Search

Self-supervised learning effectively leverages the information from unlabeled data to extract spatial-spectral features that are both representative and discriminative, partially addressing the challenge of high data annotation costs in hyperspectral image classification. Inspired by the success of redundancy reduction-based self-supervised learning in other domains, we introduce it into HSIC. We proposed a spatial-spectral feature extraction network, HyperBT, to more effectively reduce redundancy. Specifically, we added the off-diagonal terms of the cross-covariance matrix to the loss function and new data augmentation methods, including band bisection and edge weakening. Experimental results demonstrate that our method achieves high accuracy in classification, surpassing many state-of-the-art methods. Through ablation experiments, we validate the effectiveness of each component in the loss function.

Hyperspectral images (HSIs) encounter an inherent challenge due to spectral variability (SV), which directly impacts the accuracy of endmember extraction and abundance estimation. Most unmixing networks based on autoencoder (AE) overlook the impact of SV and instead focus more on exploring the features of endmembers. In this article, we propose a feedback information-guided SV attention network (FSVA-Net), an AE-based neural network that utilizes a feedback information-guided structure to exploit and model the SV factors. We develop a feedback enhancement module (FEM) that utilizes residuals from a linearly reconstructed image to reweight and emphasize pixels affected by SV. And an SV attention (SVA) mechanism is proposed to explore interference-affected information in a global perspective according to the enhanced feature. With the extracted SV-related feature, we design a generative augmented linear mixing model (ALMM)-based decoder to model SV from the perspective of scaling and perturbation in a more reliable way. In a more concern on SV architecture, the proposed network achieves excellent performance on both synthetic and real datasets.

As an important task in hyperspectral image (HSI) processing, hyperspectral anomaly detection has gained increasing attention and has been extensively studied. However, most existing methods mostly focus on the statistics spectral distribution of pixels and the spatial correlation of pixels, while ignoring pixels’ correlation in the spectral space. Compared with statistics spectral distribution, the spectral correlation represents not only the spectral distribution but also the relations between different pixels in the spectral space. In this article, we propose a novel graph-based hyperspectral anomaly detection method which exploits the correlation of pixels in spectral space. We model the spectral correlation through the structure of the graph which is constructed by a KNN-like strategy. Meanwhile, edge modification (EM) operations including the edge adding (EA) and edge deleting (ED) operations are adopted to further optimize the graph structure. Once the graph structure is determined, the vertices with fewer edges connected will be masked iteratively. This is because vertices with fewer connections have a higher probability of being anomalous compared to others. Then, all the masked vertices will be reconstructed using an iterative graph propagation updating strategy. Since background pixels exhibit higher correlation, they can be effectively reconstructed through graph propagation. However, anomalies are more challenging to reconstruct due to their low spectral correlation. Experiments are conducted on four real HSI datasets which demonstrate the superior performance of the proposed iterative graph propagation anomaly detector (IGPAD) method in detecting small-size anomalies compared with other graph-related and state-of-the-art anomaly detection methods.

In recent years, tensor representation (TR)-based hyperspectral anomaly detection (HAD) approaches have attracted more and more attention. However, two urgent issues still need to be addressed: 1) existing tensor decomposition approaches for HAD cannot make full use of the spectral–spatial correlation of background components in hyperspectral images (HSIs) and 2) most approaches based on TR overlook the piecewise-smooth of background components that exist simultaneously in the spectral and spatial domains. To this end, with the aid of an advanced multiscale entanglement renormalization ansatz (MERA) tensor network, this article proposes an algorithm based on MERA decomposition and enhanced total variation regularization (MERAETV) for HAD. Specifically, MERA decomposes the background tensor by contracting a top-level factor with the remaining semiorthogonal and orthogonal factors. Due to the intricate interplay between semiorthogonal (low-rank) and orthogonal factors, low-rank MERA approximation exhibits a robust representational capacity that effectively captures the spectral–spatial correlation of the background component. Meanwhile, an enhanced total variation (ETV) regularization is devised to capture the inherent piecewise-smooth of the background component in both spectral and spatial domains. Furthermore, our algorithm incorporates group sparsity constraint and Gaussian noise term to enhance the discrimination between anomalies and background. Finally, a highly efficient update scheme based on the alternating direction method of multipliers (ADMM) is designed. A large number of experiments on one synthetic and seven real HSIs demonstrate the superiority of our proposed approach.

Autoencoders (AEs) have demonstrated excellent performance in the field of hyperspectral unmixing (SU), due to their self-supervised nature and ease of implementation. Recently proposed AE-based networks contend that local spatial information limits further improvement in unmixing accuracy and tends to explore and utilize global information, which improves unmixing accuracy at the expense of increased computational complexity. However, we believe that precise unmixing can be achieved by fully leveraging local information. In this article, we propose a unidirectional local-attention AE network (ULA-Net) that explores spatial information pixel by pixel and achieves accurate spatial–spectral feature fusion. ULA-Net utilizes unidirectional local attention (ULA) module to calculate the correlation between neighboring pixels and the central pixel within local regions, extracting discriminative local information. Moreover, ULA-Net effectively extracts relevant spatial information and suppresses irrelevant information based on a double fusion strategy (DFS) module. This process achieves more accurate control over the contribution of spatial information by implementing information fusion in both the pixel and feature dimensions. To address spectral variability, we implement the extended linear mixing model (ELMM) in the decoder part to improve unmixing accuracy without increasing the number of parameters. We conduct ablation experiments to investigate the roles of each module. Experimental results on both synthetic and real datasets demonstrate the effectiveness of the proposed network.

With the advantages of being scalable, labeled-sample free, and structurally simple, autoencoder (AE) is suitable for hyperspectral target detection (HTD) with discriminative attributes extraction ability. However, there are some bottlenecks in current AE-based approaches, like network complexity, parameter determination, and dictionary reliability problems. In this article, a novel target-driven AE (TAE) with a constrained latent code layer and improved loss function is proposed. To achieve effective noise and interference annihilation, a simple AE-based network is designed with the adaptively determined number of hidden neurons by estimating the ranks of the background (BKG) and the target component. In addition, to untangle the dependence on the BKG dictionary, a novel strategy is imposed on the traditional Kullback–Leibler (KL) divergence, called the truncated KL divergence (TKL), is designed and imposed by the loss function of TAE, guiding the propensity for network reconstruction of the targets over BKG with probabilities rather than dictionaries. Furthermore, unlike AE-based methods with forward strategy, a novel target-driven iterative AE (ITAE) framework is developed to further exploit the potential of prior and posterior knowledge. The iterative system repeatedly trains the same TAE and adds the weighted constrained energy minimization (CEM) detection map back to the current data matrix for the next iteration. ITAE provides a general classic-deep learning collaborative framework that can train any AE with a traditional detector through an iterative process to improve the diversity between the targets and the BKG. To validate the effectiveness of ITAE, six other state-of-the-art HTD methods are chosen for comparison under six datasets of various scenarios. Based on the 3-D receiver operating characteristic (ROC) curve-derived evaluation metric, we conducted experiments for parameter analysis, detection performance comparison, noisy study, and ablation study. Results show that ITAE outperforms other methods, especially in target detectability.

Infrared (IR) small target detection is a crucial aspect in both military and civilian fields. Due to the poor quality and complex scenarios, the existing technologies lack superior real-time and overall detection performance. Tensor analysis has shown superiority, but there are three key issues. The first pertains to appropriate tensor-structured data, the second concerns a more comprehensive tensor decomposition architecture, and the third is to achieve more satisfactory real-time performance. This article proposes an IR small target detection method named 4-D spatial–temporal tensor decomposition with block term decomposition-based norm and multidirectional derivative-based priors (4DST-BTMD). It converts the target detection task into a low-rank and sparse decomposition (LRSD) optimization problem of decomposing background, target, and noise tensors from 4-D spatial–temporal domain. First, we construct a 4-D spatial–temporal tensor, followed by projecting it into a constructed data-sphered space, which lays the data foundation for decomposition. Then, based on the 4-D sphered tensor, a low-rank surrogate named block term decomposition-based norm (BTDN) for background estimation is proposed, which fully integrates global information across different dimensions. Meanwhile, more effective salient 4-D prior tensors based on multidirectional derivatives are designed to effectively guide the model to focus on significant areas. Finally, an efficient ADMM-based solving framework is designed for the LRSD model. Furthermore, with seven state-of-the-art competitive methods, extensive experiments and analysis illustrate that the proposed 4DST-BTMD not only demonstrates the superiority in background suppressibility (BS) and target detectability (TD) but also has satisfactory real-time detection performance.

The poor quality images with scarce target characteristics and complicated textures make it tough to enhance target detectability (TD) and background suppressibility (BS) simultaneously in infrared (IR) small target detection. Representation learning has gradually shown superiority, but two key issues arise. The first is constructing an appropriate dictionary. The second is a comprehensive and accurate representation learning model. This article introduces a low-rank and sparse spatial-temporal tensor representation learning model based on local binary contrast measure (STRL-LBCM), which transforms the target detection into an optimization task for representation and decomposition of background, target, and noise components from spatial-temporal aspects. First, we construct an IR spatial-temporal tensor from the original sequence, followed by improving GoDec plus to obtain a sparse part and a background dictionary, which lays the foundation for representation learning model. Then, a rank surrogate named fully connected tensor nuclear norm (FCTNN) is proposed for low-rank coefficient, aiming at low-rank background estimation. Meanwhile, with the sparse part, we design a salient prior called local binary contrast measure, which extracts more effective prior information for guiding target detection. Furthermore, not only Frobenius norm but also a 3-D weight total variation regularization is utilized to preserve background details and remove noise. Finally, an efficient solving framework is designed for the proposed model. Comprehensive experiments reveal that STRL-LBCM provides superior TD, BS, and overall performance compared with 16 target detection methods.