Hyperspectral anomaly detection (HAD) aims at effectively separating the anomaly target from the background. The low-rank and sparse matrix decomposition (LRaSMD) technique has shown great potential in HAD tasks. However, some LRaSMD models need to convert the hyperspectral data into a two-dimensional matrix. This cannot well maintain the characteristics of the hyperspectral image (his) in each dimension, thus degenerating its representation capacity. In this context, this article proposes a tensor-based Go decomposition (GODEC) model, called TGODEC. The TGODEC model supports the idea of GODEC, representing the HSI data as a combination of background tensor, anomaly tensor, and noise tensor. In detail, the background tensor is solved by the tensor singular value hard thresholding decomposition. The anomaly tensor is solved by a mapping matrix using the corresponding sparse cardinality. Interestingly, the obtained background and anomaly tensors can also be developed for HAD, thus a TGODEC-based anomaly detector is established, called TGODEC-AD. Specifically, the TGODEC-AD method combines the typical RX-AD and R-AD with the above decomposition result of the TGODEC model and constructs different modal operator detectors. Experimental results on multiple real hyperspectral datasets verify the effectiveness of the TGODEC and TGODEC-AD methods. It means that the proposed TGODEC model can effectively characterize the spatial structural features of HSI. As a result, the pure decomposed components can be obtained, contributing to detecting the anomaly target and suppressing the background better in HAD tasks.

Due to high similarity interference of photovoltaic images, repetitive EL image detection and photovoltaic image stitching pose a great challenge. This article proposes two application algorithms and a registration method, which are repetitive EL image detection and photovoltaic image stitching based on the proposed AKAZE-LATCH+GMS registration method. Considering the existence of duplicate EL images acquired indoors, a repetitive EL image detection based on global feature consistency and local cell's texture similarity is proposed. The AKAZE-LATCH+GMS is used to calculate global consistent feature ratio and dHash is used to quantify local cell's similarity to determine whether it is repetitive or not. To solve the problem of high similarity of outdoor photovoltaic images, which leads to dislocation of stitching, an image stitching algorithm based on geographic information is proposed in this article. Compared with AKAZE+GMS and ORB+GMS, the registration experimental results show that the AKAZE-LATCH+GMS method can obtain more initial matching features and higher correct matching rate. Indoor EL images with geometric and illumination differences are used to verify the proposed detection method, which can effectively eliminate repetitive EL images. Aerial photovoltaic images, acquired from the same or different view, are used to verify the validity of the stitching method, and the photovoltaic images can be stitched correctly compared with the stitching method without geographic information.

Band selection (BS) is a critical topic in hyperspectral image dimensionality reduction, focusing on identifying representative bands that can convey the essential information of the full bands without significant loss. Recently, BS based on multiobjective optimization (MO) have become predominant methods. However, there are still two problems that need to be solved. Firstly, the majority of multiobjective band selection methods predominantly optimize accuracy in classification tasks, neglecting the emphasis on anomaly detection. Secondly, in the process of addressing the combinatorial optimization problem of multiobjective band selection using evolutionary algorithms, insufficient consideration is given to the impact of bands on the recognition capability of anomaly when devising solution strategies and determining optimal solutions. To handle the above-mentioned problems, a novel anomaly-oriented multiobjective optimization band selection (AOMOBS) is developed to better suppress background and identify anomaly. Specifically, for the first problem, by calculating the degree of deviation of band, the noise estimates of band and the degree of redundancy between bands, an unsupervised band selection algorithm for anomaly detection tasks, based on multiobjective optimization, is designed. For the second problem, a multiobjective optimization band similarity sorting strategy for anomaly detection tasks is designed for non-dominated sorting, and for decision makers to choose the most appropriate solution from the trade-off solutions. Experiments conducted on real hyperspectral datasets demonstrate that the algorithm effectively identifies a subset of bands with high representational power for anomaly detection. Moreover, AOMOBS outperforms current state-of-the-art methods in both effectiveness and robustness.

Anomaly detection is susceptible to complex background and interference noise. Local anomaly detection and collaborative representation detection can effectively suppress background. However, they both encounter the problem of window size determination due to unexpectable target size, which will strongly limit their detection performance. On the other hand, some interferences show similar properties with anomalies, namely small, weak and in low probability, which would be wrongly reputed as anomalies and degrade the detection accuracy too. A global information and structure tensor guided collaborative representation method (GSGCRD) is proposed in this paper. Firstly, the global and local information are combined into the collaborative representation to obtain a more valuable background dictionary, preventing the big targets from being missed. Secondly, to address the problem that anomaly targets are susceptible to interference spectra, an improved structure tensor is introduced into the band selection and detection refining. Bands of better spatial representation for anomaly are selected for detecting, and interference spectra are removed from representation residual to refine the detection result. Experiments on the real datasets show that the method performs better than classical and some recently proposed methods in background and interference suppression, and achieves satisfactory anomaly detection precision.

Hyperspectral anomaly detection (HAD) aims to identify targets deviating from normal patterns of background. However, the lack of labeled samples pose significant challenges to the task. Self-supervised deep learning methods have shown promising results in this scenario, but they often fall into partial anomaly representations, causing high false alarm rates. Effectively leveraging task knowledge has become a key point. In this paper, we propose a hyperspectral anomaly detection method based on blind-block masked data augmentation and self-supervised contrastive learning. First, the hyperspectral data is divided into patches, followed by patch classification and data augmentation to generate accurate prior knowledge for the learning model. Second, patch pairs with their pseudo-label and corresponding masks are used to train a lightweight residual network integrating contrastive learning, in order to improve detection accuracy while avoiding high computational cost of pixel- wise reconstruction. Finally, pseudo-label is introduced into the detection process again to further guide the anomaly localization. Experimental results show that the proposed method achieves outstanding performance on various hyperspectral datasets, demonstrating strong background-anomaly separability and superior detection accuracy.

High dimensionality, nonlinearity and complexity of hyperspectral image pose significant challenges to target detection using spectral characteristics. Spatial features play an important beneficial role for the task, which can be extracted and integrated in a decision-level or feature-level fusion pattern. However, pixel-wise feature variability and fusion parameter determination remain difficult to handle. This paper proposes an adaptive spatial-spectral decision fusion framework and a detection method with structural enhancement. In the framework, fusion terms and parameters can be generated automatically. The proposed method further suppresses background and noise effectively while preserving target structural information. First, the relationship between decision fusion and feature fusion is analyzed based on CEM detector and an Adaptive Spatial-Spectral Decision Fusion CEM (ASSDF-CEM) algorithm is developed, allowing automatic determination of fusion terms and coefficients. Second, a background smoothness-constrained regularization is added to the CEM’s objective function, and a detection operator BCCEM is deduced. And then, a Representative Structural feature Extraction Strategy (RSES) is designed to obtain global structural features for both targets and backgrounds. Finally, RSES is combined with BCCEM and ASSDF-CEM, leading to the proposed Adaptive Three-feature Decision Fusion target detection algorithm named A3FDF-BCCEM. Experimental results on four hyperspectral image datasets demonstrate the superiority of the proposed method in target detection.