In view of the practical value of open datasets of hyperspectral images (HSIs), HSI open-set classification (OSC) has attracted more and more attention. Existing HSI OSC methods are usually based on learning labeled samples to identify unknown classes. However, due to the complex high-dimensional characteristics of HSIs and the limited number of labeled samples, the recognition of unknown classes based only on limited labeled samples often has low and unstable accuracy. To address this problem, we propose a semi-supervised graph constraint dual classifier network (SSGCDCN) that can achieve efficient and stable OSC by learning unknown class features and relationships among samples. First, a dual classifier consisting of a multi-classifier and multiple binary classifiers is constructed, which has the ability to discover the unknown class samples by assigning and enabling pseudo-labels to participate in model training to achieve unknown class feature learning. Then, to improve the classification accuracy of both known and unknown classes, a homogeneous graph constraint is imposed on SSGCDCN to learn the relationship information among samples (including labeled and unlabeled samples). This constraint can bring the features of similar samples closer while pushing apart features of dissimilar samples. Experiments evaluated on three datasets demonstrate that the proposed method can obtain superior OSC performance than other state-of-the-art classification methods.

Hyperspectral image open-set classification has gained increasing attention due to its practical significance. However, existing approaches face two major challenges: (1) poor and unstable classification performance under limited labeled samples, and (2) the lack of end-to-end open-set classification frameworks. To address these issues, we propose an Incremental Multi-Task Contrastive Learning Network (IMTCLN), which integrates four learning tasks to achieve end-to-end open-set classification under few-shot conditions through feature sharing and multi-task collaboration. First, we introduce an expanded class labeling method in the model’s output layer, enabling end-to-end open-set classification. Second, among the four learning tasks, the supervised classification task learns the mapping between known-class samples and their labels using limited labeled data. To enhance classification performance under few-shot conditions, we design a semi-supervised Euclidean contrastive learning task, which improves intra-class compactness and inter-class separability by modeling homogeneous and heterogeneous sample relationships. Additionally, for effective unknown-class recognition, we propose a supervised Mahalanobis contrastive learning task, optimizing the Mahalanobis distance among known classes to identify unknown-class samples. Finally, to further enhance classification stability, we introduce an incremental learning task, which leverages pseudo-labeled unknown-class samples to learn their discriminative features, enabling robust discrimination between known and unknown classes. Extensive experiments on three public datasets demonstrate that IMTCLN significantly outperforms existing methods, particularly under extremely limited labeled samples, showcasing superior open-set classification performance and stability.

Infrared small-target detection (ISTD) in a single frame is an essential, yet challenging task due to its small size of targets, weak energy, and clutter background. Current methods either design complex network architectures to facilitate multilevel information interaction (e.g., DNA-Net and UIU-Net) or introduce structural texture priors to enhance feature discrimination (e.g., SRNet and CSRNet). However, both methods fail to explicitly distinguish or suppress the interference of complex background from infrared small targets, which makes them easy to “get lost” in clutter background with insufficient attention to the targets. In this work, we innovatively propose a novel background-masking approach (denoted as BGM) for ISTD. The proposed BGM aims to force the network to focus exclusively on the target by masking out irrelevant background information, thereby enhancing the network’s ability to detect weak and small infrared targets. Specifically, we present a new ISTD method that leverages a proxy training task with masking, enabling the network to simultaneously predict on both the original input and the masked data, where the background is randomly masked/forgotten. This strategy allows for a better concentration of the model on the shapeless targets rather than the cluttered background. The method is flexible with a simple U-shaped network without complicated manipulation and also computationally efficient without increasing the overall computational burden during inference. Extensive experiments demonstrate that our proposed BGM effectively enhances the detection performance of infrared small targets and achieves 70.8% mean intersection over union (mIoU) on IRSTD-1K. The source code would be available athttps://github.com/ZhihaoMa123/BGM

Existing methods for hyperspectral image classification (HSIC) typically assume a closed-set scenario, where all target types are known, and the classifier can assign only predefined classes to samples (pixels). However, in real remote sensing applications, open-set scenarios are common, where unknown classes exist. To address this problem, we propose a graph-constrained deep multi-task approach for open-set HSIC. Our method tackles the challenge of detecting unknown classes by integrating multiple-class classifiers and multiple binary classifiers. Additionally, to handle the limited labeled samples issue in HSIC, we propose utilizing homogeneous and heterogeneous graphs to constrain the two types of classifiers, thereby improving the accuracy of unknown class detection and known class classification. Experimental results on the Pavia University dataset demonstrate that our proposed method outperforms other closed-set and open-set classification methods significantly.

Nowadays, various types of foundational models have emerged, showcasing remarkable performance across a multitude of downstream tasks. However, in the domain of hyperspectral image classification (HSIC), substantial research is still required to effectively leverage the advantages of foundational models and adapt them to hyperspectral data. Consequently, we propose a HSIC algorithm based on a fixed-parameter SAM encoder. Specifically, the global feature extraction subnetwork integrates global patch information to obtain processed features. Subsequently, the semantic information extraction subnetwork is trained using cross-entropy to extract semantic features of categories, culminating in pixel-level classification. Experiments on two HSI datasets indicate that the proposed method can obtain better classification performance when compared with seven state-of-the-art methods.

Nowadays, foundation models have demonstrated exceptional performance across numerous downstream tasks. However, the effective application of these models to hyperspectral image classification (HSIC) is challenged by the unique characteristics of hyperspectral data, including high dimensionality, high variability, and high spatial structure complexity. Therefore, methods need to be developed, which leverage the advantages of foundation models while addressing these challenges. First, a novel HSIC algorithm based on a frozen-parameter segment anything model (SAM) encoder, called SAGFFNet, is proposed. This framework represents the first attempt to use a frozen SAM encoder for global feature extraction and to use spectral dimension data as prompts, enabling precise global spatial-spectral feature extraction with the aid of spectral information. Second, by introducing the self-adaptive padding mechanism and the global feature extraction subnetwork (GFEsNet), the model is enabled to extract distinctive and discriminative features for each category from hyperspectral data through varying padding sizes, thereby enhancing the feature extraction and generalization capabilities of the foundation model. Subsequently, the spectral feature prompt subnetwork (SFPsNet) is designed to extract spectral feature information from samples of different classes as prompt features, assisting the framework in better understanding the global features extracted by GFEsNet. Finally, the semantic information decoder subnetwork (SIDsNet) is introduced as a semantic information decoder, achieving efficient fusion of global spatial-spectral features and spectral prompt features, which significantly improves classification performance. Experiments conducted on four hyperspectral image datasets show that the proposed method outperforms nine existing approaches in terms of classification accuracy.