BockNet: Blind-Block Reconstruction Network With a Guard Window for Hyperspectral Anomaly Detection

IEEE Transactions on Geoscience and Remote Sensing, Год журнала: 2023, Номер 61, С. 1 - 16

Опубликована: Янв. 1, 2023

Hyperspectral anomaly detection (HAD) aims to identify anomalous targets that deviate from the surrounding background in unlabeled hyperspectral images (HSIs). Most existing deep networks exploit reconstruction errors detect anomalies are prone fit pixels, thus yielding small for anomalies, which is not favorable separating HSIs. In order achieve a superior network HAD purposes, this paper proposes self-supervised blind-block (termed BockNet) with guard window. BockNet creates (guard window) center of network's receptive field, rendering it unable see information inside window when reconstructing central pixel. This process seamlessly embeds sliding dual-window model into our BockNet, inner and outer field outside Naturally, utilizes only predict/reconstruct pixel perceptive field. During pixels varying sizes, typically fall window, weakening contribution results so those reconstructed converge distribution area. Accordingly, HSI can be deemed as pure HSI, error will further enlarged, improving discrimination ability anomalies. Extensive experiments on four datasets illustrate competitive satisfactory performance compared other state-of-the-art detectors.

Язык: Английский

---

Enhanced Autoencoders With Attention-Embedded Degradation Learning for Unsupervised Hyperspectral Image Super-Resolution

IEEE Transactions on Geoscience and Remote Sensing, Год журнала: 2023, Номер 61, С. 1 - 17

Опубликована: Янв. 1, 2023

Recently, unmixing-based networks have shown significant potential in unsupervised multispectral-aided hyperspectral image super-resolution task (MS-aided HS-SR). Nevertheless, the representation ability of and design loss functions still not been fully explored, leaving large room for further improvement. To this end, we propose an enhanced unmixing-inspired network with attention-embedded degradation learning, EU2ADL short, to realize MS-aided HS-SR. First, two coupled autoencoders serve as backbone simultaneously decompose input modalities into abundances corresponding endmembers, whose encoder part is composed a spatial-spectral two-stream subnetwork modality-salient learning parameter-shared one-stream modality-interacted enhancement. More importantly, hybrid model-constrained containing perceptual abundance term degradation-guided introduced eliminate latent distortions. Since built on model, additionally present adaptively estimate unknown parameters. Extensive experimental results four datasets demonstrate effectiveness our proposed methods when compared state-of-the-arts.

Язык: Английский

---

BS³LNet: A New Blind-Spot Self-Supervised Learning Network for Hyperspectral Anomaly Detection

IEEE Transactions on Geoscience and Remote Sensing, Год журнала: 2023, Номер 61, С. 1 - 18

Опубликована: Янв. 1, 2023

Recent years have witnessed the flourishing of deep learning-based methods in hyperspectral anomaly detection (HAD). However, lack available supervision information persists throughout. In addition, existing unsupervised learning/semisupervised learning to detect anomalies utilizing reconstruction errors not only generate backgrounds but also reconstruct some extent, complicating identification original image (HSI). order train a network able background pixels (instead anomalous pixels), this article, we propose new blind-spot self-supervised (called BS3LNet) that generates training patch pairs with blind spots from single HSI and trains fashion. The BS3LNet tends high for low due fact it adopts architecture, i.e., receptive field each pixel excludes itself reconstructs using its neighbors. above characterization suits HAD task well, considering spectral signatures targets are significantly different those neighboring pixels. Our can be considered superb generator, which effectively enhances semantic feature representation distribution weakens expression anomalies. Meanwhile, differences between reconstructed by our used measure degree so separated background. Extensive experiments on two synthetic three real datasets reveal is competitive regard other state-of-the-art approaches.

Язык: Английский

---

X-Shaped Interactive Autoencoders With Cross-Modality Mutual Learning for Unsupervised Hyperspectral Image Super-Resolution

IEEE Transactions on Geoscience and Remote Sensing, Год журнала: 2023, Номер 61, С. 1 - 17

Опубликована: Янв. 1, 2023

Hyperspectral image super-resolution can compensate for the incompleteness of single-sensor imaging and provide desirable products with both high spatial spectral resolution. Among them, unmixing-inspired networks have drawn considerable attention owing to their straightforward unsupervised paradigm. However, most do not fully capture utilize multi-modal information due limited representation ability constructed networks, hence leaving large room further improvement. To this end, we propose an X-shaped interactive autoencoders network cross-modality mutual learning between hyperspectral multispectral data, XINet short, cope problem. Generally, it employs a coupled structure equipped two autoencoders, aiming at deriving latent abundances corresponding endmembers from input correspondence. Inside network, novel architecture is designed by coupling disjointed U-Nets together via parameter-shared strategy, which only enables sufficient flow modalities but also leads informative spatial-spectral features. Considering complementarity across each modality, module transfer knowledge one modality another, allowing better utilization Moreover, joint self-supervised loss proposed effectively optimize our XINet, enabling manner without external triplets supervision. Extensive experiments, including super-resolved results in four datasets, robustness analysis, extension other applications, are conducted, superiority method demonstrated.

Язык: Английский

---

A systematic review of data fusion techniques for optimized structural health monitoring

Information Fusion, Год журнала: 2023, Номер 103, С. 102136 - 102136

Опубликована: Ноя. 10, 2023

Advancements in structural health monitoring (SHM) techniques have spiked the past few decades due to rapid evolution of novel sensing and data transfer technologies. This development has facilitated simultaneous recording a wide range data, which could contain abundant damage-related features. Concurrently, age omnipresent started with massive amounts SHM collected from large-size heterogeneous sensor networks. The abundance information diverse sources needs be aggregated enable robust decision-making strategies. Data fusion is process integrating various produce more useful, accurate, reliable about system behavior. paper reviews recent developments applied systems. theoretical concepts, applications, benefits, limitations current methods challenges are presented, future trends discussed. Furthermore, set criteria proposed evaluate contents original review papers this field, road map provided discussing possible work.

Язык: Английский

---

Tensor Decompositions for Hyperspectral Data Processing in Remote Sensing: A comprehensive review

IEEE Geoscience and Remote Sensing Magazine, Год журнала: 2023, Номер 11(1), С. 26 - 72

Опубликована: Фев. 2, 2023

Owing to the rapid development of sensor technology, hyperspectral (HS) remote sensing (RS) imaging has provided a significant amount spatial and spectral information for observation analysis Earth's surface at distance data acquisition devices. The recent advancement even revolution HS RS techniques offer opportunities realize potential various applications while confronting new challenges efficiently processing analyzing enormous data. Due maintenance 3D inherent structure, tensor decomposition aroused widespread concern spurred research in tasks over past decades. In this article, we aim present comprehensive overview decomposition, specifically contextualizing five broad topics processing: restoration, compressive (CS), anomaly detection (AD), HS–multispectral (MS) fusion, unmixing (SU). For each topic, elaborate on remarkable achievements models RS, with pivotal description existing methodologies representative exhibition experimental results. As result, remaining follow-up directions are outlined discussed from perspective actual practices merged advanced priors deep neural networks. This article summarizes different decomposition-based methods categorizes them into classes, simple adoptions complex combinations other algorithm beginners. We expect that survey provides investigations trends experienced researchers some extent.

Язык: Английский

---

Model-Informed Multistage Unsupervised Network for Hyperspectral Image Super-Resolution

IEEE Transactions on Geoscience and Remote Sensing, Год журнала: 2024, Номер 62, С. 1 - 17

Опубликована: Янв. 1, 2024

By fusing a low-resolution hyperspectral image (LrMSI) with an auxiliary high-resolution multispectral (HrMSI), super-resolution (HISR) can generate (HrHSI) economically. Despite the promising performance achieved by deep learning (DL), there are still two challenges remaining to be solved. First, most DL-based methods heavily rely on large-scale training triplets, which reduces them limited generalization and poor practicability in real-world scenarios. Second, existing pursue higher designing complex structures from off-the-shelf components while ignoring inherent information degradation model, hence leading insufficient integration of domain knowledge lower interpretability. To address those drawbacks, we propose model-informed multi-stage unsupervised network, M2U-Net for short, leveraging both prior (DIP) model information. Generally, is built three-stage scheme, i.e., (DIL), initialized establishment (IIE), generation (DIG) stages. The first stage exploit via tiny network whose parameters outputs will serve as guidance following Instead feeding uninformed noise input three, IIE aims establish expressive HrHSI-relevant resorting spectral mapping thus facilitating extraction further magnifying potential DIP high-quality reconstruction. Last, dual U-shape powerful regularizer capture statistics, U-Nets coupled together cross-attention (CAG) module separately achieve spatial feature final generation. CAG incorporate abundant into reconstruction process guide toward more plausible Extensive experiments demonstrate effectiveness our proposed terms quantitative evaluation visual quality. code available at https://github.com/JiaxinLiCAS.

Язык: Английский

---

Coastline extraction using remote sensing: a review

GIScience & Remote Sensing, Год журнала: 2023, Номер 60(1)

Опубликована: Авг. 10, 2023

Coastlines are important basic geographic elements and mapping their spatial attribute changes can help monitor, model manage coastal zones. Traditional studies focused on the accuracy of extraction methods evolution characteristics coastlines. Thanks to advances in remote sensing for earth observations, recent coastline reveal detailed ocean-land interaction changes. In this review, we aim identify key milestones using by associating emergence major research topics with occurrence multiple application fields, data sources, algorithms. Specifically, define coastlines that be applied different summarize products, analyze principles, advantages disadvantages methods. On basis, discussed development direction challenges involved. This study provides practical insights incorporated into future approaches technologies.

Язык: Английский

---

Model-Guided Coarse-to-Fine Fusion Network for Unsupervised Hyperspectral Image Super-Resolution

IEEE Geoscience and Remote Sensing Letters, Год журнала: 2023, Номер 20, С. 1 - 5

Опубликована: Янв. 1, 2023

Fusing a low-resolution hyperspectral image (LrHSI) with an auxiliary high-resolution multispectral (HrMSI) is burgeoning technique to realize super-resolution, in which learning-based methods have dominated the mainstream direction. However, underutilization of degradation models and strong dependence on large-scale training triplets severely impedes their applicability performance. Considering these issues, we reformulate fusion task as spectral mapping problem hence propose unsupervised model-guided coarse-to-fine network. Specifically, knowledge learning first performed fully excavate latent model information, will serve guidance for better learning. Following that, network constructed multi-scale attentional module head structure tail. The former deployed achieve more informative compression, latter adopted capture relationship, including degradation-guided subnetwork group-by-group coarse reconstruction refinement inter-group correlation dependencies. Finally, HSI can be recovered via established mapping. Extensive experiments simulated real datasets verify superiority our proposed method. code available at https://github.com/JiaxinLiCAS/UMC2FF_GRSL.

Язык: Английский

---

A Multilevel Multimodal Fusion Transformer for Remote Sensing Semantic Segmentation

IEEE Transactions on Geoscience and Remote Sensing, Год журнала: 2024, Номер 62, С. 1 - 15

Опубликована: Янв. 1, 2024

Accurate semantic segmentation of remote sensing data plays a crucial role in the success geoscience research and applications. Recently, multimodal fusion-based models have attracted much attention due to their outstanding performance as compared conventional single-modal techniques. However, most these perform fusion operation using convolutional neural networks (CNN) or vision transformer (Vit), resulting insufficient local-global contextual modeling representative capabilities. In this work, multilevel scheme called FTransUNet is proposed provide robust effective backbone for by integrating both CNN Vit into one unified framework. Firstly, shallow-level features are first extracted fused through layers feature (SFF) modules. After that, deep-level characterizing information spatial relationships well-designed Fusion (FVit). It applies Adaptively Mutually Boosted Attention (Ada-MBA) Self-Attention (SA) alternately three-stage learn cross-modality representations high inter-class separability low intra-class variations. Specifically, Ada-MBA computes SA Cross-Attention (CA) parallel enhance intra- simultaneously while steering distribution towards semantic-aware regions. As result, can fuse manner, taking full advantage accurately characterize local details global semantics, respectively. Extensive experiments confirm superior with other approaches on two fine-resolution datasets, namely ISPRS Vaihingen Potsdam. The source code work available at https://github.com/sstary/SSRS.

Язык: Английский

---