Saliency And Sparse Representation Learning Based Optical Remote Sensing Image Object Detection And Classification

Optical remote sensing technology has been widely applied and played an important role in various fields that are closely related to the national welfare and the people's livelihood,such as ecological protection,hydrogeology,disaster monitoring and urban planning.With the explosive growth of remote sensing data,it also brings severe challenges to the analysis and interpretation of optical remote sensing data: how to extract the crucial ground objects information effectively in the face of massive multi-source heterogeneous remote sensing data,and thus provide efficient data support for multiple application fields.Given that different types of optical remote sensing data obtained by different sensors not only have different data characteristics,but also present diverse distribution characteristics of the ground objects,the interpretation methods for different types of optical remote sensing data should be different.Aiming at the main problems and challenges in the interpretation of panchromatic remote sensing images(RSIs)and hyperspectral remote sensing images(HSIs),this dissertation studies the region-of-interest(ROI)detection for panchromatic RSIs and hyperspectral image classification based on the theories of saliency analysis and sparse representation learning,and also proposes effective ROI detection and classification methods.The innovative research content of the dissertation mainly includes:1.Considering the difficulty of extracting high-quality regions-of-interest from the panchromatic RSIs for the lack of color information,we propose an effective approach for regions-of-interest detection based on the statistical distinctiveness.In order to compensate for the absence of color information,we take advantage of the statistically meaningful representations for panchromatic RSIs to capture the significant information about the inherent structures and unusual characteristics that are distinguished from the natural images.Then we identify the regions that are highly distinctive from the rest of the scene in terms of statistical characteristics as salient regions.For the existing saliency analysis models,the most widely used center bias and boundary prior are not applicable to panchromatic RSIs,while the proposed method incorporates the local saliency analysis based on the lower-order statistical distinctiveness and the global saliency analysis based on the higher-order statistical distinctiveness,thus highlighting the regions-of-interest that are highly distinctive from the rest of the scene without relying on priors.Experimental results demonstrate that the proposed method can not only detect multiple complete ground objects of interest in RSIs,but also detect regions-of-interest with clear boundaries,which can consistently and uniformly highlight the whole regions of interest and suppress the interference such as background clutters and noise effectively,thus producing high-quality detection results.2.In view of the phenomena of “the same spectral signatures with different materials” and “different spectral signatures with the same material” in hyperspectral image classification and the complex nonlinear data structure of hyperspectral images,we propose a kernel fused representation-based classification(KFRC)method via a spatial-spectral composite kernel with ideal regularization(CKIR).By introducing the ideal regularization strategy into the spatial-spectral composite kernel learning framework,the proposed method can take full advantage of the spectral information,the spatial information and the label information for hyperspectral image classification,thus significantly improving the classification accuracies and avoiding the salt and pepper appearance of the resulting classification maps.In addition,by embedding the optimized CKIR into the kernel sparse representation-based classifier and the kernel collaborative representation-based classifier,the classification accuracies of the nonlinear representation-based classifiers can be significantly improved.Meanwhile,the linearly inseparable problem of data samples in the low-dimensional space caused by the complex nonlinear data structure of the hyperspectral images can be effectively solved through the kernel technology.Considering that the fused residual has stronger discriminant ability than the individual residual produced by sparse representation(SR)or collaborative representation(CR),the CKIR-based KFRC method is proposed to further enhance the discrimination ability of the proposed classifier.Experimental results on real hyperspectral data sets confirm that the proposed method can significantly improve the class separability and outperform the other state-of-the-art classifiers.3.Aiming at the scarce availability of the labeled data samples in hyperspectral image classification,a multifeature correlation adaptive representation-based classification method is developed.In the case of the very limited available training samples for hyperspectral image classification,it is difficult to reconstruct the test sample accurately by using sparse representation learning methods.Therefore,the proposed method introduces a data correlation adaptive penalty into the representation model to make the model balance between SR and CR adaptively according to the precise correlation structure of the dictionary,thus constructing a more appropriate representation model for the test sample.Furthermore,the proposed model has the ability of performing sample selection and grouping correlated samples together simultaneously,which overcomes the intrinsic limitations of the traditional SR and CR models.In order to improve the classification accuracy under the small size samples situations,the proposed method designs a multi-task representation mechanism to integrate the discriminative capabilities of complementary features.Considering that the relationship between the test sample and the training samples is also discriminative for representation,we further introduce the locality constraint to encode the local structure information of the data samples and the correlation between the test sample and the dictionary atoms into the model,thus building a more accurate representation model for the test sample.Experimental results demonstrate that the proposed method can still achieve favorable classification results in the case of the very limited available training samples,which reflects the superiority of the proposed method for hyperspectral image classification under the small size samples situations.