which is further combined with semi-supervised learning for face and digit image classification.The proposed method is derived from the original low-rank representation by incorporating the local information of data.Rank constraint has the capacity to capture the global structure of data.Therefore,the proposed method is able to capture both the global structure by low-rank representation and the local structure by the locally constrained regularization term,simultaneously.The regularization term is induced by the locality assumption that similar samples have large similarity coefficients.The measurement of similarity among all samples is obtained by low-rank representation.Considering the non-negativity restriction of the coefficients in physical interpretation,the regularization term can be written as a weighted?₁-norm.Then a semi-supervised learning framework based on local and global consistency is used for the classification task.Experimental results show that the proposed method provides better representation of data structure and achieves higher classification accuracy in comparison with the state-of-the-art graphs on three real face and one digit databases.?4?We propose a novel unsupervised saliency and kmeans clustering based SAR image change detection method.Salient areas of an image always are discriminative and different from other areas,which make them can be easily noticed.The strong visual contrast of local areas makes saliency suitable to guide the change detection of SAR images,where exist a difference between the two images.Firstly we use the saliency to extract the interest regions from the initial difference image,and then by using a simple thresholding method to extract the most informative regions and neglect and noninterest regions.With the extracted saliency map,we use it to extract the corresponding regions from the two initial SAR iamge,and use PCA to extract the features of local patches to incorporate the contextual information.Finally,kmeans clustering is used to get the final change map.We demonstrate the effectiveness of the proposed method on five real and two simulated SAR image data sets.?5?We propose a SAR image change detection method based on log-ratio and non-local low-rank matrix decomposition.Log-ratio has been widely used in generating the difference image of SAR images.The multiplicative noise existed in SAR images can be transformed into an additive noise by using the log-ratio,for which satisfies the mechanism of SAR image.However,the pixel-wise operator generates many isolated pixels in the difference image which may greatly degrade the performance of change detection algorithms.By stacking the most similar patches from a large region to construct a nonlocal patch matrix which has a clear low-rank structure,we propose to decompose it into three parts:low-rank component,sparse isolated pixels and small level and redundant pixels,in which the low-rank component corresponds to a clean difference image and the others to speckle noise reduced isolated pixels.Finally,kmeans clustering and PCA are used to obtain the final change map.The proposed method can well resolve the isolated pixels exist in change map which are caused by the speckle noises.