Feature Extraction of Hyperspectral Images With Semi-supervised Sparse Graph Learning

Gene expression profiles have been used to predict cancer recurrence or other clinical outcomes of cancer patients. However, clinical information of cancer patients is often incomplete, which yields many unlabeled samples that cannot be used in supervised learning. In this is paper, we develop a novel semi-supervised leaning (SSL) method that uses both labeled and unlabeled patient samples to predict cancer recurrence. Our new SSL algorithm employs a sparse representation approach where a labeled sample is represented as a combination of a small number of properly chosen unlabeled samples. Experiments with a set of gene expression data from patients with colorectal cancer(CRC) demonstrate that our SSL algorithm can improve prediction accuracy compared to other two SSL methods including TSVM and T3VM, and the traditional support vector machine.

In the health monitoring of electromechanical transmission systems, the collected state data typically consist of only a minimal amount of labeled data, with a vast majority remaining unlabeled. Consequently, deep learning-based diagnostic models encounter the challenge of scarcity in labeled data and abundance in unlabeled data. Traditional semi-supervised deep learning methods based on pseudo-label self-training, while alleviating the issue of labeled data scarcity to some extent, neglect the reliability of pseudo-label information, the accuracy of feature extraction from unlabeled data, and the imbalance in sample selection. To address these issues, this paper proposes a novel semi-supervised fault diagnosis method under imbalanced unlabeled sample class information screening. Firstly, an information screening mechanism for unlabeled data based on active learning is established. This mechanism discriminates based on the variability of intrinsic feature information in fault samples, accurately screening out unlabeled samples located near decision boundaries that are difficult to separate clearly. Then, combining the maximum membership degree of these unlabeled data in the classification space of the supervised model and interacting with the active learning expert system, label information is assigned to the screened unlabeled data. Secondly, a cost-sensitive function driven by data imbalance is constructed to address the class imbalance problem in unlabeled sample screening, adaptively adjusting the weights of different class samples during model training to guide the training of the supervised model. Ultimately, through dynamic optimization of the supervised model and the feature extraction capability of unlabeled samples, the recognition ability of the diagnostic model for unlabeled samples is significantly enhanced. Validation through two datasets, encompassing a total of 12 experimental scenarios, demonstrates that in scenarios with only a small amount of labeled data, the proposed method achieves a diagnostic accuracy increment exceeding 10% compared to existing typical methods, fully validating the effectiveness and superiority of the proposed method in practical applications.

Part 3 of this Special Section on High-Performance Computing in Applied Remote Sensing continues to present the state-of-the-art research in incorporating high-performance computing facilities and algorithms for effective and efficient remote sensing applications. “Land cover classification in multispectral imagery using clustering of sparse approximations over learned feature dictionaries” by Moody et al. presents results from an ongoing effort to extend neuromimetic machine vision algorithms to multispectral data using adaptive signal processing combined with compressive sensing and machine learning techniques. “Graphics processor unit accelerated finite-difference time domain method for electromagnetic scattering from onedimensional large scale rough soil surface at low grazing incidence” by Jia et al. applies the Compute Unified Device Architecture (CUDA) technology to compute an electromagnetic scattering problem on GPUs and achieves significant speedup factors. “High-performance meshing processing of remote sensing data on large displays” by Hsieh et al. presents a technique that visualizes LIDAR point clouds on a tiled display wall, termed highly interactive-parallelized display (HIPerDisplay), which has twenty 24-inch liquid-crystal displays with a total resolution of 46 Mpixels. The results also showed that the HIPerDisplay offered superior performance for the processing of large LIDAR datasets. “Real-time implementation of optimized maximum noise fraction transform for feature extraction of hyperspectral images” by Wu et al. presents a parallel implementation of the optimized maximum noise fraction (G-OMNF) transform algorithm for feature extraction of hyperspectral images on commodity graphics processing units (GPUs). The proposed approach explored the algorithm data-level concurrency and optimized the computing flow. “Hyperspectral band selection based on parallel particle swarm optimization and impurity function band prioritization schemes” by Chang et al. presents a framework for band selection in hyperspectral imagery that uses two techniques, referred to as particle swarm optimization (PSO) band selection and the impurity function band prioritization (IFBP) method. The experimental results demonstrate that the proposed PPSO/IFBP band selection method can not only improve computational speed but also offer a satisfactory classification performance. “Graphics processing unit implementation of the maximum likelihood solution to the inverse problem for retrieval of geophysical parameters from high-resolution sounder data” by Wei et al. explores the use of GPU in speedup of the maximum likelihood solution to an illposed retrieval problem. For the infrared atmospheric sounding interferometer high-resolution sounder having 8641 channels, the use of GPU on maximum likelihood estimation shows a promising speedup of 1986× compared to a single-threaded native CPU version.

In this letter, we propose a new semisupervised learning (SSL) algorithm for remotely sensed hyperspectral image classification. Our main contribution is the development of a new soft sparse multinomial logistic regression model which exploits both hard and soft labels. In our terminology, these labels respectively correspond to labeled and unlabeled training samples. The proposed algorithm represents an innovative contribution with regard to conventional SSL algorithms that only assign hard labels to unlabeled samples. The effectiveness of our proposed method is evaluated via experiments with real hyperspectral images, in which comparisons with conventional semisupervised self-learning algorithms with hard labels are carried out. In such comparisons, our method exhibits state-of-the-art performance.

It is important to select representative patterns for successful classification of remotely sensing images. In the paper, a superpixel-based active learning (shorted as SPAL) algorithm is proposed to iteratively select informative unlabeled samples for labeling by experts. In particular, new representation of a remote sensing image is generated by grouping the data into perceptually meaningful atomic regions, i.e., superpixels which have both spectral similarity and spatial adjacency. Then, active learning is iteratively exploited to select informative samples to enlarge the training dataset. At each iteration, the unlabeled samples are firstly sorted according to a prefixed uncertain strategy. Then, the few highest uncertain samples are selected on the assumption that all the samples selected at each iteration are from different superpixels in order to guarantee the diversity. The proposed SPAL algorithm is validated by real hyperspectral images. Experimental results show that the SPAL can improve the classification accuracies as well as significant enhancement of computational efficiency.

Remotely sensed hyperspectral imaging allows for the detailed analysis of the surface of the Earth using advanced imaging instruments which can produce high-dimensional images with hundreds of spectral bands. Supervised hyperspectral image classification is a difficult task due to the unbalance between the high dimensionality of the data and the limited availability of labeled training samples in real analysis scenarios. While the collection of labeled samples is generally difficult, expensive, and time-consuming, unlabeled samples can be generated in a much easier way. This observation has fostered the idea of adopting semisupervised learning techniques in hyperspectral image classification. The main assumption of such techniques is that the new (unlabeled) training samples can be obtained from a (limited) set of available labeled samples without significant effort/cost. In this paper, we develop a new approach for semisupervised learning which adapts available active learning methods (in which a trained expert actively selects unlabeled samples) to a self-learning framework in which the machine learning algorithm itself selects the most useful and informative unlabeled samples for classification purposes. In this way, the labels of the selected pixels are estimated by the classifier itself, with the advantage that no extra cost is required for labeling the selected pixels using this machine–machine framework when compared with traditional machine–human active learning. The proposed approach is illustrated with two different classifiers: multinomial logistic regression and a probabilistic pixelwise support vector machine. Our experimental results with real hyperspectral images collected by the National Aeronautics and Space Administration Jet Propulsion Laboratory's Airborne Visible–Infrared Imaging Spectrometer and the Reflective Optics Spectrographic Imaging System indicate that the use of self-learning represents an effective and promising strategy in the context of hyperspectral image classification.

Feature extraction is a very important preprocessing step for classification of hyperspectral images. The linear discriminant analysis (LDA) method fails to work in small sample size situations. Moreover, LDA has poor efficiency for non-Gaussian data. LDA is optimized by a global criterion. Thus, it is not sufficiently flexible to cope with the multi-modal distributed data. We propose a new feature extraction method in this paper, which uses the boundary semi-labeled samples for solving small sample size problem. The proposed method, which called hybrid feature extraction based on boundary semi-labeled samples (HFE-BSL), uses a hybrid criterion that integrates both the local and global criteria for feature extraction. Thus, it is robust and flexible. The experimental results with three real hyperspectral images show the good efficiency of HFE-BSL compared to some popular and state-of-the-art feature extraction methods.

<strong class="journal-contentHeaderColor">Abstract.</strong> Hyperspectral images contain abundant spectral and spatial information about the earth's surface, labeling data processing and analysis more difficult, as well as the problem of sample labeling. In this paper, local binary pattern (LBP), sparse representation and mixed logistic regression model are introduced, and a sample labeling method based on neighborhood information and priority classifier discrimination is presented. Then, a hyperspectral remote sensing image classification method based on texture features and semi-supervised learning is implemented. The LBP is employed to extract features of spatial texture information from remote sensing images and enrich the feature information of samples. Then the multivariate logistic regression model is used to select the unlabeled samples with the largest amount of information, and the unlabeled samples with neighborhood information and priority classifier tags are selected to obtain the pseudo-labeled samples after learning. By making full use of the advantages of sparse representation and mixed logistic regression model, a new hyperspectral remote sensing image classification model based on semi-supervised learning is constructed to effectively achieve accurate classification of hyperspectral images. The data of Indian Pines, Salinas scene and Pavia University are selected to verify the validity of the proposed method. The experiment results show that the proposed classification method can obtain higher classification accuracy and show stronger timeliness and generalization ability.

In this paper, we proposed a new semi-supervised multi-manifold learning method, called semi- supervised sparse multi-manifold embedding (S3MME), for dimensionality reduction of hyperspectral image data. S3MME exploits both the labeled and unlabeled data to adaptively find neighbors of each sample from the same manifold by using an optimization program based on sparse representation, and naturally gives relative importance to the labeled ones through a graph-based methodology. Then it tries to extract discriminative features on each manifold such that the data points in the same manifold become closer. The effectiveness of the proposed multi-manifold learning algorithm is demonstrated and compared through experiments on a real hyperspectral images.

In this paper, we propose a manifold geometry based projective nonnegative matrix factorization linear dimensionality reduction method, called neighborhood preserving orthogonal projective nonnegative matrix factorization (NPOPNMF), for feature extraction of hyperspectral image. By adding constraints on projective nonnegative matrix factorization (PNMF) that each data point can be represented as a linear combination of its neighbors, NPOPNMF preserves local neighborhood geometrical structure of hyperspectral data in the reduced space, and overcomes the Euclidean limitation of PNMF. The metric structure of original high-dimensional hyperspectral data space is preserved due to the orthogonality of projection matrix. NPOPNMF can be performed in either supervised or unsupervised mode according to the construction of adjacency graph and it can improve the discriminant performance of PNMF. Theoretical analysis and experimental results on hyperspectral data sets demonstrate that the proposed method is an effective and promising method for hyperspectral image feature extraction.

Feature extraction (FE) is a crucial research area in hyperspectral image (HSI) processing. Recently, due to the powerful ability of deep learning (DL) to extract spatial and spectral features, DL-based FE methods have shown great potentials for HSI processing. However, most of the DL-based FE methods are supervised, and the training of them suffers from the absence of labeled samples in HSIs severely. The training issue of supervised DL-based FE methods limits their application on HSI processing. To address this issue, in this paper, a novel modified generative adversarial network (GAN) is proposed to train a DL-based feature extractor without supervision. The designed GAN consists of two components, which are a generator and a discriminator. The generator can focus on the learning of real probability distributions of data sets and the discriminator can extract spatial-spectral features with superior invariance effectively. In order to learn upsampling and downsampling strategies adaptively during FE, the proposed generator and discriminator are designed based on a fully deconvolutional subnetwork and a fully convolutional subnetwork, respectively. Moreover, a novel min-max cost function is designed for training the proposed GAN in an end-to-end fashion without supervision, by utilizing the zero-sum game relationship between the generator and discriminator. Besides, the proposed modified GAN replaces the original Jensen-Shannon divergence with the Wasserstein distance, aiming to mitigate the unstability and difficulty of the training of GAN frameworks. Experimental results on three real data sets validate the effectiveness of the proposed method.

A novel manifold learning feature extraction approach—preserving neighborhood discriminant embedding (PNDE) of hyperspectral image is proposed in this paper. The local geometrical and discriminant structure of the data manifold can be accurately characterized by within-class neighboring graph and between-class neighboring graph. Unlike manifold learning, such as LLE, Isomap and LE, which cannot deal with new test samples and images larger than 70×70, the method here can process full scene hyperspectral images. Experiments results on hyperspectral datasets and real-word datasets show that the proposed method can efficiently reduce the dimensionality while maintaining high classification accuracy. In addition, only a small amount of training samples are needed.

In this paper, we propose a novel semi-supervised learning classification framework using box-based smooth ordering and multiple 1D-embedding-based interpolation (M1DEI) in [J. Wang, Semi-supervised learning using multiple one-dimensional embedding-based adaptive interpolation, Int. J. Wavelets Multiresolut. Inf. Process. 14(2) (2016) 11 pp.] for hyperspectral images. Due to the lack of labeled samples, conventional supervised approaches cannot generally perform efficient enough. On the other hand, obtaining labeled samples for hyperspectral image classification is difficult, expensive and time-consuming, while unlabeled samples are easily available. The proposed method can effectively overcome the lack of labeled samples by introducing new labeled samples from unlabeled samples in a label boosting framework. Furthermore, the proposed method uses spatial information from the pixels in the neighborhood of the current pixel to better catch the features of hyperspectral image. The proposed idea is that, first, we extract the box (cube data) of each pixel from its neighborhood, then apply multiple 1D interpolation to construct the classifier. Experimental results on three hyperspectral data sets demonstrate that the proposed method is efficient, and outperforms recent popular semi-supervised methods in terms of accuracies.

Deep learning-based methods have made significant progress in hyperspectral image (HSI) classification in recent years. However, deep learning-based methods usually rely on a large number of samples, and in many cases, it is difficult to label HSI and only limited training samples are available. To solve this problem, an HSI classification method based on adaptive spatial pyramid constraint (ASPC) is proposed to make full use of the global spatial neighborhood information of the labeled samples, which can improve the generalization ability of the classification model. The main steps of the proposed method are as follows. First, an HSI complexity evaluation method based on edge detection is proposed to assess the homogeneity of the objects in the HSI. Second, an HSI pyramid segmentation method based on spatial pyramid is proposed to generate multiscale subregions, where HSI complexity is used to adaptively determine the scale of the segmentation. Third, a spatial supervised constraint is proposed to generate the loss function of labeled subregions. Fourth, a spatial unsupervised constraint is proposed to generate the loss function of unlabeled subregions. The proposed method fully explores the spatial-spectral correlation between unlabeled samples and labeled samples, and add corresponding constraints to the training objective according to the correlation. By adding the ASPC, the trained model becomes more robust and can make full use of the limited training samples. To verify the effectiveness of the proposed method, three benchmark hyperspectral datasets are used to verify the performance of the proposed method. Experimental results show that the performance of this method is better than the existing state-of-the-art methods.

A graph convolutional network (GCN) employs a graph filtering kernel tailored for data with irregular structures. However, simply stacking more GCN layers does not improve performance; instead, the output converges to an uninformative low-dimensional subspace, where the convergence rate is characterized by the graph spectrum— this is the known over-smoothing problem in GCN. In this paper, we propose a sparse graph learning algorithm incorporating a new spectrum prior to compute a graph topology that circumvents over-smoothing while preserving pairwise correlations inherent in data. Specifically, based on a spectral analysis of multilayer GCN output, we derive a spectrum prior for the graph Laplacian matrix L to robustify the model expressiveness against over-smoothing. Then, we formulate a sparse graph learning problem with the spectrum prior, solved efficiently via block coordinate descent (BCD). Moreover, we optimize the weight parameter trading off the fidelity term with the spectrum prior, based on data smoothness on the original graph learned without spectrum manipulation. The output L is then normalized for supervised GCN training. Experiments show that our proposal produced deeper GCNs and higher prediction accuracy for regression and classification tasks compared to competing schemes.