As a key task in hyperspectral image processing, hyperspectral anomaly detection has garnered significant attention and undergone extensive research. Existing methods primarily relt on two prior assumption: low-rank background and sparse anomaly, along with additional spatial assumptions of the background. However, most methods only utilize the sparsity prior assumption for anomalies and rarely expand on this hypothesis. From observations of hyperspectral images, we find that anomalous pixels exhibit certain spatial distribution characteristics: they often manifest as small, clustered groups in space, which we refer to as cluster sparsity of anomalies. Then, we combined the cluster sparsity prior with the classical GoDec algorithm, incorporating the cluster sparsity prior into the S-step of GoDec. This resulted in a new hyperspectral anomaly detection method, which we called Turbo-GoDec. In this approach, we modeled the cluster sparsity prior of anomalies using a Markov random field and computed the marginal probabilities of anomalies through message passing on a factor graph. Locations with high anomalous probabilities were treated as the sparse component in the Turbo-GoDec. Experiments are conducted on three real hyperspectral image (HSI) datasets which demonstrate the superior performance of the proposed Turbo-GoDec method in detecting small-size anomalies comparing with the vanilla GoDec (LSMAD) and state-of-the-art anomaly detection methods.

Self-supervised learning has made significant strides in hyperspectral image classification. For contrastive learning, traditional data augmentation methods will lead to biases with real-world spectra, limited variability, loss of information. To address those limitations, this study introduces diffusion models to generate realistic synthetic samples as views input to contrastive learning. This diffusion-based augmentation enables our contrastive learning network to learn more robust spectral-spatial representations. Experimental results on two publicly available datasets indicate that this approach outperforms numerous existing methods. And the ablation study demonstrates the effectiveness of employing diffusion models.

In recent years, self-supervised learning has made significant strides in hyperspectral image classification [1]. However, different approaches come with distinct strengths and limitations. Contrastive learning excels at extracting key information from large volumes of redundant data, but its training objective can inadvertently increase intra-class feature distance. To address this limitation, we leverage diffusion models for their proven ability to refine and aggregate features by modeling complex data distributions. Specifically, diffusion models’ inherent denoising and generative process are theoretically well-suited to enhance intra-class compactness by learning to reconstruct clean, representative features from perturbed inputs. We propose the new method - ContrastDM. This approach generates synthetic features, improving and enriching feature representation, and partially addressing the issue of sample sparsity. Classification experiments on three publicly available datasets demonstrate that ContrastDM significantly outperforms state-of-the-art methods.

As a low-rank regularization technique, the tensor nuclear norm (TNN) has been extensively employed for hyperspectral anomaly detection (HAD). However, traditional TNN and its variants often suffer from rank estimation bias and are difficult to effectively capture the spectral-spatial correlation of complex backgrounds. In addition, the extracted sparse component is often submerged in the background, resulting in abnormal targets being insignificant. To overcome these problems, this letter proposes an enhanced low-rank and joint saliency prior (ELRJSP) method for HAD. Specifically, within the framework of tensor singular value decomposition (t-SVD), we propose an improved weighted tensor nuclear norm (IWTNN), defined as a weighted combination of the weighted tensor nuclear norm (WTNN) and the weighted nuclear norm of the core matrix. This novel norm effectively integrates information from both the original background tensor and its core tensor, thereby leveraging spectral-spatial structural characteristics more comprehensively and leading to enhanced accuracy in background estimation. Furthermore, in order to accurately detect and highlight abnormal targets, under the constraint of the tensor ℓF,1-norm, a visual saliency sparse weight tensor to constrain the abnormal tensor is designed by combining the visual saliency weight graph and the sparse optimized weight graph and extending them to the tensor domain. Comparative experiments on three real hyperspectral datasets demonstrate that the developed ELRJSP algorithm outperforms several advanced algorithms.

Recently, Mamba has made significant strides in sequence modeling, with its global receptive field, dynamic weighting strategy and linear growth in computational complexity. In remote sensing (RS) change detection (CD), several studies have demonstrated that Mambas leverage a unique scanning mechanism to traverse images from various directions, showcasing excellent long-range modeling capabilities. However, as the network depth increases, Mamba often struggle to retain shallow textures and local features effectively. In particular, modern RS images frequently capture complex surface scenes, including seasonal climate variations and densely built environments, making local contextual details crucial for effective CD. Therefore, a difference enhancement and inter-scale interactive fusion Mamba (DEIF-Mamba) is proposed to alleviate the issue. This entire network framework integrates CNN and Mamba, utilizing CNN to capture local feature information, while Mamba employs a cross-scanning mechanism to integrate global information. To address the interference caused by mixed texture features and the missed detection of subtle changes in complex scenes, a differential feature enhancement module (DFEM) is proposed to enrich local contextual details and improve feature representation. In addition, we propose an inter-scale interactive fusion (ISIF) strategy to fully utilize the cross-scale interactive information and minimize information redundancy. Extensive experiments on four CD datasets demonstrate that the proposed DEIF-Mamba achieves an average F1 of 85.87%, and shows superior performance compared with other state-of-the-art (SOTA) methods. Code will be available online (https://github.com/Jyl199904/DEIF-Mamba).

In recent years, researchers have focused on mitigating the impact of spectral variability (SV) on unmixing performance, leading to the development of various deep-learning-based unmixing networks. Currently, most unmixing networks that account for SV mainly rely on probabilistic generative models, which lacks specific constraints for SV and suffers from the instability of solutions generated by probabilistic models. This results in the generated endmembers or SV components lacking clear physical meaning. To avoid the problems above, we propose an endmember-oriented Transformer network (EOT-Net) that leverages the advantages of endmember bundles to introduce variability while providing stable endmember results with clear physical meaning. We design an endmember-oriented Transformer (EOT) to capture endmember-specific features through directional subspace projection and a low-redundancy attention (LRA) mechanism. Subsequently, the proposed network is divided into two branches: endmember generation and abundance estimation, to process endmember-specific features. In the endmember generation branch, endmember-specific features are transformed into intraclass weights that are used to combine signatures within the bundles, and a set of endmembers is generated for each pixel. In the abundance estimation branch, endmember-specific features are integrated using a heterogeneous information fusion (HIF) module that leverages the spatial distribution heterogeneity of the endmembers, ultimately producing the abundance results. We applied the proposed algorithm to both synthetic and real datasets, and the experimental results demonstrated the model’s superiority.

With the advantages of deep learning (DL) techniques, various infrared (IR) small target detection networks have been proposed. While many networks aim at single-frame detection through supervised learning, ignoring abundant spatial-temporal information and causing heavy labeling costs. In this paper, we develop a 3-D spatial-temporal knowledge aware-based unsupervised network for IR target detection (STUTD). Specifically, we transform IR sequences into 3-D spatial-temporal tensors as data foundation. Based on the designed spatial-temporal Swin Transformer block (ST-STB), we introduce a multiscale feature extraction and aggregation (MFEA) module for effective feature extraction. And a Variational Autoencoder (VAE)-style background reconstruction module with a multihead gating mechanism is designed for background reconstruction. Besides, a designed sparse cardinality selection of residuals performs element-wise filtering on the residuals between the original tensors and the reconstructed background to obtain a pure target tensor. By an unsupervised learning approach, STUTD can achieve IR small target detection. Comprehensive experiments illustrate the superiority of STUTD among state-of-the-art methods. It can be concluded that STUTD has satisfactory overall performance and real-time performance.

The need for precise and efficient monitoring of the Earth’s surface has become increasingly urgent. However, existing methods primarily rely on Siamese-based neural network, which both restrict the extraction of change information and reduce computational efficiency. To address these issues, we propose a One-Stream neural network (OSMT) based on Multi-level feature Fusion for hyperspectral image (HSI) change detection (CD). The proposed method leverages the multi-level semantic change features of the bi-temporal images. First, a band-wise image fusion strategy is utilized to obtain the fused difference image of the bi-temporal HSIs. Then, a residual connection block with spatial-spectral attention mechanism is employed to extract multi-level change features. Next, a multi-level feature fusion module composed of Transformer encoders is used to fuse the change features of different levels, including spatial details at the lower level and semantic information at the higher level. Finally, a classifier is employed to predict the detection results. Experimental results on one public dataset validate the effectiveness of our proposed method.

Noise corruption commonly exists in hyperspectral images (HSIs) and severely affects the accuracy of hyperspectral un-mixing algorithms. The noise formulation of HSIs is relatively complex and would change in conjunction with different devices and imaging settings. For real applications, applying denoising approaches without accurate close-to-reality noise modeling before unmixing may not improve, but rather degrade the unmixing performance. This study formulates a close-to-reality noise model and proposes a learning-based hyperspectral image denoising method for hyperspectral un-mixing. In the experiments, several widely used unmixing algorithms were employed to verify the effect of the proposed method. The experimental results on both synthetic and real demonstrated that our proposed method can handle HSI data with various gain settings and helps to improve the unmixing performance effectively.

Hyperspectral unmixing is a crucial step in hyperspectral image processing. Hyperspectral images in real scenes are saturated with spectral variability, and unmixing performance is limited. We propose the Spectral Variability Attention Net (SVA-Net). We have separately designed a Complementary Feature Enhancement Module (CFE) and a Spectral Variability Attention Mechanism to capture both the original material features in the image and other easily overlooked features. In addition, we design the improved mixing model based on augmented linear mixing model (ALMM) to better cope with the effects of spectral variability. Experiments on real datasets demonstrate the effectiveness of our model.

Recently, convolution neural networks (CNNs) have flourished in hyperspectral image (HSI) change detection (CD). However, these approaches typically rely on single-scale and single-level features to obtain change information, limiting the further boost detection accuracy. To solve the above problem, we propose a novel multiscale and multilevel spatial–spectral features aggregated neural network ($\text{M}^{2}\text{S}^{2}$Net) for HSI CD. The proposed method leverages the multiple spatial–spectral (SS) features of bi-temporal images to mine temporal change information. First, the HSIs are cropped into patches with different spatial and spectral sizes to provide various scales SS information and alleviate the problem of spectral redundancy. Then, the patches are utilized to capture the multiscale SS features in the residual connection block (ResBlock). These bi-temporal features are inputted into the Transformer encoder-based feature fusion module to learn the discriminative change representations. Via the global receive field of the self-attention mechanism, various fine-grained change information is learned from the features at different scales. Finally, the change features of each level are aggregated to produce the change map. The experiments on two public HSI datasets verify the effectiveness of the proposed method. Specifically, the overall accuracy (OA) of the$\text{M}^{2}\text{S}^{2}$Net surpasses the second-best method by 0.56% and 0.10% on the Farmland and River datasets, respectively.

Models based on low-rank and sparse decomposition (LRaSD) have been rapidly developed in the hyperspectral anomaly detection (HAD) task. However, traditional LRaSD models usually impose multiple complex regularizers on the background components, which inevitably increases the computational cost and fails to maximize the effectiveness between them. In addition, regularizers for abnormal components mostly penalize each pixel with the same intensity. To tackle these challenges, we propose a model based on enhanced 3-D total variation (TV) and sparse reweighted regularization, referred to as E-3DTVSR. Specifically, an enhanced 3-D TV (E-3DTV) regularization is adopted to simultaneously characterize the low-rank and piecewise smoothness of the background. Since E-3DTV applies sparse regularization to subspace base maps of gradient maps along all bands of a hyperspectral image (HSI), rather than the gradient maps themselves, this effectively removes noise and improves the detection efficiency of the model. Meanwhile, in order to enhance the sparsity of abnormal targets and distinguish sparse nonabnormal pixels, combined with the log-sum function and the reweighted$\ell _{1}$minimization strategy, a sparse reweighted regularization is designed to adaptively assign weight to each target. Experiments demonstrate that E-3DTVSR reaches the area under the ROC curve (AUC) scores of 99.86%, 99.59%, and 99.72% on three public HSI datasets, respectively, outperforming the current advanced approaches.

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 image change detection (HSI CD) refers to identifying and analyzing differences between two HSIs acquired in the same area but at different times. However, current deep learning (DL)-based methods have limitations in fully exploiting the change information between bitemporal images and utilizing multilevel features. To address these issues, we propose a novel Multi-level features Fused and Change information Enhanced neural Network (MFCEN) for HSI CD. The proposed MFCEN method leverages the hierarchical low- and high-level features of bitemporal images, allowing for the direct capture and enhancement of change features. First, a Siamese-based network is employed to extract multilevel features from the bitemporal images, including low-level spatial details and high-level semantic features. Within the temporal change information branch (TCIB) at each level, the change features are reinforced by the semantic features of each image, and the change features act as a guiding force to direct the feature extraction of each bitemporal image to focus more on the change region. Next, the enhanced change features of each level are fed into the multilevel features fusion module (MFFM) to aggregate the fine-grained details and high-level semantics. Finally, the fused features, enriched with multilevel change information, are utilized for CD. The experiments demonstrate that our proposed MFCEN outperforms existing methods on three public datasets. The code will be available athttps://github.com/Ding201901/MFCEN.

Infrared (IR) small target detection plays an essential role in many civilian and military fields. And low-rank and sparse decomposition-based detection techniques have gradually shown superiority. However, there are two crucial issues. The first is accurate low-rank background estimation, the second is effective target enhancement and background suppression. To address the abovementioned issues, we propose a method named feedback spatial-temporal infrared small target detection with framelet- and Log-based improved tensor nuclear norm (FST-FLNN). Specifically, we transform the current frame to be detected and its adjacent frames into a spatial-temporal tensorD, followed by projectingDonto the principal component analysis (PCA)-driven orthogonal complement subspace to achieveD⊥VPCAmfor preliminarily eliminating principal background components. To further enhance the low-rank property of residual background components and estimate it more accurately, we fully utilize the spatial-temporal information of the background and the corresponding core to improve tensor nuclear norm using Log operation in framelet domain, which is called framelet-and Log-based improved tensor nuclear norm (FL-ITNN). In addition, we introduce a posterior knowledge extraction method based on extended 3-D morphology as feedback mechanisms towards the model. According to the different feedback objects of posteriors, we provide two different feedback mechanisms and their corresponding LRSD-based target detection models FST-FLNN. Finally, two efficient ADMM-based solving frameworks are designed. Compared with nine state-of-the-art competitive methods, comprehensive qualitative and quantitative experiments and analysis on five real complex IR sequences illustrate the satisfactory target detectability (TD), background suppressibility (BS) and overall performance of the two versions of FST-FLNN.

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.

In various civilian and military applications, infrared (IR) small target detection is faced with difficulties because of complex backgrounds, low signal-to-clutter ratio, and clutter interference. Although low-rank and sparse theories based on tensor analysis have been widely employed, three crucial issues persist. The first concerns accurate estimation of target and background, the second involves the development of a more comprehensive target detection model, and the last one is real-time detection performance. This article proposes an IR small target detection method named clustering and tracking-guided infrared small target spatial-temporal prediction completion model (CTSTC). Specifically, a 3-D spatial-temporal tensor is constructed in the high-frequency domain, incorporating target detection priors and subsequent yet-to-be-detected IR frames as foundational data for the prediction completion model. Secondly, we improve K-means clustering algorithm for multiple low-rank background clusters, followed by designing an IKC-derived rank surrogate, resulting in more accurate low-rank background estimation. Furthermore, a Bayesian-inference-derived tracking regularization is incorporated into the completion model to describe the motions and state transitions of targets, thereby improving the robustness of interferences. Meanwhile, an efficient ADMM-based optimization scheme is designed for solving the completion model. Additionally, spatial-temporal evolution-based fusion strategies are proposed, which utilize the past and future spatial-temporal knowledge for the assistance of the detection of current frames and enhance both TD and BS of the prediction completion model. Compared with ten state-of-the-art competitive methods, extensive experiments on five practical datasets have illustrated the superiority of CTSTC in terms of target detectability (TD), background suppressibility (BS), and overall performance.

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.

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.

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.

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.

In recent years, tensor representation (TR) based hyperspectral anomaly detection approaches have attracted more and more attention. However, two urgent issues still need to be addressed: 1) existing tensor decomposition approaches for hyperspectral anomaly detection (HAD) cannot make full use of the spectral-spatial correlation of background components in hyperspectral images (HSIs); 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 multi-scale entanglement renormalization ansatz (MERA) tensor network, this paper 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 semi-orthogonal and orthogonal factors. Due to the intricate interplay between semi-orthogonal (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.

The large number of bands contained in hyperspectral images (HSIs) can provide a wealth of information on regions of interest but also causes the "curse of dimensionality" and information redundancy. To address this issue, this paper proposes a novel band selection method based on triple constraints and attention network (TCANet-BS) for HSIs. The proposed TCANet-BS can make good use of band representativeness, band informativeness, and inter-band correlation when selecting bands. Specifically, TCANet-BS calculates band representativeness by taking advantage of the attention reconstruction network. Moreover, TCANet-BS obtains band informativeness and inter-band correlation through spectral information divergence and orthogonal subspace projection technique, respectively. Experimental results on the real-world hyperspectral dataset show that TCANet-BS can effectively improve classification accuracy compared with other advanced band selection methods.

Due to the poor quality image with low signal-to-clutter ratio (SCR), infrared (IR) small target detection is faced with great challenges in the remote sensing field. Despite the fact that the local contrast measure (LCM) has been widely applied for IR target detection, the existing LCM-based methods suffer from weak target detectability (TD) or background suppressibility (BS) in complicate background. In this paper, we propose a novel IR small target detection method based on improved tri-layer window local contrast measure (TrLCM). With an additional isolation circle in TrLCM, the influence of background on target detection is reduced to a certain extent. Besides, background suppressibility is also promoted through a designed adaptive adjustment coefficient. Comprehensive experiments and analysis on three datasets verify that the proposed TrLCM achieves advanced TD, BS and overall performance.

The common purpose of HTD is to distinguish the target from the BKG in HSIs using either fully-known or partly-known prior knowledge. Traditional HTD methods generally rely on distance or similarity measurement and statistical techniques to identify discriminating characteristics. Inspired by the simplicity and efficiency of AE in feature extraction and dimension reduction, and considering the susceptibility of limited prior knowledge to spectral variability, a novel HTD method based on hidden dimension-restricted iterative AE guided by posterior knowledge is proposed. The latent space mapping from the original data is constrained with the concept of VD by the algorithm of NWHFC, which allows the reconstruction more tendentious to target and purified BKG. The posterior knowledge, obtained by CEM on a subpixel level, is iteratively added to the input of AE to improve the diversity of the target and BKG collaboratively. The results show that the BKG is effectively suppressed and the target is prominently highlighted layer by layer. Comprehensive experiments and analysis conducted on the public dataset demonstrate that the proposed method is structurally simple yet efficient, outperforming five state-of-the-art algorithms rank by 3-D ROC. The key indicator, target detectability, exceeds the comparison methods by 51.1%.

Band selection (BS) is an effective approach to alleviate the spectral redundancy of a hyperspectral image (HSI). The emerging deep-learning-based BS methods have become a hot topic due to their ability to model nonlinear relationships between spectral bands. However, existing deep-learning-based BS methods fail to accurately extract the representativeness of each band as a result of the limitation of interpretation networks. Moreover, existing deep-learning methods cannot fully utilize the interband correlation and the spatial information of HSIs for BS. To solve these issues, in this letter, we propose a novel Transformer reconstruction network for unsupervised BS, termed BSFormer. Specifically, the Transformer reconstruction network, which contributes to leveraging the spectral–spatial information of the HSI, consists of a Transformer-based band attention (TBA) module and a convolutional autoencoder (CAE)-based reconstruction module. On this basis, we design a novel band evaluation criterion composed of representative metric and redundancy metric, which are interpreted with the help of the multihead self-attention layer in the TBA module. The designed criterion can fully use the band representativeness and interband correlation for BS. Experimental results on three well-known hyperspectral datasets verify that the proposed BSFormer can yield better classification performance than the competitors.

Nowadays, some tensor-based hyperspectral anomaly detection (HAD) approaches are still insufficient in utilizing the spatial-spectral structure information of hyperspectral images (HSIs), resulting in the inability to isolate the background and abnormal targets well. In this letter, an enhanced tensor low-rank representation learning (ETLR) model is proposed for HAD. Specifically, the original 3-D hyperspectral image (HSI) data is firstly decomposed into a structural background component, an anomaly component and a noise component. Among them, with the help of multi-subspace learning technology, the structural background component is reformulated by the t-product of the background dictionary tensor and the corresponding coefficient tensor. Then, tensor nuclear norm (TNN) is adopted to preserve the global low-rank property of the background component in both spatial and spectral dimensions. For the abnormal component, an ℓ2,1,1-norm is designed to enhance the group sparsity of abnormal pixels. For the noise component, a tensorF-norm constraint is imposed to suppress the confusion of noise and anomalies. Meanwhile, a robust dictionary tensor that can adequately characterize the background is constructed by using tensor robust principal component analysis (TRPCA). Furthermore, to reduce the interference of redundant information on detection accuracy, the optimal clustering framework (OCF) method is utilized for band selection. Finally, extensive experiments on one simulated and three real HSI datasets confirm that our algorithm is superior than current HAD algorithms.

Recently, hyperspectral anomaly detection (HAD) methods based on tensor low-rank representation (TLRR) have received widespread attention. However, most of them tend to emphasize the utilization of multiple types of prior knowledge to characterize background components, while the prior information about anomaly components is limited. Additionally, the constructed background dictionary is also susceptible to noise and outliers. To address these challenges, this paper focuses on both the background and abnormal components, proposing a robust tensor low-rank sparse representation with saliency prior (RTLSR-SP) method for HAD. Specifically, for the background component described by the dictionary tensor and the corresponding coefficient tensor, tensor nuclear norm (TNN) constraint and sparsity constraint are imposed on the coefficient tensor simultaneously to capture the global and local spatial-spectral structure information of the hyperspectral image (HSI), respectively. For the anomalous component, we design a sparse saliency prior weight tensor to enhance the saliency of anomalous targets. Meanwhile, the tensor ℓF,1-norm is also integrated into the model to better separate abnormal targets from the background. Furthermore, combining tensor robust principal component analysis (TRPCA) and skinny tensor singular value decomposition (skinny t-SVD), a robust background dictionary is constructed. Finally, an efficient iterative algorithm based on the alternating direction method of multipliers (ADMM) is derived to optimize the RTLSR-SP model. Comprehensive experimental findings on one simulated dataset and six real hyperspectral datasets demonstrate the effectiveness and superiority of the proposed algorithm compared with eight state-of-the-art HAD algorithms.

Low rank and sparse representation (LRaSR)-based approaches have been widely used for hyperspectral anomaly detection (HAD) by effectively decomposing a hyperspectral data set into a low - rank component for background (BKG), a sparse component for anomalies, and a residual component for noise. This paper proposed a novel hyperspectral anomaly detection (AD) method based on data sphering (DS) and unsupervised target detection with sparse cardinality (DS-UTSSC). First of all, DS-UTS uses data sphering to remove BKG, which is characterized by 1 st and 2ndstatistics, for original data$X$. Second, potential anomaly component$\mathrm{S}_{\text{DS}-\mathrm{U}\text{TS}}$is generated via unsupervised target detection and subspace projection for the sphered data$X$. To further reduce the impact of noise on anomaly detection, sparse cardinality (SC) is incorporated to obtain$\mathrm{S}_{\text{DS}-\text{UTSSC}}$. Finally, RX-AD is implemented on$\mathrm{S}_{\mathrm{D}\text{S-UTSSC}}$to detect anomalies. The experimental results validate that DS-UTSSC is very competitive against the LRaSR-based models and AE-based method.

Hyperspectral images (HSIs) have been widely used in remote sensing change detection for environment monitoring and urban studies. Aiming to leverage the rich information of HSIs, we proposed a joint spatial-spectral-temporal attention method for HS image change detection in this paper. To this end, we put together the convolutional block attention module (CBAM) and the recurrent neural network (RNN) to the end-to-end network. The CBAM can get spatial and spectral feature representation, and the latter can use the temporal information by a long short-term memory. Experiments on one HSIs change detection dataset demonstrated that our proposed method could get effective performance in HSIs for change detection.

Hyperspectral image (HSI) change detection (CD) has gained much attention in remote sensing. However, most deep learning methods are restricted by a limited receptive field, without leveraging temporal information, and the need for many training samples. In this letter, we proposed a Transformer Encoder-based HSI CD framework called CDFormer. First, space and time encodings are added to the pixel sequence to guide transformers to exploit change information of space and time by the pixel embedding (PE) module. Second, the self-attention component of Transformer Encoder module has a global space-time receptive field to mine the correlation and interaction between bi-temporal features, enhancing the utilization of temporal dependencies. Next, the multi-head attention mechanism learns several attentions and extracts the joint weighted spatial-spectral-temporal features, which improves the feature discrimination ability of the changes. Finally, the detection result is predicted using a fully connected network. It is notable to mention that the proposed method only uses a few labeled samples to train the network. Experiments on two HSI datasets demonstrate that our proposed method can get effective performance in HSI CD.

Spatial information can play a supporting role in spectral unmixing. In this letter, we propose a dual branch autoencoder network to incorporate spatial-contextual information for spectral-spatial unmixing. The two branches leverage different architectures to efficiently extract spatial information and spectral information. In the first branch, we use fully connected layers to extract spectral information, where the neuron in each layer can capture all spectral features. In the second branch, 2-D convolution is adopted to exploit spatial features, which does not require hand-crafted assumptions compared with conventional methods. Then the extracted features are concatenated and propagated to generate the abundance and reconstruct the pixel. Moreover, to solve the drawbacks of the existing reconstruction functions, we propose a new function termed squared sine distance to improve the convergence quality of the proposed network. Experimental results reveal the effectiveness of our proposed method on both synthetic data and real-world data.

With the development of deep learning, the benchmark of hyperspectral imagery classification is constantly improving, but there are still significant challenges for hyperspectral imagery classification of few-shot scenes. This letter proposes an active-learning-based prototypical network (ALPN), which uses the prototypical network to extract representative features from a few samples. Moreover, it combines semisupervised clustering and active learning methods to select and request labels from valuable examples actively. In this way, the feature extraction ability of the network is gradually optimized. The experimental results validated that the classification accuracy and robustness of ALPN significant exceeded the comparison baselines. Furthermore, because it can be regarded as a sample selection method, ALPN can be easily combined with other models to obtain better classification results.

With the development of deep learning, hyperspectral image classification (HSIC) has improved rapidly in recent years. Unsupervised feature learning algorithms play an important role in extracting features from hyperspectral images (HSIs). This letter proposed a random-occlusion-based Bootstrap-Your-Own-Latent network (ROBYOL), combining a new augmentation method and a superior contrastive learning algorithm for feature extraction. The proposed method consists of a self-supervised learning part for feature extracting and a classifier part as the downstream task. It can be proved by the experimental results that the feature extraction ability of the network is effective in this way. Furthermore, the influence of different occlusion strategies is also studied, including changing occlusion area and occlusion value, and we proposed translucent occlusion. Our results with two well-known HSIs reveal that proper occlusion strategies can improve hyperspectral classification results effectively.

Band selection (BS) is an effective means to solve the problems of spectral redundancy and Hughes phenomenon in hyperspectral images (HSIs). However, existing BS methods fail to take into account the representativeness, redundancy, and information content of the selected bands simultaneously, and most of them lack consideration of the inherent nonlinear relationship between bands. To address these problems, we propose a novel unsupervised BS framework that can comprehensively consider band representativeness, redundancy, and information content (RRI) in this letter. The band representativeness is estimated by a three-dimensional convolutional autoencoder, which can capture the inherent nonlinear relationship between the bands and leverage the spatial information of the HSI. The redundancy and the information content of a band subset are restricted and enhanced by the correlation coefficient and the information divergence, respectively. Subsequently, RRI combines these three indicators as the subset evaluation criterion and utilizes immune clone selection algorithm to search for the desired band subset. Experimental results verify that the proposed RRI method can provide higher classification accuracy than the competitors and is robust to noisy bands.

The technique of detecting multiple dim and small targets with low signal-to-clutter ratios (SCR) is essential for infrared search and tracking systems. In this letter, we establish a multiple small targets detection method derived from hierarchical maximal entropy random walk (HMERW). The HMERW revolves the limitation of strong bias to the most salient target of the primal maximal entropy random walk (MERW) based on a proposed graph decomposition theory. To enhance the characteristics of small targets and suppress strong clutters, we design a specific weight matrix for HMERW instead of using the conventional weight matrix in MERW. First, a stationary distribution map is obtained by importing the filtered infrared image into the HMERW. Second, a coefficient map is constructed based on the designed weight matrix to fuse the stationary distribution map. Then, an adaptive threshold is used to segment multiple small targets from the fusion map. Extensive experiments on practical datasets demonstrate that the proposed method is superior to the state-of-the-art methods in terms of target enhancement, background suppression, and multiple small targets detection.

Although remote sensing image registration has been studied for several years, achieving accurate image registration remains a challenging task due to the complicated conditions surrounding remote sensing images. To improve the accuracy and robustness of image registration, we proposed a registration method based on the global mixed structure similarity (GMSIM) measure. This measure mixes the structure similarity in both the frequency domain and the intensity domain because phase-based structure similarity in the frequency domain is sensitive to intensity contrast and spatial translation, and gray-based structure similarity in the intensity domain is efficient to structure change. Feature-based registration methods are used to generate the initial registration parameters. After that, we calculate the final registration parameters by maximizing GMSIM. Quantum-behaved particle swarm optimization (QPSO) is utilized to solve the optimal results of GMSIM due to its high efficiency. The proposed method has been evaluated on several remote sensing images differing in scale, gray, and scene and compared with three state-of-the-art registration methods. Experimental results demonstrate the high accuracy of the proposed scheme.

Low-rank and sparse representation (LRaSR)-based approaches have been widely used for anomaly detection (AD). Their central ideas are to minimize the rank of the low-rank space constrained to predetermined values, while using various regularization parameters to control the sparse representation. Three key issues arise from LRaSR. The first is how to determine the constrained rank. The second is an appropriate selection of regularization parameters. The third one is the detector used for AD. This article presents a new but rather simple competing model, called component decomposition analysis (CDA) which represents a data space X as a linear orthogonal decomposition of three components, X = PC$^{{m}} +$IC$^{{j}} +$N with${m}$principal components, PC$^{{m}}$, generated by principal component analysis (PCA) and${j}$independent components, IC$^{{j}}$, generated by independent component analysis (ICA) plus a noise component N. CDA offers several advantages over LRaSR. First, CDA uses well-known component analysis techniques to decompose the dataset without solving constrained optimization problems. Second, the values of${m}$and${j}$can be automatically determined by virtual dimensionality (VD) and a minimax-singular value decomposition (MX-SVD). To better extract anomalies from the IC$^{{j}}$component space, the concept of sparsity cardinality (SC) is further incorporated into CDA to derive a CDASC anomaly detector (CDASC-AD). The experimental results demonstrate that CDASC-AD is very competitive against the LRaSR-based models and performs well in hyperspectral AD.

Despite that many state-of-the-art methods have been proposed for infrared (IR) small target detection, target detectability (TD) and background suppressibility (BS) cannot be significantly improved simultaneously, especially in complex situations. This article proposes a novel IR small target detection method named improved multimode nuclear norm joint local weighted entropy contrast (IMNN-LWEC), which represents the IR target detection task as an optimization problem for tensor decomposition of three components in the spatial–temporal domain, including background tensor, target tensor, and sparse structure tensor. First, to utilize the spatial and temporal information in an IR sequence effectively, we transform the original IR sequence into a nonoverlapping spatial–temporal patch tensor. Second, a nonconvex approximation of tensor rank called improved multimode weighted tensor nuclear norm (IMWTNN) is proposed to estimate background tensor rank, which is of benefit to separate the background component more completely from the original image. Third, based on the structure tensor theory, we introduce a new sparse prior map called LWEC via a designed image entropy operator and a new prior information filter, which can further preserve the target and suppress the background simultaneously. Besides, a novel tubewise sparse regularization term is designed to identify linear sparse structures. The Frobenius norm is used to characterize noise. Finally, to solve the proposed model, an efficient optimization scheme utilizing the alternating direction method of multipliers (ADMM) is designed to retrieve the small targets. Comprehensive experiments on five datasets witness the superior TD and BS performance of the proposed method compared with nine state-of-the-art detection methods.

Autoencoder is an efficient technique for unsupervised feature learning, which can be applied to hyperspectral unmixing. In this letter, we present an autoencoder network with adaptive abundance smoothing (AAS) to solve the challenges of previous techniques. Specifically, the proposed method uses a multilayer encoder to obtain the abundance and a single-layer decoder to reconstruct the image. The AAS algorithm tackles the outliers by exploiting the spatial-contextual information and can be adaptive for each pixel. Moreover, the softmax function is used as the encoder output function with the help of L1/2regularization to produce sparse output. Experimental results of the synthetic and real data reveal the superior performance of the proposed method against other competitors.

In recent years, low-rank representation (LRR) has attracted considerable attention in the field of hyperspectral anomaly detection. The main objective of LRR-based methods is to extract anomalies from the complex background. However, the presence of anomalies in the background dictionary can lower the detection performance. In this letter, a novel method is proposed for hyperspectral anomaly detection based on the LRR model. This method facilitates the discrimination between the anomalous targets and background by utilizing a novel dictionary and an adaptive filter based on the local outlier factor (LOF). In order to exclude the potential anomalies from the dictionary, the ranking of LOF scores for each pixel is adapted to select the potential background pixels as dictionary atoms. A filter that explores the intrinsic spatial structure is designed to enhance the differences between the anomalies and the background pixels. The experimental results that conducted on three real-world data sets demonstrate that the proposed method achieves a better performance than several state-of-the-art hyperspectral anomaly detection methods.

Due to significant geometric distortions and illumination differences, developing techniques for high precision and robust multisource remote sensing image registration poses a great challenge. This article presents an iterative image registration approach, called iterative scale-invariant feature transform (ISIFT) for remote sensing images, which extends the traditional scale-invariant feature transform (SIFT)-based registration system to a close-feedback SIFT system that includes a rectification feedback loop to update rectified parameters in an iterative manner. Its key idea uses consistent feature point sets obtained by maximum similarity to calculate new alignment parameters to rectify the current sensed image and the resulting rectified sensed image is then fed back to update and replace the current sensed image as a new sensed image to reimplement SIFT for next iteration. The same process is repeated iteratively until an automatic stopping rule is satisfied. To evaluate the performance of ISIFT, both the simulated and real images are used for experiments for the validation of ISIFT. In addition, several data sets are particularly designed to conduct a comparative study and analysis with existing state-of-the-art methods. Furthermore, experiments with different rotation are also performed to verify the adaptability of ISIFT under different rotation distortions. The experimental results demonstrate that ISIFT improves performance and produces better registration accuracy than traditional SIFT-based methods and existing state-of-the-art methods.

Hyperspectral unmixing (HU) is a crucial technique for exploiting remotely sensed hyperspectral data, which aims at estimating a set of spectral signatures, called endmembers and their corresponding proportions, called abundances. The performance of HU is often seriously degraded by various kinds of noise existing in hyperspectral images (HSIs). Most of existing robust HU methods are based on the assumption that noise or outlier only exists in one kind of formulation, e.g., band noise or pixel noise. However, in real-world applications, HSIs are unavoidably corrupted by noisy bands and noisy pixels simultaneously, which require robust HU in both the spatial dimension and spectral dimension. Meanwhile, the sparsity of abundances is an inherent property of HSIs and different regions in an HSI may possess various sparsity levels across locations. This article proposes a correntropy-based spatial-spectral robust sparsity-regularized unmixing model to achieve 2-D robustness and adaptive weighted sparsity constraint for abundances simultaneously. The updated rules of the proposed model are efficient to be implemented and carried out by a half-quadratic technique. The experimental results obtained by both synthetic and real hyperspectral data demonstrate the superiority of the proposed method compared to the state-of-the-art methods.

The orthogonal-projection-based band selection (OPBS) algorithm is one of the newly proposed band selection methods. In this letter, we present a nonlinear version of the OPBS method, which is denoted as the Kernel-OPBS method. The OPBS method selects the desired bands one by one, and in each round of lookup, it chooses the band that has the maximum distance to the hyperplane spanned by the currently selected bands. Extending this algorithm to a feature space associated with the original input space through a certain nonlinear mapping function can provide a nonlinear version of the OPBS algorithm. Although it is basically intractable to compute the mapped bands due to the high dimensionality of the feature space produced by the nonlinear mapping function, the selection criterion of the Kernel-OPBS method is actually related to only the inner products of the mapped bands; thus, the kernel function can be applied and it is unnecessary to define the nonlinear mapping function. Experimental results on different data sets demonstrate that the selected bands obtained by the Kernel-OPBS method can achieve higher pixel classification performances than that by the OPBS method.

Infrared small target detection is one of the most important parts of infrared search and tracking (IRST) system. Generally, the small and dim target is of low signal-to-noise ratio and buried in the complicated background and heavy noise, which makes it extremely difficult to be detected with low false alarm rates. To solve this problem, we propose a small target detection method based on multiscale local contrast measure. Different from conventional methods, we novelly measure the local contrast from two aspects: local dissimilarity and local brightness difference. First, we present a new dissimilarity measure called the local energy factor (LEF) to describe the dissimilarity between the small targets and their surrounding backgrounds. Second, the feature of the brightness difference between the small targets and the backgrounds is utilized. Afterward, the local contrast is measured by taking both features of the above into account. Finally, an adaptive segmentation method is applied to extract the small targets from the backgrounds. Extensive experiments on real test data set demonstrate that our approach outperforms the state-of-the-art approaches.

Four-band multispectral remote sensing imagery is widely used for a variety of purposes and has a long historical data record. However, most of the existing bathymetry inversion methods are either unable to determine the optimal solution through theoretical or semianalytical models due to the limited number of bands, or their application is limited by the difficulty in obtaining in situ data. In this article, a log-ratio model and a semianalytical model are combined to develop a new shallow water depth inversion method (L-S model) using four-band multispectral remote sensing images without the need for supporting truth data. A case study was conducted for Ganquan Island in the South China Sea, using four-band multispectral imagery from the GeoEye-1, WorldView-2 (four bands selected), Sentinel-2 (four bands selected), and Gaofen-1 satellites; a sound range of 30 m was achieved. When compared to LiDAR-measured water depth data, GeoEye-1, WorldView-2, Sentinel-2, and Gaofen-1 data have root-mean-square errors (RMSEs) of 1.33-1.97 m. In addition, compared with the results of the log-ratio model trained using 200 LiDAR-based depth readings, the L-S model results obtained for the four satellite types are similar in the RMSE. These results show that the L-S model can achieve results that are close to those of the log-ratio model without the need for external inputs. This provides a feasible new method for bathymetry inversion in areas without truth data using only four-band imagery.

In the field of deep learning, finetuning the pretrained networks to get a good classifier is a common way of transfer learning. Unlike the traditional way, we insert deformable convolutional layers into the pretrained networks, and finetune the new networks. As a result, we find it performs as well as the normal one in classification, and when we construct a plane detection pipeline based on the two classifiers respectively, the one with deformable convolution shows a better result than the other.

Spectral unmixing is a fundamental issue that needs to be addressed in the application of hyperspectral images. Due to the complex imaging conditions in remote sensing, it is common for the same object to have different spectral signatures. In this paper, we present a novel endmember bundle extraction method based on pixel purity index and superpixel segmentation to deal with this problem, where each material is represented by a set of similar endmember spectra. This method improves the accuracy of endmember extraction and focuses on the removal of redundant endmembers, leading to less spectral unmixing error than existing endmember bundle extraction algorithms. The experimental results on both synthetic dataset and real dataset demonstrate that the proposed method performs effectively in extracting variable endmember sets.

Nonnegative matrix factorization (NMF) and its various robust extensions have been widely applied to hyperspectral unmixing. Most existing robust NMF methods consider that noises only exist in one kind of formulation. However, hyperspectral images (HSI) are unavoidably corrupted by noisy bands and noisy pixels simultaneously in the real application s. This paper presents a robust NMF using ℓ1,2norm and further proposes a two-dimensional robust NMF model by incorporating ℓ2,1norm and ℓ1,2norm, which is robust to noises in both spatial dimension and spectral dimension simultaneously. In addition, the Huber's M-estimator is integrated into the model to achieve better assignations of weights for each pixel and band with various noise intensities, which avoids the singularity problem and effectively improves the unmixing performance. The elegant updating rules of the proposed model are also efficiently learnt and provided. Experiments are conducted on both synthetic and real hyperspectral data sets. The experimental results demonstrate the effectiveness of the proposed methods in unmixing performance.

This paper proposes a new unsupervised band selection method for hyperspectral imagery anomaly detection. Main background subset is identified using global RX detector. An effective criterion of band subset selection is designed based on minimum background variations using differential data of the main background subset. The particle swarm optimization algorithm is used as the search strategy to find the best solution for band selection with the proposed criteria. The proposed method is evaluated by experiments of hyperspectral anomaly detection, and the results show that the performance of anomaly detection is significantly improved after band selection.

Hyperspectral unmixing (HU) is a crucial technique for exploiting remotely sensed hyperspectral data, which aims to estimate a set of spectral signatures, called endmembers and their corresponding proportions, called abundances. Nonnegative matrix factorization (NMF) and its various robust extensions have been widely applied to HU. Most existing robust NMF methods consider that noises only exist in one kind of formulation. However, the hyperspectral images (HSIs) are unavoidably corrupted by noisy bands and noisy pixels simultaneously in the real applications. This paper proposes a novel spectral-spatial robust NMF model by incorporating 12,1 norm and 11,2 norm, which achieves robustness to band noise and pixel noise simultaneously. The Huber's M-estimator is integrated into the proposed model to achieve better assignations of weights for each pixel and band with various noise intensities, which avoids the singularity problem and effectively improves the unmixing performance. The elegant updating rules of the proposed spectral-spatial robust model are also efficiently learned and provided. Experiments are conducted on both synthetic and real hyperspectral data sets. The experimental results demonstrate the effectiveness of the proposed methods in unmixing performance.

Band selection (BS) tries to find a few useful bands to represent the whole hyperspectral image cube. This letter proposes a novel unsupervised BS method based on the band correlation analysis (BCA). The BCA method tries to find a subset of bands that can well represent the whole image data set. To avoid the exhaustive search, the BCA method iteratively adds the band with the good representative ability and low redundancy into the selected band set, until the sufficient quantity of bands has been obtained. The redundancy and the representative ability of one band are computed by its correlation with the currently selected bands and the remaining unselected bands, respectively. Through constructing a correlation matrix of total bands, the BCA method can find the bands that with large amounts of information and low redundancy, which ensures that the selected bands are useful for the further applications like pixels classification. Experimental results on three different data sets demonstrate that the proposed method is very effective and can achieve the best performance among the competitors.

Kernel nonnegative matrix factorization (KNMF) is an extension of NMF designed to capture nonlinear dependence features in data matrix through kernel functions. In KNMF, the size of the kernel matrices is closely associated with the input data matrix, of which the calculation consumes a large amount of memory and computing resource. When applied on large-scale hyperspectral data, KNMF often meets the bottleneck of memory and may cause the overflow of memory. And when dealing with dynamically acquired data, KNMF requires recomputation of the whole data set when newly acquired data arrived, which produces huge memory and computing resource requirements. To reduce the usage of memory and improve the computational efficiency when applying KNMF on large scale and dynamic hyperspectral data, we extend KNMF by introducing partition matrix theory and considering the relationships among dividing blocks. The decomposition results of hyperspectral data are derived from much smaller scale matrices containing the formerly achieved results and the newly data blocks incrementally. In this paper, we propose an incremental KNMF (IKNMF) to reduce the computing requirements for large-scale data in hyperspectral unmixing. An improved IKNMF (IIKNMF) is also proposed to further improve the abundance results of IKNMF. Experiments are conducted on both synthetic and real hyperspectral data sets. The experimental results demonstrate that the proposed methods can effectively save memory resources without degrading the unmixing performance and the proposed IIKNMF can achieve better abundance results than IKNMF and KNMF.

Band selection (BS) is a special case of the feature selection problem, and it tries to remove redundant bands and select a few informative and distinctive bands to represent the whole image cube. The maximum ellipsoid volume (MEV) method regards the band subset with the maximum volume as the optimal band combination. However, the MEV method cannot be directly applied for hyperspectral imagery due to the high dimensionality of the data sets. Therefore, we first combine MEV with the sequential forward search (SFS) and propose a new unsupervised BS method called MEV-SFS. Furthermore, a subtle relationship between the ellipsoid volume of the band set and the orthogonal projections (OPs) of the candidate bands is observed. Based on this relationship, we propose another equivalent method, namely, the OP-based BS (OPBS) method. OPBS is the fast version of MEV-SFS, and it has a better computational efficiency and the potential to determine the number of bands to be selected. We specifically explain the rationality of the MEV-based methods (MEV-SFS and OPBS) and illustrate their theoretical significance and physical meaning from different aspects. Theoretical analysis also demonstrates that OPBS can be regarded as a model or framework for BS, and thus, we further propose a third novel BS method named the OPBS-information divergence (OPBS-ID) method, which is a variant of OPBS. OPBS-ID can achieve a better classification performance than OPBS in many cases. Experimental results on different hyperspectral data sets demonstrate that the proposed methods have high computational efficiency, and the selected bands can achieve satisfactory classification performances.

This paper proposed a novel local feature descriptor for multisource remote sensing image matching that is robust to significant geometric and illumination differences. In the proposed registration method, traditional SIFT algorithm is applied for local feature extraction and a novel descriptor, named local order pattern based self-similarity descriptor, LOPSS descriptor, is constructed for each extracted feature point. Then, a matching process followed by a reliable outlier removal procedure is implemented for feature matching and mismatch elimination. Finally, registration parameters are estimated by least square method in the affine transformation. The proposed method is applied for matching multisource remote sensing image pairs and the results verify its robustness and discriminability.

Hyperspectral unmixing aims to estimate a set of endmembers and their corresponding percentages in pixels. NMF and its extensions with various constraints have been widely applied to hyperspectral unmixing. L1/2 regularizer and L2 regularizer can be added into NMF to enforce sparseness and smoothness respectively. In practice, an rigion in hyperspectral image may possesses different sparse level across locations. It remains a problem how to impose constraints accordingly when the level of sparse varies. We propose a novel nonnegative matrix factorization with data-guided constraints (DGC-NMF). The DGC-NMF assigns sparseness or smoothness constraints on abundance of each pixel individually according to their mixed level. Experiments on the synthetic data validate the proposed algorithm.

This paper proposes a novel Bayesian Lasso model with spatially related sparsity for hyperspectral linear unmixing. Based on the sparsity hypothesis and the spatial correlation between pixels, we introduce Lasso penalty and spatially constrained prior distributions to the original Beyesian model. This prior information takes into consideration that different regions in the hyperspectral possess various sparse level and the pixels in a same region tends to share a similar mixed level. We assign relatively slighter constraint to the transition areas which generally have higher mixed levels, by measuring spatial correlation adapting to the prior of abundance vector. Empirical experiments show attractive results of the proposed method via extensive simulation studies and comparisons with other algorithms.

The minimization problem of reconstruction error over large hyperspectral image data is one of the most important problems in unsupervised hyperspectral unmixing. A variety of algorithms based on nonnegative matrix factorization (NMF) have been proposed in the literature to solve this minimization problem. One popular optimization method for NMF is the projected gradient descent (PGD). However, as the algorithm must compute the full gradient on the entire dataset at every iteration, the PGD suffers from high computational cost in the large-scale real hyperspectral image. In this paper, we try to alleviate this problem by introducing a mini-batch gradient descent based algorithm, which has been widely used in large-scale machine learning. In our method, the endmember can be updated pixel set by pixel set while abundance can be updated band set by band set. Thus, the computational cost is lowered to a certain extent. The performance of the proposed algorithm is quantified in the experiment on synthetic and real data.

High-precision and robust matching of multi-source remote sensing image matching is not easy to achieve because of non-linear intensity differences and significant geometric distortions. A new registration method is proposed by integrating the scale-invariant feature transform (SIFT) and optimization of local self-similarity mutual information (LSS_MI). This method consists of two main steps. In the first step, SIFT approach with a reliable outlier removal procedure is implemented. By repeatedly fine turning several selected matched feature point coordination, a series of registration parameters are estimated by least square method and used to construct initial particle swarms. Then, a local self-similarity descriptor (LSS) is computed for pre-matching image pairs and the optimal match parameters are obtained by optimizing LSS_MI based on QPSO. The experimental results verify that the LSS-MI is more robust and accurate than regional mutual information in multi-source remote sensing image.

In this paper, we proposed an incremental kernel non-negative matrix factorization (IKNMF) to reduce the computing scale in hyperspectral unmixing. Kernel non-negative matrix factorization (KNMF) is an extended non-negative matrix factorization (NMF) able to capture nonlinear dependency features in data matrix through kernel functions. In KNMF algorithm, the size of kernel matrices is closely associated with the input data scale. To reduce calculation and storage of large matrices, we extend KNMF by introducing partition matrix theory. The decomposition results of data matrices are derived from smaller scale matrices incrementally. Experiments are conducted on synthetic hyperspectral images with multiple sizes, and the experimental results show that the proposed algorithm have effect in saving calculation and memory resource without degrading the unmixing performance.

Most nonlinear unmixing algorithms are based on the nonlinear mixing models with different forms. This paper focuses on the well-known generalized bilinear model (GBM). Though the GBM has shown interesting and promising for nonlinear unmixing, currently almost all the GBM-based unmixing algorithms are supervised. That is, the endmembers must be assumed known in advance. This paper develops an unsupervised nonlinear unmixing method based on the GBM, which can obtain the endmember, abundances and nonlinearity coefficients simultaneously. In the proposed method, the projected-gradient (PG) algorithm are utilized to alternately solve two nonnegative matrix factorization problems. The former updates the endmembers while the latter updates the abundances as well as the nonlinearity coefficients. Experimental results show that the proposed algorithm provide good performance in term of both endmember estimation and abundances estimation comparing with other state-of-the-art algorithms.

To select a minimal and effective subset from a mass of bands is one key issue in hyperspectral image processing. This paper proposed a novel band selection approach using mutual information and K-L divergence. Mutual information (MI) is usually used to measure the statistical dependence between two random variables and can be used to evaluate the relativity of each band. Firstly all bands are grouped into a number of subsets using mutual information. Then, retain only specified number of bands that has maximum information amount which defined by K-L divergence in every subset by removing the others. At last, the retained bands consist of the final bands collection. Experimental results of real hyperspectral dataset show that the proposed algorithm reduces the dimensionality of the data significantly, as well as keeps a better precision.

Endmember extraction is a key step in the hyperspectral image analysis process. The kernel new simplex growing algorithm (KNSGA), recently developed as a nonlinear alternative to the simplex growing algorithm (SGA), has proven a promising endmember extraction technique. However, KNSGA still suffers from two issues limiting its application. First, its random initialization leads to inconsistency in final results; second, excessive computation is caused by the iterations of a simplex volume calculation. To solve the first issue, the spatial pixel purity index (SPPI) method is used in this study to extract the first endmember, eliminating the initialization dependence. A novel approach tackles the second issue by initially using a modified Cholesky factorization to decompose the volume matrix into triangular matrices, in order to avoid directly computing the determinant tautologically in the simplex volume formula. Theoretical analysis and experiments on both simulated and real spectral data demonstrate that the proposed algorithm significantly reduces computational complexity, and runs faster than the original algorithm.

Nonlinear unmixing, which has attracted considerable interest from researchers and developers, has been successfully applied in many real-world hyperspectral imaging scenarios. Hopfield neural network (HNN) machine learning has already proven successful in solving the linear mixture model; this study utilized an HNN machine learning approach to solve the generalized bilinear model (GBM) optimization problem. Two HNNs were constructed in a successive manner to solve respective seminonnegative matrix factorization problems intended for abundance and nonlinear coefficient estimation. In the proposed HNN-based GBM unmixing method, both HNNs evolve to stable states after a number of iterations to obtain unmixing results related to the states of neurons. In experiments on synthetic data, the proposed method showed more efficient performance in regard to abundance estimation accuracy than other GBM optimization algorithms, especially when given reliable endmember spectra. The proposed method was also applied to real hyperspectral data and still demonstrated notable advantages despite the obvious increase in unmixing difficulty.

Nonlinear spectral mixture analysis is important when the light suffers multiple interactions among distinct materials. Few attempts have been conducted to incorporate spatial information to improve the performance of nonlinear unmixing algorithms. In this letter, local windows are adopted in the preliminary classification map to search the relevant endmembers for each pixel. Virtual endmembers, resulting from the relevant endmembers, represent the multiple-scattering effects in each pixel, and the corresponding abundances are estimated based on a modified bilinear model. Experiments on simulated and real hyperspectral images demonstrate that the proposed method provides a competitive or even better performance over some existing algorithms.

Endmember extraction is an important and challenging step to solve the spectral unmixing problem. Most existing endmember extraction algorithms (EEAs) usually find image pixels as endmembers assuming the presence of pure pixels in an image scene or generate virtual endmembers without pure-pixel assumption. When some prevalent materials have pure-pixel representation and pure pixels of other less prevalent materials are absent in the image, it would be more appropriate to extract the endmembers of both prevalent and less prevalent materials, respectively. Therefore, a novel two-stage EEA is presented in this paper. In the first stage, conventional pure-pixel-based EEAs are applied to generate a candidate pixel set, and then spatial information of the candidate pixels is exploited to determine the endmembers of prevalent materials. In the second stage, given known endmembers of prevalent materials, a modified algorithm based on nonnegative matrix factorization is performed to generate the endmembers of less prevalent materials. The validity of the proposed algorithm is demonstrated by experiments based on synthetic mixtures and a real image scene.