Article
Geochemistry & Geophysics
Gulsen Taskin, E. Fatih Yetkin, Gustau Camps-Valls
Summary: Feature selection (FS) is crucial in various fields, and graph-embedding (GE) techniques have been found efficient for FS due to their ability to preserve the geometric structure of the original feature space. However, the computational cost of GE is high, especially for large-scale problems. This article addresses this issue by combining the GE framework and representation theory to propose a novel FS method that reduces the computational complexity significantly and outperforms its counterparts in high-dimensional hyperspectral image processing.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2023)
Article
Geochemistry & Geophysics
Weiming Li, Qikang Liu, Shuaishuai Fan, Hongyang Bai, Mingrui Xin
Summary: The article proposes a multi-stage superpixel-guided sparse GAT method for hyperspectral image classification tasks, which utilizes spatial topology and spectral sparsity to improve classification accuracy, and feature fusion and pixel-level feature refinement further enhance the performance.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2023)
Article
Optics
Chenming Li, Meiling Wang, Xinyu Sun, Min Zhu, Hongmin Gao, Xueying Cao, Inam Ullah, Qin Liu, Peipei Xu
Summary: Medical hyperspectral imagery (HSI) has become a promising auxiliary diagnostic tool, but the curse of dimensionality and computational complexity can be problems. Therefore, we proposed a novel dimensionality reduction method called TWMDA to extract more discriminative information from high-dimensional HSI data.
OPTICS AND LASER TECHNOLOGY
(2023)
Article
Geochemistry & Geophysics
Peng Wang, Chengyong Zheng, Shengwu Xiong
Summary: This letter introduces a method for reducing the dimension of hyperspectral images by jointly considering spectral redundancy and spatial continuity through graph embedding in core tensor space. The method embeds a graph to the core tensor space during Tucker decomposition to maintain the distance property between intraclass samples. By constraining the projected matrices with orthogonality to increase stability and discriminative features, the method outperforms many other tensor methods in experimental results.
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS
(2021)
Article
Engineering, Electrical & Electronic
Kai Chen, Guoguo Yang, Jing Wang, Qian Du, Hongjun Su
Summary: In this article, a novel unsupervised dimensionality reduction (DR) method called MFS-PE is proposed for hyperspectral image classification. It combines spatial and spectral features, implements sample augmentation and neighbor selection, and maximizes global features to reveal the intrinsic structure of HSIs. Experimental results demonstrate that MFS-PE outperforms state-of-the-art DR methods in terms of classification accuracy.
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
(2023)
Article
Automation & Control Systems
Jan Niklas Boehm, Philipp Berens, Dmitry Kobak
Summary: Neighbor embeddings are a family of methods for visualizing high-dimensional data sets using kNN graphs. By changing the balance between attractive and repulsive forces, a spectrum of embeddings with different trade-offs can be obtained. UMAP and ForceAtlas2 algorithms represent different levels of attraction on this spectrum.
JOURNAL OF MACHINE LEARNING RESEARCH
(2022)
Article
Computer Science, Interdisciplinary Applications
Hongmin Gao, Meiling Wang, Xinyu Sun, Xueying Cao, Chenming Li, Qin Liu, Peipei Xu
Summary: In this study, a novel method called unsupervised dimensionality reduction via tensor-based low-rank collaborative graph embedding (TLCGE) is proposed to preserve the intrinsic 3-D data structure of medical hyperspectral imagery (HSI) while reducing dimensions. The proposed TLCGE introduces the entropy rate superpixel (ERS) segmentation algorithm to generate superpixels, constructs a low-rank collaborative graph weight matrix on each superpixel, and fully explores the local and global structures within each superpixel. Experimental results demonstrate the computational efficiency and effectiveness of the proposed TLCGE, making it suitable for preprocessing in real medical HSI classification or segmentation.
COMPUTER METHODS AND PROGRAMS IN BIOMEDICINE
(2023)
Article
Computer Science, Theory & Methods
Laurent Amsaleg, James Bailey, Amdeie Barbe, Sarah M. Erfani, Teddy Furon, Michael E. Houle, Milos Radovanovic, Xuan Vinh Nguyen
Summary: This paper investigates the impact of adversarial perturbation on the ranking of classification and retrieval related objects, revealing that as the dimensionality of the data domain increases, the amount of perturbation needed to subvert neighborhood rankings decreases. Theoretical analysis and experiments demonstrate that the vulnerability to adversarial attack rises with increasing dimensionality.
IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY
(2021)
Article
Telecommunications
Monika Sharma, Mantosh Biswas
Summary: The paper introduces a Collaborative Representation based K Closest Neighbor classes (CRKCN) classification algorithm that shows promising improvements in classification performance for hyperspectral images compared to traditional methods.
WIRELESS PERSONAL COMMUNICATIONS
(2022)
Article
Physics, Multidisciplinary
Hao Li, Yuanshu Zhang, Yong Ma, Xiaoguang Mei, Shan Zeng, Yaqin Li
Summary: The study introduces two methods, PENRC and J-PENRC, based on representation for hyperspectral image classification. By combining different norm penalties, similarity matrix, adaptive dictionary, and neighbor information, these methods aim to improve the classification performance and accuracy.
Article
Biology
Hongmin Gao, Mengran Yang, Xueying Cao, Qin Liu, Peipei Xu
Summary: In this paper, a dimensionality reduction algorithm named enhanced discriminant local constraint preserving projection (EDLCPP) is proposed to address the issue of data redundancy and noise in microscopic hyperspectral images. The proposed method effectively reduces the spectral features and improves the recognition and classification performance through global spectral attention mechanism, high discriminability sample selection, graph construction, and graph embedding modules.
COMPUTERS IN BIOLOGY AND MEDICINE
(2023)
Article
Computer Science, Information Systems
Wan-Lei Zhao, Hui Wang, Peng-Cheng Lin, Chong-Wah Ngo
Summary: This paper addresses the issue of merging k-nearest neighbor (k-NN) graphs in two different scenarios. A symmetric merge algorithm is proposed to combine two approximate k-NN graphs, facilitating large-scale processing. A joint merge algorithm is also proposed to expand an existing k-NN graph with a raw dataset, enabling the incremental construction of a hierarchical approximate k-NN graph.
IEEE TRANSACTIONS ON BIG DATA
(2022)
Article
Geochemistry & Geophysics
Youqiang Zhang, Guo Cao, Bisheng Wang, Xuesong Li, Prince Yaw Owusu Amoako, Ayesha Shafique
Summary: In this study, a graph-based method called dual sparse representation graph-based collaborative propagation (DSRG-CP) is proposed for hyperspectral image (HSI) classification. DSRG-CP combines spectral and spatial graphs to carry out label propagation iteratively, achieving competitive results in HSI classification.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2022)
Article
Geochemistry & Geophysics
Xinwei Jiang, Liwen Xiong, Qin Yan, Yongshan Zhang, Xiaobo Liu, Zhihua Cai
Summary: This letter proposes an unsupervised dimensionality reduction method, LRCRP, which introduces Laplacian regularization and local enhancement into collaborative representation to reduce the spectral dimension in a graph embedding framework. Experimental results demonstrate the effectiveness of the proposed model.
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS
(2022)
Article
Geochemistry & Geophysics
Yule Duan, Hong Huang, Yuxiao Tang
Summary: The LC-SMHL algorithm is a novel dimensionality reduction method that can simultaneously discover the manifold-based sparse structure and multivariate discriminant sparse relationship of HSI. By constructing sparse hypergraphs and learning optimal projections, the LC-SMHL method demonstrates better performance in HSI classification.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2021)
Article
Geochemistry & Geophysics
Yong Chen, Jinshan Zeng, Wei He, Xi-Le Zhao, Ting-Zhu Huang
Summary: The proposed factor smoothed TR decomposition model demonstrates superior performance in capturing the spatial-spectral continuity of HR-HSI and improving the quality of reconstructed images, leading to the best results compared to state-of-the-art methods.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2022)
Article
Geochemistry & Geophysics
Hongyan Zhang, Jingyi Cai, Wei He, Huanfeng Shen, Liangpei Zhang
Summary: This article proposes a new method that simultaneously explores the low-rank characteristic of noise-free HSI and the low-rank structure of stripe noise on each band of the HSI, to achieve separation of noise-free HSI, stripe noise, and other mixed noise within one unified framework.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2022)
Article
Geochemistry & Geophysics
Xianwei Zheng, Xiujie Wu, Linxi Huan, Wei He, Hongyan Zhang
Summary: The proposed unified gather-to-guide network (G2GNet) for remote sensing semantic segmentation utilizes a gather-to-guide module (G2GM) to calibrate RGB features and improve segmentation performance. By generating cross-modal descriptors and using channel-wise guide weights, the G2GM preserves informative features while suppressing redundant and noisy information. Extensive experiments demonstrate the robustness of G2GNet to data uncertainties and its ability to enhance the semantic segmentation of RGB and auxiliary remote sensing data.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2022)
Article
Computer Science, Artificial Intelligence
Wei He, Yong Chen, Naoto Yokoya, Chao Li, Qibin Zhao
Summary: In this paper, a new coupled tensor ring factorization (CTRF) model is proposed for hyperspectral super-resolution (HSR). The CTRF model can effectively learn the tensor ring core tensors of high-resolution HSI while exploiting the low-rank property of each class, outperforming previous coupled tensor models.
PATTERN RECOGNITION
(2022)
Article
Geography, Physical
Ziqiao Wang, Hongyan Zhang, Wei He, Liangpei Zhang
Summary: This paper proposes a novel framework called the Phenology Alignment Network (PAN) for cross-phenological-region (CPR) crop mapping. PAN utilizes deep recurrent networks and unsupervised domain adaptation to align temporal-spectral features and improve mapping accuracy. Experimental results show that PAN achieves significant improvement in crop mapping and has the ability to correct inaccurate predictions.
ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING
(2022)
Article
Computer Science, Artificial Intelligence
Yuting Wan, Ailong Ma, Wei He, Yanfei Zhong
Summary: Due to noise during data acquisition, the quality of hyperspectral images is degraded and their applications are limited. Traditional methods use low-rank and sparse matrix decomposition to restore pure data, but optimization of the l0-norm is difficult. This article proposes an accurate multiobjective low-rank and sparse denoising framework for accurate modeling without sensitive regularization parameters, achieving effective optimization through a subfitness strategy.
IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION
(2023)
Article
Geochemistry & Geophysics
Hao Zhang, Ting-Zhu Huang, Xi-Le Zhao, Wei He, Jae Kyu Choi, Yu-Bang Zheng
Summary: In this article, we propose a transformed domain model called T-RSTR, which combines sparse and low-tensor-ring (TR)-rank priors for hyperspectral image denoising. T-RSTR integrates the transform-based low-TR-rank and sparse regularizers to capture the global low-rankness and sparsity of the transformed tensors, resulting in significantly improved denoising performance. An elaborately designed proximal alternating minimization-based algorithm is used to solve the T-RSTR model, and its convergence is theoretically proved. Extensive numerical results demonstrate the superiority of T-RSTR over competing methods.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2023)
Article
Agriculture, Multidisciplinary
Xiangyu Liu, Wei He, Hongyan Zhang
Summary: Plastic greenhouse (PG) is important in protected agriculture for enhancing crop yields and quality. However, accurately counting and mapping PGs based on high-resolution remote sensing images is challenging due to their variability and interconnection. This paper proposes a novel Cross-Regional Segmentation and Counting framework (CRSC) that integrates the unsupervised Style Transfer Network (STNet) and dual task-based Segmentation Counting Network (SCNet) to simultaneously perform PG segmentation and counting across different regions. The CRSC framework improves cross-regional segmentation and counting of PGs, especially in cases with limited labeled samples in the source region.
COMPUTERS AND ELECTRONICS IN AGRICULTURE
(2023)
Article
Geochemistry & Geophysics
Mengyuan Wang, Wei He, Hongyan Zhang
Summary: To address the weak ability of convolutional neural network (CNN) models to capture long-distance correlation, which leads to spectral distortion and edge blurring, we propose a spatial-spectral transformer network for hyperspectral image (HSI) denoising. By introducing the shifted window-based transformer method, we model image content correlation while preserving the local inductive bias for denoising HSIs. Experimental results demonstrate the superiority of our proposed method, called spatial-spectral transformer denoising (SSTD), over other mainstream learning-based HSI denoising algorithms on simulated and real data.
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS
(2023)
Article
Engineering, Electrical & Electronic
Wei He, Tatsumi Uezato, Naoto Yokoya
Summary: This paper introduces an interpretable network architecture, which incorporates self-attention mechanism to create a deep attention prior (DAP) for solving inverse imaging problems. Compared to the deep image prior (DIP), DAP provides a better understanding of the prior and adopts more stable input and early stopping handling. Experiments demonstrate that DAP performs effectively as an image prior for various inverse imaging tasks, such as denoising, inpainting, and pansharpening, and also shows potential applications in higher-level processing, such as interactive segmentation and selective colorization.
IEEE TRANSACTIONS ON COMPUTATIONAL IMAGING
(2023)
Proceedings Paper
Geosciences, Multidisciplinary
Weinan Cao, Hongyan Zhang, Wei He, Hongyu Chen, Ewe Hong Tat
Summary: This paper proposes a novel method for anomaly detection in hyperspectral images based on a low-rank module embedded autoencoder network. The method purifies the background and reconstructs the image background to separate anomalies. It fully considers the differences between anomalies and backgrounds in spatial and spectral dimensions to better reconstruct the background.
2022 IEEE INTERNATIONAL GEOSCIENCE AND REMOTE SENSING SYMPOSIUM (IGARSS 2022)
(2022)
Proceedings Paper
Geosciences, Multidisciplinary
Weinan Cao, Hongyan Zhang, Wei He, Hongyu Chen, Ewe Hong Tat
Summary: The primary purpose of anomaly detection in hyperspectral images is to separate different anomaly targets from their surrounding backgrounds. In this paper, a multi-scale background reconstruction network with low-rank embedding (MSBRNet) is proposed to extract low-rank background features using neural networks and reconstruct the background to effectively separate anomalies from the background.
2022 IEEE INTERNATIONAL GEOSCIENCE AND REMOTE SENSING SYMPOSIUM (IGARSS 2022)
(2022)
Proceedings Paper
Computer Science, Artificial Intelligence
Wei He, Quanming Yao, Naoto Yokoya, Tatsumi Uezato, Hongyan Zhang, Liangpei Zhang
Summary: In this study, a neural network architecture for hyperspectral image restoration is proposed. By disentangling 3D convolution into 2D spatial and spectral convolutions and building a spectrum-aware search space, the most efficient architecture is automatically learned using neural architecture search strategy to fully exploit the spatial-spectral information. Experimental results demonstrate the remarkable performance of the searched architecture in various reconstruction tasks.
COMPUTER VISION, ECCV 2022, PT XIX
(2022)
Article
Geochemistry & Geophysics
Jingyi Cai, Wei He, Hongyan Zhang
Summary: This article proposed an HSI denoising and destriping method based on anisotropic spatial and spectral total variation regularized double LR approximation (ATVDLR), which can achieve superior performance in complex mixed noise and image structural information protection.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2022)
Article
Geochemistry & Geophysics
Yong Chen, Wei He, Xi-Le Zhao, Ting-Zhu Huang, Jinshan Zeng, Hui Lin
Summary: The TLNLGS method proposed in this article combines global spectral correlation and nonlocal sparse prior in denoising HSI images, resulting in better noise removal and preservation of image smooth structure.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2022)
