Search

Graph convolutional networks (GCN) have emerged as a powerful alternative tool for analyzing hyperspectral images (HSIs). Despite their impressive performance, current works strive to make GCN more sophisticated through either elaborate architecture or fancy training tricks, making them prohibitive for HSI data in practice. In this paper, we present a Graph Convolutional RVFL Network (GCRVFL), a simple but efficient GCN for hyperspectral image classification. Specifically, we generalize the classic RVFL network into the graph domain by using graph convolution operations. This not only enables RVFL to handle graph-structured data, but also avoids iterative parameter adjustment by employing an efficient closed-form solution. Unlike previous works that perform HSI classification under a transductive framework, we regard HSI classification as a graph-level classification task, which makes GCRVFL scalable to large-scale HSI data. Extensive experiments on three benchmark data sets demonstrate that the proposed GCRVFL is able to achieve competitive results with fewer trainable parameters and adjustable hyperparameters and higher computational efficiency. In particular, we show that our approach is comparable to many existing approaches, including deep CNN models (e.g., ResNet and DenseNet) and popular GCN models (e.g., SGC and APPNP). [ABSTRACT FROM AUTHOR]

Randomized neural networks (NNs), such as random vector functional link (RVFL) and extreme learning machine (ELM), have been widely applied in various classification problems owing to their computational efficiency and universal approximation capability. However, such approaches are designed for regular Euclidean data and lack the ability to generalize to complex structured data. Moreover, their randomly generated parameters often lead to a suboptimal decision boundary with a growing requirement of hidden neurons. In this article, we first propose a plain framework, termed randomized graph convolutional networks (RGCNs), to generalize the classic randomized NNs to the non-Euclidean domain. Then, a hybrid framework called evolution-driven RGCN (EvoRGCN) is presented by using adaptive differential evolution with novelty search strategy to seek the globally optimal graph embedding for the plain RGCN. Finally, we recast the classic ELM and RVFL under the proposed frameworks, resulting in four novel semi-supervised models, including the plain models [i.e., graph convolutional extreme learning machines (GCELMs) and graph convolutional RVFL (GCRVFL)] and the optimized models (i.e., O-GCELM and O-GCRVFL). We show that our approaches are the natural generalization of the traditional randomized NNs in the non-Euclidean domain. Furthermore, our approaches not only retain the advantages of the classic approaches but also enable them to handle graph data. We compare our approaches against many existing methods across regular datasets and graph benchmarks, demonstrating that the proposed approaches dramatically outperform the compared methods with better generalization ability and robustness. Particularly, we quantitatively show the performance ranking of different randomized NNs, i.e., O-GCRVFL $> $ O-GCELM $\approx $ GCRVFL $> $ GCELM $\approx $ RVFL $> $ ELM. [ABSTRACT FROM AUTHOR]

Deep subspace clustering (DSC) has achieved remarkable performances in the unsupervised classification of hyperspectral images. However, previous models based on pixel-level self-expressiveness of data suffer from the exponential growth of computational complexity and access memory requirements with an increasing number of samples, thus leading to poor applicability to large hyperspectral images. This article presents a neighborhood contrastive subspace clustering (NCSC) network, a scalable and robust DSC approach, for unsupervised classification of large hyperspectral images. Instead of using a conventional autoencoder, we devise a novel superpixel pooling autoencoder to learn the superpixel-level latent representation and subspace, allowing a contracted self-expressive layer. To encourage a robust subspace representation, we propose a novel neighborhood contrastive regularization to maximize the agreement between positive samples in subspace. We jointly train the resulting model in an end-to-end fashion by optimizing an adaptively weighted multitask loss. Extensive experiments on three hyperspectral benchmarks demonstrate the effectiveness of the proposed approach and its substantial advancement of state-of-the-art approaches. [ABSTRACT FROM AUTHOR]

Hyperspectral image (HSI) clustering is a challenging task due to the high complexity of HSI data. Subspace clustering has been proven to be powerful for exploiting the intrinsic relationship between data points. Despite the impressive performance in the HSI clustering, traditional subspace clustering methods often ignore the inherent structural information among data. In this article, we revisit the subspace clustering with graph convolution and present a novel subspace clustering framework called graph convolutional subspace clustering (GCSC) for robust HSI clustering. Specifically, the framework recasts the self-expressiveness property of the data into the non-Euclidean domain, which results in a more robust graph embedding dictionary. We show that traditional subspace clustering models are the special forms of our framework with the Euclidean data. On the basis of the framework, we further propose two novel subspace clustering models by using the Frobenius norm, namely efficient GCSC (EGCSC) and efficient kernel GCSC (EKGCSC). Each model has a globally optimal closed-form solution, making it easier to implement, train, and apply in practice. Extensive experiments strongly evidence that EGCSC and EKGCSC dramatically outperform current models on three popular HSI data sets consistently. [ABSTRACT FROM AUTHOR]

Extreme Learning Machine (ELM) has gained lots of research interests due to its universal approximation capability and fast learning speed. Although several prior works have focused on developing deep ELM, it is still an open problem to design effective deep ELM. Stacking random layers will result in overfitting and accumulation of random errors. To address this issue, this paper presents a simple yet effective deep ELM called Densely Connected Convolutional ELM (DC2ELM) for hyperspectral image spectral-spatial classification. First, we introduce dense connection into ELM to make full use of intermediate feature maps produced by randomized convolutional layers, which is beneficial to reduce the random error. Secondly, stacked ELM auto-encoders are employed to generate reduced representation, leading to a deeper architecture. The proposed approach consists of fewer trainable parameters than traditional convolutional neural networks and can easily be trained without any iterative parameters tuning, making it easier to implement and apply in practice. We compare the proposed approach with many prior arts over three real hyperspectral images, showing that the proposed approach can achieve superior performance using limited training data and with a reduced risk of overfitting the training data. [ABSTRACT FROM AUTHOR]

Hyperspectral image (HSI) consists of hundreds of continuous narrowbands with high spectral correlation, which would lead to the so-called Hughes phenomenon and the high computational cost in processing. Band selection (BS) has been proven to be effective in avoiding such problems by removing redundant bands. However, many existing BS methods separately estimate the significance for every single band and cannot fully consider the nonlinear and global interaction between spectral bands. In this article, by assuming that a complete HSI band set can be reconstructed from its few informative bands, we propose a unified BS framework, BS Network (BS-Net). The framework consists of a band attention module (BAM), which aims to explicitly model the nonlinear interdependences between spectral bands, and a reconstruction network (RecNet), which is used to restore the original HSI from the learned informative bands, resulting in a flexible architecture. The resulting framework is end-to-end trainable, making it easier to train from scratch and to combine with many existing networks. We implement two versions of BS-Nets, respectively, using fully connected networks (BS-Net-FC) and convolutional neural networks (BS-Net-Conv), and extensively compare their results with popular existing BS approaches on three real hyperspectral data sets, showing that the proposed BS-Nets can accurately select informative band subset with less redundancy and outperform the competitors in terms of classification accuracy with competitive time cost. [ABSTRACT FROM AUTHOR]

Highlights • A novel hierarchical ensemble of ELM integrated with representation learning and ensemble learning. • Deep cascade structure is used for re-representing features. • Sparse connection and feature bagging are used to encourage individual diversity. • HE-ELM significantly outperforms many existing ensemble and representation learning methods. Abstract Extreme Learning Machine (ELM), which is proposed for generalized single-hidden layer feedforward neural networks, has become a popular research topic due to its unique characteristics. However, the random nature inherent in ELM's hidden layer results in unstable performance and a large number of hidden neurons is required, making the risk of overfitting increased. In this paper, we propose a simple but effective ensemble approach, called Hierarchical Ensemble of Extreme Learning Machine (HE-ELM), to improve ELM. To encourage the diversity of component ELMs, two strategies are taken into account, namely, the sparse connection to component ELMs and feature bagging. The resulting architecture is able to integrate both representation learning and ensemble learning with relatively fewer parameters and consists of independent component ELMs, making it easy to implement, train, and apply in practice. We compare results of the proposed HE-ELM with existing methods for 22 classification problems, showing that HE-ELM is able to achieve significant improvement in terms of classification accuracy, with a reduced risk of overfitting the training data. [ABSTRACT FROM AUTHOR]

Extreme Learning Machine (ELM) has been widely used for various classification problems. However, the traditional ELMs are typically based on the regular Euclidean data, thus ignoring the intrinsic structured information among data and resulting in poor robustness. In this paper, we present a novel semi-supervised learning framework, termed Graph Convolutional Extreme Learning Machines (GCELM), based on extending the traditional ELM to the non-Euclidean domain. Technically, we recast ELM layers into a randomized graph convolutional embedding layer followed by a graph convolutional regression layer, which endows ELM with the capability to handle graphs in the non-Euclidean domain. Benefiting from the diversity of the randomized graph convolution, we further propose an enhanced GCELM, i.e., Voting-based GCELM (V-GCELM), by using a simple voting ensemble strategy. The proposed methods preserve the advantages of ELMs, thus being more efficient than the gradient-based graph convolutional networks but without loss of graph learning ability. Extensive experiments on 36 benchmark datasets demonstrate that the proposed methods significantly outperform many previous semi-supervised classification methods. • Classic ELM is generalized into the non-Euclidean domain by using graph convolution. • An ensemble-based graph convolutional ELM is proposed for semi-supervised learning. • The proposed methods are benchmarked on 36 standard UCI datasets. • Our methods significantly outperform many existing methods by a large margin. [ABSTRACT FROM AUTHOR]

Hyperspectral image (HSI) consists of hundreds of narrow spectral band components with rich spectral and spatial information. Extreme Learning Machine (ELM) has been widely used for HSI analysis. However, the classical ELM is difficult to use for sparse feature leaning due to its randomly generated hidden layer. In this paper, we propose a novel unsupervised sparse feature learning approach, called Evolutionary Multiobjective-based ELM (EMO-ELM), and apply it to HSI feature extraction. Specifically, we represent the task of constructing the ELM Autoencoder (ELM-AE) as a multiobjective optimization problem that takes the sparsity of hidden layer outputs and the reconstruction error as two conflicting objectives. Then, we adopt an Evolutionary Multiobjective Optimization (EMO) method to solve the two objectives, simultaneously. To find the best solution from the Pareto solution set and construct the best trade-off feature extractor, a curvature-based method is proposed to focus on the knee area of the Pareto solutions. Benefited from the EMO, the proposed EMO-ELM is less prone to fall into a local minimum and has fewer trainable parameters than gradient-based AEs. Experiments on two real HSIs demonstrate that the features learned by EMO-ELM not only preserve better sparsity but also achieve superior separability than many existing feature learning methods. [ABSTRACT FROM AUTHOR]

Due to prior knowledge deficiency, large spectral variability, and high dimension of hyperspectral image (HSI), HSI clustering is extremally a fundamental but challenging task. Deep clustering methods have achieved remarkable success and have attracted increasing attention in unsupervised HSI classification (HSIC). However, the poor robustness, adaptability, and feature presentation limit their practical applications to complex large-scale HSI datasets. Thus, this article introduces a novel self-supervised locality preserving low-pass graph convolutional embedding method (L2GCC) for large-scale hyperspectral image clustering. Specifically, a spectral–spatial transformation HSI preprocessing mechanism is introduced to learn superpixel-level spectral–spatial features from HSI and reduce the number of graph nodes for subsequent network processing. In addition, locality preserving low-pass graph convolutional embedding autoencoder is proposed, in which the low-pass graph convolution and layerwise graph attention are designed to extract the smoother features and preserve layerwise locality features, respectively. Finally, we develop a self-training strategy, in which a self-training clustering objective employs soft labels to supervise the clustering process and obtain appropriate hidden representations for node clustering. L2GCC is an end-to-end training network, which is jointly optimized by graph reconstruction loss and self-training clustering loss. On Indian Pines, Salinas, and University of Houston 2013 datasets, the clustering accuracy overall accuracies (OAs) of the proposed L2GCC are 73.51%, 83.15%, and 64.12%, respectively. [ABSTRACT FROM AUTHOR]

Deep Learning models have been successfully applied to the classification of hyperspectral images (HSIs), among which Convolutional Neural Network (CNN) can extract more efficient spatial–spectral features to improve classification accuracy compared to traditional classifiers such as Support Vector Machine. However, in the case of limited labelled samples with inevitable noise, simply applying CNN for HSIs classification could be unsatisfactory due to the issue of overfitting. To address the problem, data preprocessing techniques such as filtering methods can be used to effectively remove the noise and enhance the spatial–spectral features. Compared to typical filters like Gabor, Propagation Filter (PF) can effectively smooth the HSIs and preserve the edges without causing the cross-region mixing problem, which could significantly compromise the performance of CNN based HSIs classification models. Therefore, in this paper a new model which combine Propagation Filter with CNN (PF-CNN) is proposed. It makes use of the advantages of PF and CNN to effectively handle the cross-region mixing issue for HSIs classification. The experimental results show that the proposed algorithm outperforms other state-of-the-art HSIs classification models. The source code of the proposed model is available at . [ABSTRACT FROM AUTHOR]

Hyperspectral anomaly detection (HAD) aims to distinguish anomalies from background-by-background modeling. Deep learning has been applied to HAD and achieves promising detection results. However, there exist several issues that need to be addressed: 1) unrealistic Gaussian assumption on the latent representations may limit its application; 2) deep features are not well-suited to anomaly detection due to the separation between feature learning and anomaly detection; 3) lack of adequate exploitation of spectral-spatial features; 4) negative effect caused by spectral band redundancy. In this article, we propose an end-to-end trainable deep one-class classification network for HAD. Specifically, a minimal enclosing hypersphere is trained to involve the deep features of background samples. These background samples are selected by a density clustering-based method. In this way, feature learning and anomaly detection are incorporated into a unified framework. Meanwhile, there is no explicit Gaussian assumption on the background features. Moreover, due to the complementarity of spectral and spatial features, a novel feature fusion strategy is proposed to fuse spectral and spatial features extracted by a two-stream deep convolutional autoencoder network. Finally, a band attention module is used to automatically learn small weights for redundant bands and thus reduce the negative effect caused by redundant bands. Experimental results on five public datasets demonstrate the superiority of the proposed method compared to several state-of-the-art HAD methods in the detection performance. [ABSTRACT FROM AUTHOR]

Currently, deep neural networks (DNNs) are an important method for handling hyperspectral image (HSI) classification because of their good performance in image processing. However, DNNs’ performance depends on a massive number of training data and hyperparameters that are carefully fine-tuned, which results in structural complexity and a time-consuming process. Deep forest is a novel deep learning method that does not need much training data and has a simple structure. In this paper, we first design a deep forest for spectral-based HSI classification and then propose an improved deep forest algorithm, named deep multigrained cascade forest (dgcForest), for spatial-based HSI classification. On the one hand, the cascade forest in dgcForest is used in multigrained scanning, which enhances the performance; on the other hand, a pooling layer is added after the multigrained scanning to reduce the output dimensions. To demonstrate that our proposed algorithm presents a good performance in HSI classification, we analyze the hyperparameters of deep forest and dgcForest and compare them with other methods on the biased and unbiased data sets, which illustrates that our method is superior to other state-of-the-art deep learning methods. [ABSTRACT FROM AUTHOR]

Highlights • We propose a multiobjective optimization-based sparse extreme learning machine (MO-SELM) for classification and regression tasks. • The sparse connecting structure of ELM is designed for learning a more compact network. • We improve MOEA/D to optimize the proposed multiobjective model and make decisions by ensemble learning. • Experimental results reveal the superior performance of the proposed MO-SELM. Abstract Extreme Learning Machine (ELM) is a popular machine learning method and has been widely applied to real-world problems due to its fast training speed and good generalization performance. However, in ELM, the randomly assigned input weights and hidden biases usually degrade the generalization performance. Furthermore, ELM is considered as an empirical risk minimization model and easily leads to overfitting when dataset exists some outliers. In this paper, we proposed a novel algorithm named Multiobjective Optimization-based Sparse Extreme Learning Machine (MO-SELM), where parameter optimization and structure learning are integrated into the learning process to simultaneously enhance the generalization performance and alleviate the overfitting problem. In MO-SELM, the training error and the connecting sparsity are taken as two conflicting objectives of the multiobjective model, which aims to find sparse connecting structures with optimal weights and biases. Then, a hybrid encoding-based MOEA/D is used to optimize the multiobjective model. In addition, ensemble learning is embedded into this algorithm to make decisions after multiobjective optimization. Experimental results of several classification and regression applications demonstrate the effectiveness of the proposed MO-SELM. [ABSTRACT FROM AUTHOR]

• MARP can perform excellent performance with low computational complexity. • Nodes impaction on classification is suppressed via proposed ARP mechanism. • LSTM is adopted to update nodes and preserve local convolutional feature. • Avoiding local optimization problem by multi-scale perception mechanism. In recent years, the applications of graph convolutional networks (GCNs) in hyperspectral image (HSI) classification have attracted much attention. However, hyperspectral classification faces problems such as complex noise effects, spectral variability, labelled training sample deficiency, and high spectral mixing between materials. Furthermore, the available GCN-based methods are computationally complex and cannot automatically adjust aggregate paths. To mitigate these issues, we propose a novel multiadaptive receptive field-based graph neural framework (MARP) for HSI classification. In our method, an adaptive receptive path aggregation (ARP) mechanism is proposed to suppress the impact of noise nodes on classification and automatically explore an adaptive receptive field, where a graph attention (GAT) neural network is introduced to learn the importance of different-sized neighbourhoods and a long short-term memory (LSTM) method is adopted to update the nodes and preserve the local convolutional features of the nodes. To address the problem that ARP may fall into a local optimum, we design a multiscale receptive mechanism. Extensive experimental results obtained on four public HSI datasets demonstrate that the proposed MARP method can mitigate oversmoothing and reduce computational complexity while achieving competitive performance when compared to several state-of-the-art methods. [ABSTRACT FROM AUTHOR]