Autoencoders (AEs) have gained widespread application in the field of hyperspectral anomaly detection, largely due to their notable effectiveness in efficiently reconstructing backgrounds within hyperspectral images (HSIs). However, the absence of prior knowledge and constraints imposed by spectral information capacity hinder the accuracy of anomaly detection by allowing AEs to reconstruct both anomalous targets and backgrounds simultaneously. To address this limitation, a spatial–spectral cross-guided masked autoencoder (SSCMAE) has been proposed. The guided mask is generated based on the spectral difference between the anomaly and the background. This mask effectively suppresses the reconstruction of anomalous targets while enhancing the accuracy of background reconstruction. Moreover, a dual-branch structure operates, encompassing spatial and spectral dimensions, effectively capturing the inherent three-dimensional characteristics present in HSIs. Ingeniously designed cross-connection layers within the architecture enhance the spatial and spectral branches' capability of extracting internal spatial and spectral features of images. In order to capture a more comprehensive range of background features, a lightweight three-dimensional convolutional autoencoder is introduced. This addresses the issue of local feature loss during background reconstruction and overcomes the limitations that visual transformers face when learning local image structures. The proposed method has been systematically compared against several advanced methods on six real-world datasets. The results explicitly demonstrate the efficacy and superior performance of the presented SSCMAE approach.

Hyperspectral images are valuable for precise land cover change detection in a consistent area over time. Nevertheless, supervised methods for hyperspectral change detection face limitations due to insufficient labeled samples. Additionally, current deep learning-based approaches often neglect critical feature interactions across temporal, spatial, and spectral domains. To address these challenges, we introduce a semi-supervised model called the HyperMatch-based Cross Temporal and Spatial Interaction Transformer (CTSIT) for hyperspectral change detection. The key contributions of this study are as follows: 1) Introduction of the HyperMatch training schedule, which relies on weak-to-strong consistency, to enhance feature extraction using both labeled and unlabeled data. 2) Contribution of the Cross Temporal Bidirectional Attention (CTBA) module to emphasize bidirectional temporal interactions and resemblances. 3) Introduction of the Spatial and Spectral Attention (SSA) module, which includes the Dense Cross Spatial Attention (DCSA) and Spectral Attention (SA) modules, to capture long-range densely spatial interactions and internal spectral similarities. Extensive comparative experiments conducted on three mainstream hyperspectral change detection datasets confirm the effectiveness and superiority of the proposed HyperMatch-based CTSIT in utilizing both labeled and unlabeled samples. For instance, the training ablation demonstrates that this method significantly outperforms most existing state-of-the-art (SOTA) hyperspectral change detection methods, even with a significantly smaller number of samples.

Hyperspectral images can assist change-detection methods in precisely identifying differences in land cover in the same region at different observation times. However, the difficulty of labeling hyperspectral images restricts the number of training samples for supervised change-detection methods, and there are also complex real influences on hyperspectral images, such as noise and observation directions. Furthermore, current deep-learning-based change-detection methods ignore the feature reusage from receptive fields with different scales and cannot effectively suppress unrelated spatial–spectral dependencies globally. To better handle these issues, a contrastive self-supervised two-domain residual attention network (TRAMNet) with a random augmentation pool is proposed for hyperspectral change detection. The contributions of this article are summarized as follows. (1) To improve the feature extraction from hyperspectral images with random Gaussian noise and directional information, a contrastive learning framework with a random data augmentation pool and a soft contrastive loss function (SCLF) is proposed. (2) The multi-scale feature fusion module (MFF) is provided to achieve feature reusage from different receptive fields. (3) A two-domain residual attention (TRA) block is designed to suppress irrelated change information and extract long-range dependencies from both spectral and spatial domains globally. Extensive experiments were carried out on three real datasets. The results show that the proposed TRAMNet can better initialize the model weights for hyperspectral change-detection task and effectively decrease the need for training samples. The proposed method outperforms most existing hyperspectral change-detection methods.

Change detection methods using hyperspectral remote sensing can precisely identify differences of the same area at different observing times. However, due to massive spectral bands, current change detection methods are vulnerable to unrelatedspectral and spatial information in hyperspectral images with the stagewise calculation of attention maps. Besides, current change methods arrange hidden change features in a random distribution form, which cannot express a class-oriented discrimination in advance. Moreover, existent deep change methods have not fully considered the hierarchical features’ reuse and the fusion of the encoder–decoder framework. To better handle the mentioned existent problems, the parallel spectral–spatial attention network with feature redistribution loss (TFR-PS${}^2$ANet) is proposed. The contributions of this article are summarized as follows: (1) a parallel spectral–spatial attention module (${\mathrm{PS}}^2\mathrm{A}$) is introduced to enhance relevant information and suppress irrelevant information in parallel using spectral and spatial attention maps extracted from the original hyperspectral image patches; (2) the feature redistribution loss function (FRL) is introduced to construct the class-oriented feature distribution, which organizes the change features in advance and improves the discriminative abilities; (3) a two-branch encoder–decoder framework is developed to optimize the hierarchical transfer and change features’ fusion; Extensive experiments were carried out on several real datasets. The results show that the proposed ${\mathrm{PS}}^2\mathrm{A}$ can enhance significant information effectively and the FRL can optimize the class-oriented feature distribution. The proposed method outperforms most existent change detection methods.

Hyperspectral imagery is valuable for accurate detection of land cover changes within a consistent area across time. However, current training paradigms for hyperspectral change detection are usually in supervised form which are not consistent to the heavy labeling cost reality. Moreover, current deep learning methods face limitations due to insufficient temporal dependencies sharing and inadequate densely band-wise spatial position dependencies. To tackle these challenges, we introduce an innovative semi-supervised training paradigm called Conflict HyperMatch and a deep learning model called Temporal Weights-sharing and Band-wise Spatial Interacting Transformer (TWBSIT) for hyperspectral change detection. The key contributions of this study are as follows: 1) Introduction of the Conflict HyperMatch training schedule, which relies on representation discrepancy and weak-to-strong prediction consistency, to improve the sample leveraging of both labeled and unlabeled data. 2) Weights-sharing Interacting Temporal Attention (WITA) Module is contributed to capture shared temporal interactions and resemblances. 3) Band-wise Cross Spatial Attention (BCSA) module is introduced to achieve a more extensive spatial perception of the specific central image patch. Extensive experiments conducted on three real datasets validate the effectiveness of the proposed TWBSIT model based on Conflict HyperMatch in leveraging both labeled and unlabeled samples for hyperspectral change detection. This method notably decreases the requirement for labeled training samples and surpasses the performance of many existing hyperspectral change detection methods.

Abundant hyperspectral remote sensing data provides change detection (CD) technology with the opportunity to accurately distinguish land cover transformations across a consistent area over time. However, most deep learning methods for hyperspectral CD are performed in a fully supervised manner, which requires a large number of pixel-level labeled samples and is highly time-consuming. Therefore, this study combines semi-supervised learning to reduce label acquisition costs. However, there are many challenging areas, such as low contrast regions which increases extra burden to interpret general semantic information from land cover attributes. To address these challenges, we introduce an innovative a multi-view prototypical learning (MVPL) based general Semantic Explaining Temporal- Spatial-Spectral Transformer (SETSST) for hyperspectral CD. 1) First, we design self-view and cross-view prototypical similarity maps according to distinct perturbated views of input samples, which ensures appropriate feature interactions by leveraging the low-density aggregated representation. 2) Second, the semi-supervised pretraining framework based on weak-to-strong prediction alignment and MVPL is designed to simultaneously capture features from both labeled and unlabeled samples. 3) To comprehensively interpret semantic explaining information in real land cover attributes, we propose the Semantic Explaining Change Analysis (SECA) module. This module enhances feature explanation by clustering them into several change types and an unchanged type, while accounting for the influence of distance to the central pixel. Extensive experiments conducted on three real datasets demonstrate the effectiveness of the proposed MVPL-based semi-supervised pretraining framework and the SECA module in leveraging both labeled and unlabeled samples for hyperspectral CD. This method surpasses the performance of many existing hyperspectral change detection approaches and significantly reduces the need for labeled training samples.