Research On Feature Extraction And Selection Method For Hyperspectral Image Classification

Hyperspectral image characterizes hundreds of spectral features of different land-over objects in a narrow spectral range,reflecting the chemical,physical,and optical properties of various objects,and has been widely used in precision agriculture,mineral exploration,and environmental monitoring.However,the inherent imaging mechanism and complex surface environment pose serious challenges to hyperspectral image analysis: the pollution of different types of noise;the same objects with dissimilar spectra and the different objects with similar spectra;the contradiction between the lack of labeled samples and the redundant high-dimensional spectra.In addition,with the development of data acquisition technology,how to process a large number of dynamic hyperspectral data quickly and efficiently is becoming a serious challenge.Based on the analysis of sparse learning,manifold structure embedding,and loss metrics,this dissertation studies the future extraction and feature selection techniques used to label hyperspectral pixels by utilizing the intrinsic properties of hyperspectral images.The main work of this dissertation is as follows.Aiming at the phenomenon of the same objects with dissimilar spectra and the different objects with similar spectra in hyperspectral image classification,this dissertation proposes a sparse coding method that combines spectral,spatial,and label information.The method encodes the representations of high-dimensional pixels in a low-dimensional space in a distributed manner by partitioning the space.The use of the hypergraph that combines spectral and label information makes the structural distribution of low-dimensional features of pixels with the same label more compact,thereby increasing the discriminability between different labels of pixels and improving the discriminability and robustness of new features.On the real hyperspectral datasets,comparison experiments with other widely used feature extraction methods show that the method proposed in this dissertation effectively improves the classification results of different land-over objects.The application of hyperspectral images is limited because feature extraction destroys the physical meaning of the spectra.To solve this problem,based on the combination of spatial information,the dissertation presents a feature selection method based on matrix-based margin-maximization.Considering that adjacent pixels in space usually represent the same object,this method takes pixel blocks from different local spaces as input.Furthermore,the hinge loss function and the row sparse norm of the matrix are used to measure the importance of different spectral features,so that the pixel blocks with different labels have as large a distance between classes as possible in the label space.Besides,in this method,the graph is used to dynamically characterize the discrimination of different spectra,and to make the highly correlated spectra have similar discriminability,thereby enhancing the robustness to noise.Finally,a one-versus-all strategy is adopted to select a feature subset that is optimal for each class in the current context.The classification results on different hyperspectral datasets show that the proposed method can balance the difference and discriminability of selected features and effectively improve the classification performance.In hyperspectral image classification,class-related feature subsets based on one-versusall strategy are prone to ambiguity problem in classification.Moreover,current supervised feature selection methods lack the study of the conflict between insufficient labeled samples and high-dimensional spectra.Therefore,the dissertation proposes a feature selection method based on random spectral subspace.The method first randomly divides the spectral set into multiple subspaces to achieve the purpose of reducing the feature dimension and indirectly increasing the number of labeled samples.Then,the hypergraph that automatically adjusts the weights of hyperedges is used to describe the class distribution of sub-samples from local and global perspectives and to share the information of different subspaces.Moreover,based on the assumption of continuity of spectral importance,a capped loss function that is theoretically robust to isolated pixels is used in a serial or parallel manner to estimate the global importance of spectra and perform feature selection.The results of classification experiments show that the proposed method is not sensitive to redundant features,and can still obtain good performance when more features are selected.A large number of deployed hyperspectral sensors not only generate massive amounts of dynamic data,but also increase the probability of missing spectral features caused by physical system failures.Hence,methods that require all data and spectral features to be loaded at once are no longer applicable.At this time,the low efficiency of manual labeling makes the problem of insufficient labeled samples more prominent.In this dissertation,an efficient unsupervised feature selection is proposed for new problems in the context of big data.Due to the complexity of hyperspectral pixels,this method embeds feature selection into a linearly weighted self-learning representation model.Subsequently,the representativeness of the selected features is enhanced by aligning the local manifold structure between the pixels.Then,based on the cache vector technique,samples containing a portion of the features from different periods are processed in a stream.The classification results show that the proposed method has the ability to process remote sensing data at different times in real time while ensuring the performance of labeling hyperspectral pixels.