Poisson noise and Gaussian noise are widely present in high-dimensional data, which poses a challenging problem. Moreover, Poisson noise is signal-dependent and influenced by sensor parameters and the imaging environment; therefore, methods based on the additive Gaussian model are limited in handling such real-world scenarios. Meanwhile, self-supervised learning-based network frameworks can construct labels from noisy images for network training, addressing the issue of matching between noisy and clean images. However, while existing frameworks are very effective in handling Gaussian noise, the presence of a mixture of Poisson noise and Gaussian noise suppresses their performance. To address this issue, we propose a Poisson-Gaussian mixed noise removal framework, SPGM-Net, which is the first method for high-dimensional data Poisson-Gaussian mixture noise removal based on a self-supervised learning framework. First, we propose a self-supervised noise estimation network based on the generalized Anscombe transformation (GAT) noise transformation formula, which transforms noise into approximately Gaussian noise. Second, based on subspace representation theory, we build a series of training samples for self-supervised network training using a neighbor pixel sampler on the transformed noisy eigenimages. No matter in the noise estimation stage or the self-supervised denoising stage, only noisy images are used, which is suitable for real-world scenes. To verify the effectiveness of the proposed method, we test it on various data, including indoor hyperspectral images (HSIs) and real satellite images. The results of both the simulated data and the real data show that the proposed method is superior to state-of-the-art methods.

Detecting small air target is an important task in civil aviation. However, the weak characteristics of these targets make detection challenging. Hyperspectral image (HSI), provides a new approach for the small air target detection task due to its strong ability of capturing both spatial and spectral information simultaneously. In this article, we propose a spectral-spatial enhanced local contrast strategy for hyperspectral small air target detection. An unsupervised band selection step based on the local contrast strategy has been designed based on local contrast (LC-UBSM) to choose bands with better distinguish ability between the target and background in HSI. Then, we have developed an improved RX detection algorithm with combined spatial and spectral variance (CSSV-RX) to detect the target while suppressing both background and noise. Experimental results on both real GAOFEN-5 dataset and simulated dataset based on EO-1 (Earth Observing-1) satellite have validated the effectiveness and robustness of the proposed method.

Without requiring prior spectral information, hyperspectral anomaly detection (HAD) relies on the identification of pixels that significantly deviate from the surroundings. Although autoencoders (AEs) are widely adopted, recent methods integrating low-rank and sparse priors often place these constraints in separate network branches, which hinders unified optimization and increases model complexity. To overcome these limitations, we propose a unified deep framework, termed Joint Low-rank and Sparse Autoencoder (JLSAE), that embeds both priors within a single end-to-end architecture. The AE output serves as a learnable low-rank background, while an auxiliary sparse variable explicitly captures anomaly components. Crucially, we derive our optimization strategy from the Alternating Direction Method of Multipliers (ADMM). This provides a principled training mechanism with explicit low-rank and sparsity regularization, thereby establishing a strong physical interpretability by bridging deep learning with model-based optimization. Experimental results on multiple benchmark datasets show that JLSAE achieves more accurate background reconstruction and outperforms all twelve compared state-of-the-art methods in detection performance. The code is available at https://github.com/fan-jh/JLSAE.

As one of the most prominent research and application directions in hyperspectral remote sensing, anomaly detection aims to identify pixels that exhibit significant spectral differences from their surrounding background, without requiring any prior information about the targets. Traditional hyperspectral anomaly detection (HAD) methods typically characterize the background based on specific mathematical assumptions, which renders them ineffective in handling the complex environments of real-world hyperspectral images (HSIs). Anomaly detection methods based on low-rank features eliminate the need for assumptions about background distribution, making them well-suited for complex hyperspectral data. However, most existing methods based on low-rank features constrain the background component of HSIs by minimizing the nuclear norm. As a convex relaxation of the matrix rank, this norm fails to fully exploit the low-rank characteristics of the data, thereby degrading the detection accuracy of these methods. Meanwhile, solving such problems typically involves repeated singular value decomposition (SVD) operations, which incur significant computational overhead. These deficiencies restrict the application of such methods in practical anomaly detection tasks. To address these limitations, this paper proposes an anomaly detection method for HSIs based on low-rank features and factor group-sparse regularization. The proposed method employs the Schatten-p norm to constrain the background component of HSIs, thereby more efficiently exploiting the inherent low-rank features in the data. Furthermore, factor group-sparse regularization avoids repeated large-scale SVD operations during iterative optimization, thereby reducing the dominant computational burden. Extensive experimental results on five real-world hyperspectral datasets demonstrate that the proposed method achieves the highest five-dataset mean AUC while maintaining the second-shortest average detection time among the methods with valid results on all five datasets.

The rapid proliferation of remote sensing satellite constellations has ushered in an era of unprecedented access to hyperspectral (HS) and multispectral (MS) Earth observation data. HS denoising remains a critical research focus for enhancing the interpretation and application of satellite data. However, on the one hand, the available characteristics of noisy HS satellite data are often limited under real-world and unknown conditions. On the other hand, existing works ignore the complementary advantages offered by MS satellite imagery that can be employed to enhance HS denoising performance. Leveraging the synergistic potential of HS and MS data, this study pioneers a novel paradigm for the joint exploitation and optimization of multi-source remote sensing satellites. A novel Hierarchical Dual Tucker Decomposition (DTucker) framework is proposed to capitalize on the low-rank (LR) tensor property of HS and MS cross-satellite data. The inheritance of properties from the original tensor to the core tensor is explored through a manually designed model-driven constraint or a data-driven multilayer perceptron (MLP) framework. This enables robust integration of MS-derived spatial richness with HS details and significantly enhances the denoising capability. We construct HS-MS data pairs from real satellite observations, including Earth Observing-1, Sentinel-2, Gaofen-1, Gaofen-5, and Gaofen-6. Four datasets span diverse scenes such as runways, urban areas, rivers, and farmlands. The proposed method demonstrates strong generalization across various satellite combinations and application scenarios. Notably, the incorporation of MS data markedly enhances both class discrimination and structural detail in the denoised HS outputs, promoting the performance of subsequent classification and analysis tasks. The datasets and codes implemented in MATLAB and PYTHON will be available on the website: https://github.com/MinghuaWang123/DTucker, contributing to the remote sensing community.

Wetlands are critical ecosystems on Earth, playing a crucial role in climate regulation and coastal sustainability. As Asia’s largest coastal wetland, the Liao River Delta wetland is critical for biodiversity but is progressively degrading due to human activities. Accurate wetland mapping provides a reliable baseline, enabling effective resource management and monitoring. In this study, we propose a dynamic spatiotemporal fusion (DSTF) model specifically for the Liao River Delta wetland. The model employs a multiscale coextraction mechanism to capture phenological dynamics and spatial context from long time-series data. A learnable positional encoding matrix is introduced to adaptively adjust parameters, enhancing feature extraction of seasonal vegetation changes. Finally, a cascaded spatiotemporal self-attention module with gated residual connections establishes cross-dimensional dependencies for classification. Through a series of comparative experiments, the effectiveness of the DSTF model is demonstrated both quantitatively and qualitatively. Based on the classification results, high-resolution mapping of the Liao River Delta wetland was completed at a 10-m spatial resolution from 2016 to 2024. This further reveals the evolutionary characteristics of wetland land cover types in recent years, providing researchers with valuable data to support more refined management strategies.

Object-level change detection is a critical task in remote sensing with broad applications in urban traffic monitoring and smart city management. However, existing methods are primarily validated on aligned single-modal datasets, failing to adapt to real-world challenges, such as cross-view misalignment and multimodal feature discrepancy. To address these, we propose a visible-thermal cross-view object change detection (CVOCD) dataset and a corresponding cross-view multimodal object-level change detection (CMOCD) framework. The CMOCD framework integrates two core modules: the coarse-to-fine feature alignment (CFFA) module resolves spatial misalignment by first establishing global coarse correspondence and then aligning pixel-level fine consistency. The visible-thermal feature fusion (VTFF) module leverages multimodal complementarity through decoupled channel-spatial attention, enhancing feature discriminability in various environments. Comparative results show that CMOCD achieves a mAP@0.5 of 56.5% on the CVOCD dataset, outperforming six state-of-the-art single-modal detectors.

Hyperspectral anomaly detection (HAD) is a promising approach that acts as an unsupervised strategy by distinguishing anomalies from the background. Low-rank representation (LRR) based methods that exploit global correlations at the image level are effective for HAD but often fail to capture long-range correlations, resulting in the loss of important structural details. To address the limitation, we develop a novel HAD via double nonlocal priors (HADDNLP) framework that preserves critical background structure. The proposed HADDNLP method first adopts the patch-wise nonlocal low-rank tensor (NLRT) modeling to explore global correlation along spectrum (GCS) and self-similarity (SS) across distant regions in hyperspectral images (HSIs), thereby preserving the structural and contextual details of the background. Then, the nonlocal means (NLM) prior is integrated to maintain spatial distribution within the HSIs, further enhancing the model’s ability to distinguish anomalies from the background. We optimize the model with an alternating minimization (AM) algorithm for NLRT estimation and an alternating direction method of multipliers (ADMM) for joint background reconstruction and anomaly detection. Experimental results on the real satellite and aerial hyperspectral datasets demonstrate that our proposed approach outperforms state-of-the-art methods in the HAD tasks.

Coded aperture snapshot spectral imaging (CASSI) systems compressively project 3D hyperspectral data onto 2D measurements, offering high imaging speed and data efficiency. However, existing CASSI reconstruction algorithms still suffer from suboptimal reconstruction quality due to the ill-posed nature of hyperspectral compressive sensing reconstruction, which demands effective prior modeling. This paper proposes a novel reconstruction framework that integrates multiple complementary priors, jointly modeling spectral low-rankness, spatial nonlocal self-similarity, and deep image priors to comprehensively capture the intrinsic structure of hyperspectral images across spectral and spatial domains. By combining the strengths of model-based and data-driven priors, the proposed method achieves both strong generalization and expressive capacity. To tackle the optimization challenges posed by multiple regularization terms and parameters, an efficient ADMM-based solver is developed, which decomposes the problem into subproblems with closed-form solutions or those solvable via plug-and-play denoisers. In addition, an adaptive noise estimation mechanism is introduced to automatically tune the regularization parameters, eliminating the need for manual parameter adjustment. Extensive experiments demonstrate that the proposed method consistently outperforms state-of-the-art approaches in terms of reconstruction accuracy and robustness across multiple datasets.

Deep learning has shown strong performance in hyperspectral images change detection (HSIs-CD), but it typically requires a large amount of labeled data. Acquiring such labeled data is often expensive, time consuming, and infeasible in many application scenarios. Few-shot learning (FSL) addresses this limitation by training a model on a large number of training tasks and then applying it to test tasks with only a few labeled samples. However, in HSIs-CD, variations in imaging conditions lead to significant distribution gaps across datasets. Consequently, FSL models trained on these training tasks typically generalize poorly to test tasks and may fail to extract meaningful features. To address this, we propose FSDA-Former, a few-shot domain adaptation Transformer for HSIs-CD. First, it employs a generative pretraining strategy based on a masked autoencoder (FSDA-MAE) to learn domain-agnostic representations and improve feature extraction across diverse tasks. Second, an FSDA-Transformer layer (FSDA-TL) is introduced to refine the extracted features in a metalearning manner, thereby learning class-discriminative representations and constructing a task-specific metric space. Third, a two-stage domain adaptation (FSDA-TDA) strategy is developed to further adapt FSDA-former to the feature domain of the test task. FSDA-TDA further enhances the generalization ability of FSDA-former under distribution shifts between training and test tasks. Experimental results on three real-world HSIs-CD datasets demonstrate that FSDA-former consistently outperforms existing few-shot baselines. It achieves up to 3.60% improvement in kappa and 1.58% in overall accuracy (OA). These results demonstrate the effectiveness and superiority of FSDA-former in HSIs-CD, even when only a few labeled examples are available.

Tiny object detection (TOD) in remote sensing imagery remains challenging because foreground signals are extremely weak in deep feature hierarchies and are easily overwhelmed by high-response background interference. To mitigate this observed foreground-background signal modulation imbalance (FBSMI) difficulty, we propose a signal modulation network (SMN) for remote-sensing TOD. SMN comprises two complementary components. First, an adaptive Wiener filter modulator (AWFM) is inserted after backbone stages to suppress background-dominated noise while preserving weak target-related responses at multiple resolutions. Second, we introduce the novel denoising diffusion transformer (DDT), a featurespace conditional diffusion module that operates on detector feature tensors rather than image pixels. DDT generates multiple diffusion-guided semantic feature variants from high-level fused features and expands the local representation space around weak tiny object evidence. Extensive experiments on AI-TOD, SODA-A, DOTAv2.0, and DIOR-R demonstrate that SMN not only effectively mitigates the FBSMI problem, but also improves detection accuracy, particularly for very tiny and tiny objects, compared with state-of-the-art methods.

Provides society information that may include news, reviews or technical notes that should be of interest to practitioners and researchers.

Deep unfolding networks (DUNs) offer an iterative paradigm that unrolls optimization procedures into a cascaded network structure for hyperspectral image (HSI) denoising. However, existing DUNs for HSI denoising suffer from two notable limitations: 1) the ill-posed inverse problem of handling severely degraded observations, calling for a more powerful degradation-free prior that encodes the spatial–spectral structure of HSIs; and 2) existing discriminative denoisers are deterministic and tend to produce over-smoothed results. In contrast, stochastic latent diffusion models (LDMs) have demonstrated the potential to synthesize high-frequency details with high perceptual quality for HSIs, but often at the cost of sacrificing data fidelity. To this end, we propose a latent diffusion prior enhanced frequency-aware deep unfolding network (Diff-DUN) for HSI denoising. It builds on a half-quadratic splitting (HQS)-based formulation that reconciles the observation consistency of deterministic DUNs with the high perceptual quality of a stochastic LDM. Specifically, we employ a two-phase training strategy to first learn degradation-free priors from clean HSIs, followed by training the LDM to generate these priors conditioned on noisy inputs. Furthermore, we propose a frequency-aware Tetra Transformer (TT) block as the core component within the denoiser, introducing explicit frequency flow and cross-flow interactions to jointly model spatial, spectral, and frequency characteristics for more faithful recovery. Comprehensive evaluations across multiple datasets demonstrate that Diff-DUN achieves a favorable trade-off between denoising accuracy and computational overhead, performing competitively with state-of-the-art methods.

Kolmogorov-Arnold networks (KANs) represent data features by learning the activation functions and demonstrate superior accuracy with fewer parameters, FLOPs, GPU memory usage (Memory), shorter training time (TraT), and testing time (TesT) when handling low-dimensional data. However, when applied to high-dimensional data, which contains significant redundant information, the current activation mechanism of KANs leads to unnecessary computations, thereby reducing computational efficiency. KANs require reshaping high-dimensional data into a one-dimensional tensor as input, which inevitably results in the loss of dimensional information. To address these limitations, we propose weighted activation distribution KANs (WKANs), which reduce the frequency of activations per node and distribute node information into different output nodes through weights to avoid extracting redundant information. Furthermore, we introduce a multilevel tensor splitting framework (MTSF), which decomposes high-dimensional data to extract features from each dimension independently and leverages tensor-parallel computation to significantly improve the computational efficiency of WKANs on high-dimensional data. In this paper, we design SpectralKAN for hyperspectral image change detection using the proposed MTSF. SpectralKAN demonstrates outstanding performance across five datasets, achieving an overall accuracy (OA) of 0.9801 and a Kappa coefficient (K) of 0.9514 on the Farmland dataset, with only 8 k parameters, 0.07 M FLOPs, 911 MB Memory, 13.26 s TraT, and 2.52 s TesT, underscoring its superior accuracy-efficiency trade-off. The source code is publicly available at https://github.com/yanhengwang-heu/SpectralKAN.

Due to the high cost of data acquisition and annotation, it is challenging to obtain large-scale training data for hyperspectral remote sensing images (HRSIs), making it difficult to train object detectors directly. In contrast, visible-light remote sensing images (VLRSIs) are abundant and well-annotated. Hence, cross-domain object detection from VLRSI to HRSI (VLRSI2HRSI) provides a feasible solution for HRSI detection. However, existing cross-domain detection methods are developed for natural images. They neglect the shared spatial characteristics in the remote sensing scenario and fail to model the spectral properties of HRSIs. As a result, directly applying them to VLRSI2HRSI leads to significant performance degradation. To address this issue, we propose a Visible-light to Hyperspectral Cross-Domain Detection Network (VHCDN). We first observe that reconstructed residual representations in both VLRSI and HRSI suppress background and emphasize foreground objects. Based on this observation, we introduce a Residual Extracting Module (REM) to extract domain-shared spatial features. Furthermore, since objects of the same semantic class may consist of different materials and exhibit spectral variations under varying illumination, we design a Spectral Saliency Head (SSH) that models spectral uncertainty. By introducing one HRSI into the source-domain VLRSI datasets to guide the generation of spectral features, the SSH is trained with a spectral heteroscedastic loss and a spectral uncertainty loss, enabling the uncertainty learned in the labeled source domain to be stably transferred to the unlabeled target domain. Experiments on three cross-domain settings against seven representative methods demonstrate that VHCDN achieves state-of-the-art performance, while ablation studies verify the effectiveness of each proposed module.

The primary goal of fusing hyperspectral image (HSI), multispectral image (MSI), and panchromatic (PAN) images is to transfer spatial details from MSI and PAN into HSI to generate high-resolution HSIs (HR-HSIs). Nevertheless, most existing methods rely on black-box deep learning (DL) architectures and rarely leverage this fundamental principle, resulting in limited physical interpretability. To address this challenge, we introduce a novel detail injection-based model for hyperspectral, multispectral, and panchromatic imagery (DIM-HMPF). The framework is engineered to enrich the spatial detail of HSI through fusion with MSI and PAN modalities, all while maintaining high-fidelity spectral information. Therefore, we utilize the physical model of detail injection for unfolding, using DL to learn this process. We further employ an optimization problem-solving method based on the proximal gradient descent (PGD) algorithm, achieved through a series of iterative steps. By unfolding these iterative steps into several network modules, DIM-HMPF is ultimately obtained. On the representative remote sensing hyperspectral datasets of Chikusei, Houston, and Xiongan, both quantitative evaluation metrics and visual comparison results consistently demonstrate that DIM-HMPF significantly outperforms existing mainstream fusion algorithms in terms of reconstruction accuracy, especially as evidenced by its superior performance in structural preservation and spectral fidelity. These results comprehensively validate the distinct superiority, robustness, and effectiveness of DIM-HMPF within the HMPF framework, as well as its strong generalization capability and practical potential in complex scenarios. Our code is publicly available at https://github.com/Freelancefangjian/DIM-HMPF

Ship detection in optical remote sensing imagery is pivotal for various civilian and military applications. However, due to the presence of cluttered backgrounds and the signific ant variations in ship scales, accurately detecting ships in complex maritime environments remains challenging. To address these challenges, we propose the dual attention and scale-aware feature alignment network (DASFA-Net), to effectively detect multiscale ship targets under clutter interference. The DASFA-Net comprises two primary components: deformable spatial attention module with channel integration (DSAM-CI) and bidirectional flow alignment network (BiFAN). The DSAM-CI enhances discrimination between ships and distractors by jointly modeling correlations across the channel and spatial domains and applying deformable spatial attention. Concurrently, BiFAN employs a bidirectional feature flow alignment strategy to resolve spatial misalignment in feature fusion, significantly improving the quality of multiscale feature fusion. Lastly, we design the dynamic alpha complete intersection over union (DA-CIoU) based on prior knowledge of target characteristics to explicitly refine our detector for maritime ships. Extensive experiments on three challenging ship detection benchmarks, including HRSC2016, SCCOS, and FGSRCS have demonstrated the superiority of our DASFA-Net.

Deep learning methods have demonstrated remarkable performance in hyperspectral image (HSI) classification. Nonetheless, these methods may suffer from serious accuracy degradation when subjected to adversarial attacks, which pose security risks. Existing adversarial defense techniques in HSI classification primarily focus on modifying the network structure to extract global spatial features, while overlooking the differences in properties between HSIs and adversarial perturbations. To address this, by leveraging the global low-rank properties of HSIs, we develop a novel low-rank tensor representation learning network (LTRNet) for robust HSI classification against adversarial attacks. First, we devise a new deperturbation block based on low-rank tensor representation learning. This block destroys the perturbation structure by compressing the adversarial examples, mitigating the negative impact of attacks, and subsequently restores the clean images from the compressed versions via low-rank regularization for HSI classification. Second, we present a tensor decomposition-based local feature extraction block, which can achieve large receptive fields with a limited number of parameters. Then, a new tensor transformer is designed for global spatial information learning, where learnable low-rank weight factors contribute to more accurate global spatial dependency modeling. Finally, to eliminate the distribution discrepancy between original images and adversarial examples, we introduce a test-time training strategy for model optimization, thereby achieving better classification accuracy. Experimental results on three benchmark HSI data sets demonstrate that the proposed LTRNet model outperforms state-of-the-art methods under various attacks and perturbation intensities.