Remote Sensing Image Change Detection And Classification Using Spatial Information And Statistic Learning

In such an information explosion era,data has penetrated into every industry and field.Inevitably,the processing and application to these massive amounts of data is a necessary demand of people.And it becomes a great power to promote the growth of productivity and determine future of various areas.With the development of science and technology and information flow,data obtained from the reality have the characteristics of large scale and uncertain.The statistical learning methods are often the most powerful to deal such data.At the same time,to be intelligent is the inevitable trend of the development of computer.,The statistical learning is a kind of effective means to imitate human intelligence,although it has some limitations.This thesis mainly is aiming at the in-depth study of using statistical learning method to deal with remote sensing image change detection and classification problem.For the specific problems of application,evidence theory,spatial information,complex structure information and entropy information are incorporated into the statistical learning models.In summary,the major contributions are outlined as follows:(1)The Yellow River Estuary area of China is under great pressure from both human intervention and natural processes.For the analysis of the changes for this area,it presents a novel change detection based on local fit-search model and kernel-induced graph cuts in multitemporal synthetic aperture radar images.Change detection involves assigning a label to every pixel.This task is naturally formulated in terms of energy minimization,which can be effectively solved by graph cuts.The difference image is transformed implicitly by a kernel function,so that an alternative to complex modeling of the original data makes the piecewise constant model become applicable for graph cuts formulation.An issue is that graph cuts algorithm is sensitive to the initial estimate.The local fit-search model is proposed to approximate to the local histogram while selecting an optimal threshold for the initial labeling,which leads to an effective constraint for graph cuts and computational benefits as well.Visual and quantitative analyses obtained on the Yellow River Estuary data set confirm the effectiveness of the proposed method.(2)One of the main problems related to synthetic aperture radar image change detection lies in the existence of speckle.The reason is that it increases the overlap between changed and unchanged pixels in the difference image that makes the achievement of the optimal discrimination become more ditticult.In order to address the issue,we developed an unsupervised change-detection approach for multitemporal SAR image,which specifies the prior knowledge about the spatial characteristics of the classes through Dempster-Shafer fusion strategy and embeds it in the Expectation-Maximization iteration process.It considers that each pixel in the difference image is unique due to its neighborhood,even if they have the same pixel value.Then,we introduce the Dempster-Shafer fusion strategy to construct the prior model for each pixel,based on the assumption that the local and global priors are independent features.Note that a modified local prior model is developed to describe the mutual influences of neighboring sites.As consequence,within the framework of Bayes theory,the Expectation-Maximization algorithm allows one to estimate the statistical parameters associated with both the changed and unchanged classes with the merged prior model.Visual and quantitative results obtained on real multitemporal SAR image data sets confirm the effectiveness of the proposed method,comparing with state-of-the-art ones for change detection.(3)Nonlocal means algorithm has been proven to be an effective context-sensitive denoising approach,where many similar patches spatially far from a given patch could provide nonlocal constraint to the local structure.For hyperspectral image,however,the conventional nonlocal means algorithm becomes inapplicable for the high number of spectral bands.Therefore,we incorporate the image nonlocal self-similarity into the maximum a posteriori estimation for hyperspectral classification.The main novelty lies in the following two aspects:1)the nonlocal means algorithm is exploited to combine similar local structures and nonlocal averaging;2)a new class-relativity measurement is proposed to describe the self-similarity in the context of the hyperspectral classification.Several experiments on simulated and real hyperspectral data sets are provided to demonstrate the effectiveness of the proposed algorithm.(4)It presents an effective classification approach for hyperspectral image,based upon a novel discontinuity adaptive class-relative nonlocal means algorithm and embedding it in the global energy function by energy fusion strategy.Inspired from recent works related to nonlocal means,we extend this framework to label space,assuming that nonlocal similar patches have similar label structures.Thus,similar local structures and nonlocal averaging process are combined by the proposed discontinuity adaptive class-relative nonlocal means algorithm.The Shannon entropy is adopted to define the distribution of energy.The energy function is then improved by fusion strategy that selects the energy corresponding to the lowest uncertainty.As a sequence,the hyperspectral image classification task stated in term of energy minimization is efficiently solved by graph cuts algorithm.Experiments on two real hyperspectral data sets are provided to demonstrate the effectiveness of our hyperspectral classification algorithm.(5)Bayesian model with spatial prior information for hyperspectral image classification is designed.The use of spatial information can often be more accurate interpretation of the scene,and can greatly improve the performance of image classification algorithm.Using the traditional Bayesian model for image classification problem,the related prior probability is usually the global one.There are a large number of patches with similar structure distributed in an image,especially in nearest neighbor position.Thus a new algorithm is proposed to improve the traditional Bayesian model with the characteristic of local structure similarity.That is,the Bayesian model is modified with the local prior information through the evidence fusion strategy.The prior probability of each pixel in the image is fused with its neighboring who provides the spatial prior information.Thus the prior probability of each pixel can be described more accurately.The optimal classification result is obtained by maximizing the posterior probability.