最新刊期

    28 1 2024

      Reviews

    • SU Yuanchao,XU Ruoqing,GAO Lianru,HAN Zhu,SUN Xu
      Vol. 28, Issue 1, Pages: 1-19(2024) DOI: 10.11834/jrs.20243165
      Development of deep learning-based hyperspectral remote sensing image unmixing
      摘要:Hyperspectral remote sensing is an advanced technique for earth observation that combines physical imagery and spectral analysis technology. Therefore, hyperspectral remote sensing can obtain fine spectral and rich spatial information from imaged scenes, merging the spatial and spectral information into data cubes. These data cubes exhibit narrow spectral bands and a high spectral resolution, allowing different land cover objects to be distinguished. Hyperspectral remote sensing images, with their high spectral resolution and cube characteristics, have gradually become among the most essential supporting data in remote sensing engineering applications. However, due to spatial resolution limitations, the mixed pixel problem has hindered the development of hyperspectral remote sensing in fine-scale object information extraction. At present, hyperspectral unmixing is one of the most effective analytical techniques for dealing with mixed pixel problems, aiming to break through spatial resolution limitations by analyzing the components within pixels. Hyperspectral unmixing refers to any process that separates pixel spectra from a hyperspectral image into a collection of pure constituent spectra, called endmembers, and a set of corresponding abundance fractions. At each pixel, the endmembers are generally assumed to represent the pure materials in the scene, while the abundances represent the percentage of each endmember. For the fine-scale interpretation of object information, many unmixing methods have been developed for hyperspectral remote sensing images in the remote sensing field over the past 30 years, mitigating the impact of mixed pixel problems on quantitative remote sensing analysis. Currently, with the development of deep learning, an increasing number of deep learning theories and tools are used to deal with mixed pixel problems. Many new methods using deep learning for unmixing have been developed, and unmixing technology research has gradually entered a new stage of development with deep learning. Deep-learning-based methods make better use of hidden information, have a relatively lower dependence on prior knowledge, and have a stronger adaptability to complex scenes than traditional unmixing methods. Although deep learning-based unmixing methods have developed rapidly in recent years and are diverse, the analysis and summary of the work on such methods have not kept up with the pace of technological development. A timely summary of the latest research progress on developing a specific field of research has a significant role in promoting the technology. Thus, this paper sorts out the existing deep learning-based unmixing methods, classifying them according to the adopted spectral mixing models, the deep network training modes, and whether spectral variability is considered. Furthermore, this paper introduces these deep learning-based approaches and summarizes their characteristics, making the use of these methods in special works convenient for users or readers. Finally, the development of deep learning methods is summarized, referring to the current technical status, characteristics, and development prospects. In addition, some existing deep learning unmixing methods were tested in this study and organized to facilitate the research and application of unmixing technology. The development of deep learning will continue to promote the progress of unmixing techniques. In recent years, deep learning-based unmixing methods have developed rapidly and have been gradually used in vegetation distribution investigation and agricultural yield estimation, implying their good development prospect and application value. his paper can provide valuable references for researching unmixing technology in the future.  
      关键词:hyperspectral remote sensing;unmixing;deep learning;machine learning;deep neural network;remote sensing image processing;remote sensing intelligent interpretation;subpixel interpretation   
      602
      |
      201
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 48419184 false
      发布时间:2024-02-29
    • YANG Xing,FANG Leyuan,YUE Jun
      Vol. 28, Issue 1, Pages: 20-41(2024) DOI: 10.11834/jrs.20243404
      Advances in semi-supervised classification of hyperspectral remote sensing images
      摘要:Hyperspectral remote sensing technology has been widely used in remote sensing, agriculture, geological exploration, and other fields, and hyperspectral image classification is one of the most important research directions. Benefiting from sufficient label information, supervised learning has achieved good results in this field. However, in many practical applications of hyperspectral remote sensing images, sufficient label samples are difficult to obtain. One of the most important reasons is the widespread use of hyperspectral remote sensing technology, which produces huge amounts of unlabeled data. Another is the high cost of labeling. Meanwhile, unsupervised learning cannot accurately cluster unknown data, and its clustering categories are to match to real categories. Both supervised and unsupervised learning have their unavoidable disadvantages. Therefore, semi-supervised learning that uses a large number of unlabeled samples and a small number of labeled samples should be explored. In recent years, significant progress has been made in the semi supervised classification of hyperspectral remote sensing images. Researchers have proposed many innovative algorithms and technologies to address the problem of insufficient data annotation. This article reviews the progress of the semi supervised classification research on hyperspectral remote sensing images in recent years, discussing key technologies and methods.This paper starts with semi-supervised classification and hyperspectral remote sensing technologies. First, the first part of this paper introduces some basic concepts of semi-supervised learning, including semi-supervised and unsupervised learning, supervised learning, and the application of semi-supervised learning. The second part introduces the development of hyperspectral remote sensing imaging technology domestically and internationally and the application of hyperspectral remote sensing in various fields, such as land and resource surveys, agriculture and forestry remote sensing, and urban environmental monitoring. Second, the three basic assumptions of the theory, process, and data distribution of semi-supervised learning are analyzed, and four typical types are introduced: low-density separation, generative, disagreement-based (difference-based), and graph-based methods. The algorithm flow and core ideas of each method are introduced in detail. The summarized current development status, typical algorithms, and research progress of hyperspectral remote sensing image classification are analyzed. Further, the advantages and disadvantages of each algorithm are enumerated. Then, common open-source algorithms were compared on three publicly available datasets, namely, Indian Pines, Pavia University, and Houston 2013. Finally, by analyzing existing semi-supervised learning technologies and experimental results, the challenging problems and development trends of semi-supervised learning in hyperspectral remote sensing are summarized.The graph-based semi-supervised classification method performs better than other semi-supervised classification methods, which may be because the graph model can model the relationship and similarity between samples, connect similar samples, and capture the intrinsic structure and similarity in a dataset.Semi-supervised learning can efficiently utilize both labeled data and unlabeled data. The future development trend of semi-supervised classification is mainly in three aspects: how to effectively use a large number of unlabeled samples; how to fully consider multiple factors, such as performance and computational complexity; and how to select features. These aspects will affect the stability, generalization, practicability, and performance of the algorithm.  
      关键词:hyperspectral image;semi-supervised classification;Low-density separation;generative model;graph neural network   
      240
      |
      106
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 48419183 false
      发布时间:2024-02-29
    • QU Bo,ZHENG Xiangtao,QIAN Xueming,LU Xiaoqiang
      Vol. 28, Issue 1, Pages: 42-54(2024) DOI: 10.11834/jrs.20232405
      Research progress on hyperspectral anomaly detection
      摘要:The applications of remote sensing images in numerous fields have been increasing with the continuous development of aerospace and remote sensing technologies. HyperSpectral Image (HSI) is a common type of remote sensing image that comprises a series of two-dimensional remote sensing images as a 3D data cube. Each two-dimensional image in HSI can reveal the reflection/radiation intensity of different wavelengths of electromagnetic waves, and each pixel of HSI corresponds to the spectral curve reflecting the spectral information in different wavelengths. Therefore, the hyperspectral remote sensing images are characterized by “spatial-spectral integration,” which contains not only spectral information with strong discriminant but also rich spatial information. Therefore, the hyperspectral data have considerable application potential.Hyperspectral anomaly detection aims to detect pixels in a scene with different characteristics from surrounding pixels and determines them as anomalous targets without any previous knowledge of the target. Hyperspectral anomaly detection is an unsupervised process that does not require any priori information regarding the target to be measured in advance; thus, this type of detection plays a crucial role in real life. For example, anomaly target detection technology can be used to search and rescue people after a disaster, quickly determine the fire point of a forest fire, and search mineral points in mineral resource exploration. Hyperspectral anomaly detection has been a popular research direction in the area of remote sensing image processing in recent years, and a numerous researchers have conducted extensive research and achieved rich research results.However, hyperspectral anomaly detection still encounters many difficult problems. For example, the targets of the same material may exhibit various spectral characteristics due to the different imaging equipment and environment, which may interfere with the detection results and lead to the problem of “same object with different spectra.” Meanwhile, the targets of different materials may also exhibit the problem of “different objects with different spectra.” Then, most of the existing hyperspectral anomaly detection algorithms are only in the laboratory stage and with low technology maturity. Furthermore, the hyperspectral data may have numerous spectral bands that contain a considerable amount of redundant information, which increases the difficulty of data processing. Moreover, the number of publicly available hyperspectral anomaly detection datasets is insufficient and mostly old.In this paper, the main research progress of hyperspectral anomaly detection is first summarized. The existing mainstream algorithms are then classified and summarized. These algorithms are mainly divided into five categories: statistics-based anomaly detection methods, data expression-based anomaly detection methods, data decomposition-based anomaly detection methods, deep learning-based anomaly detection methods, and other methods. Through the investigation, analysis, and summary of the existing methods, three future development directions of hyperspectral anomaly detection are proposed. (1) Database expansion: new datasets with additional images and highly sophisticated remote sensing sensors are introduced. (2) Multisource data combination: the advantages of different imaging sensors and various types of remote sensing data are maximized. (3) Algorithm practicality: the anomaly detection algorithms are relayed for application on real platforms.  
      关键词:remote sensing;hyperspectral remote sensing;hyperspectral anomaly detection;deep learning;matrix factorization   
      376
      |
      327
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 41050011 false
      发布时间:2024-02-29

      Hyperspectral Target Recognition

    • JIA Sen,LIU Kuan,XU Meng,ZHU Jiasong
      Vol. 28, Issue 1, Pages: 55-68(2024) DOI: 10.11834/jrs.20242398
      Hyperspectral anomaly detection based on spatial-spectral multichannel autoencoders
      摘要:Hyperspectral anomaly detection is a type of unsupervised target detection that is crucial in the national economy and attracts the attention of numerous researchers. However, hyperspectral anomaly detection faces several challenges, such as diversified anomaly targets, difficulty in distinguishing anomalies from the background, and low detection accuracy. A hyperspectral anomaly detection method based on multichannel autoencoders is proposed to form a high-dimensional spatial-spectral feature space to address the above challenges. First, weighted spatial-spectral Gabor kernels with different scales and directions are proposed to extract the spatial-spectral features from hyperspectral images. These Gabor kernels are then redefined to increase the gap between the central and surrounding values in the kernels. The spatial-spectral features are extracted by weighted spatial-spectral Gabor kernels to form a high-dimensional spatial-spectral feature space. Second, multichannel autoencoders reduce the redundancy of multiscale spatial-spectral features in spectral dimension, extract the principal features from high-dimensional spatial-spectral feature space, and transform them into the principal feature representation space. Finally, a feature enhancement method based on hyperbolic tangent function and morphological filters is proposed to improve the distinction between abnormal targets and background noise and address the background noise in the spatial dimension. Mahalanobis distance is used to detect anomalies in the enhanced principal feature representation space. The proposed method is compared with nine state-of-the-art anomaly detection methods on five hyperspectral data sets. Anomaly Detection Maps (ADMs), Receiver Operating Characteristics (ROCs), Area Under Curves (AUCs), and box plots between abnormal and background pixels are used to evaluate the performance of the compared methods. AUC is a quantitative evaluation method, and the others are qualitative evaluation methods. The anomaly detection maps obtained by the proposed method easily locates abnormal targets compared with other methods. The ROC curves on five hyperspectral data sets show that the proposed method has a superior performance. The AUC values of five hyperspectral data sets are 0.9910, 0.9912, 0.9968, 0.9806, and 0.9812. The box plots show that the proposed method increases the gap between the anomalies and the background. The ablation experiments show that weighted spatial-spectral Gabor can extract more significant spatial-spectral features than three-dimensional Gabor. The principal feature representation space obtained by multichannel autoencoders is highly conducive to hyperspectral anomaly detection and improves detection accuracy. The feature enhancement method based on hyperbolic tangent function can improve the distinction between abnormal targets and background noise. The proposed method can extract significant spatial-spectral features from hyperspectral images to address the diversification of anomaly types and form a high-dimensional spatial-spectral feature space. The multichannel autoencoders convert the high-dimensional spatial-spectral feature space into the principal feature representation space, which can effectively reduce band redundancy in the spectral dimension and decrease the computational complexity in the anomaly detection process. The feature enhancement method based on hyperbolic tangent function can significantly improve the distinction between anomalies and background noise to locate the abnormal target.  
      关键词:hyperspectral image;anomaly detection;multichannel autoencoders;weighted spatial-spectral Gabor;hyperbolic tangent function;feature enhancement method   
      119
      |
      105
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 48419712 false
      发布时间:2024-02-29
    • LI Chenyu,HONG Danfeng,ZHANG Bing
      Vol. 28, Issue 1, Pages: 69-77(2024) DOI: 10.11834/jrs.20233075
      Deep unfolding network for hyperspectral anomaly detection
      摘要:Hyperspectral Anomaly Detection(HAD) is one of the most critical topic in hyperspectral remote sensing and has been extensively addressed in the literature over the past decade. Among them, Low-Rank Representation(LRR) models are widely used owing to their powerful separation ability for the background and targets. But their applications in practical situations still remain limited due to the extreme dependence on manual parameter selection and relatively poor generalization ability. To this end, this paper combines the LRR model with deep learning techniques to propose a new underlying network for HAD, called LRR-Net. This method efficiently solves the LRR model with the help of the Alternating Direction Method of Multipliers (ADMM) optimizer, and incorporates the solution as a priori knowledge into the deep network to guide the optimization of parameters, providing a theoretical basis for deep networks. In addition, LRR-Net converts a series of regularized parameters into learnable network parameters in an end-to-end manner, thus avoiding manual tuning of parameters. Experimental results obtained from publicly available datasets and our datasets demonstrate that the LRR-Net method outperforms many state-of-the-art model-based and deep-based algorithms of hyperspectral anomaly detection. Overall, deep learning networks are powerful in learning and are robust compared to traditional models in processing datasets with different complexity. However, despite the strong fitting ability of deep learning data, the necessary prior information is lacking, which often makes the algorithm fall into the local optima, which leads to the failure of deep learning to guarantee the stability of HAD results. The model-based algorithm can better make up for this defect, which can often get better results by improving the separability between the background and the target. Nonetheless, these LRR-based methods are unable to effectively suppress background noise due to their limited representation power, such as shadows, trees, and edges in complex scenes, with relatively large volatility in detection effects. The LRR-Net presented in this paper combines the advantages of the above two methods, and the experimental results of four typical scenarios show that the search of the optimal parameters in the neural network can effectively solve the HAD problem in an adaptive way, which is more physically meaningful.  
      关键词:hyperspectral remote sensing image;anomaly detection;deep unfolding;Low-Rank Representation (LRR);Alternating Direction Multiplier Method (ADMM)   
      115
      |
      58
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 48420065 false
      发布时间:2024-02-29
    • ZHOU Kun,XU Yang,WEI Jie,WU Zebin,WEI Zhihui
      Vol. 28, Issue 1, Pages: 78-87(2024) DOI: 10.11834/jrs.20232225
      Hyperspectral target detection based on constrained energy minimization variational autoencoder
      摘要:Target detection is an important research direction in the hyperspectral field. Hyperspectral target detection aims to distinguish pixels as background or target according to the spectral characteristics of the target. Several detection algorithms have been proposed in the past few decades. However, the complexity of background samples in hyperspectral images and the limited number of target samples lead to considerable challenges in detection algorithms. A hyperspectral target detection algorithm based on background reconstruction is proposed in this paper. Taking advantage of the large proportion of background samples in hyperspectral images, the self-representation model of the background samples is trained, and then the background is reconstructed. Simultaneously, the constrained energy minimization is used to detect the residual image, and the reconstructed background is used for the calculation of the correlation matrix. Therefore, the target sample is not involved in the calculation to affect the response energy of the target sample, and the detection accuracy is improved. Results on real hyperspectral image data are better than those of comparison experiments, which verify the effectiveness of this method.Obtaining numerous training sets of artificially labeled hyperspectral data is difficult. Therefore, using limited samples to train deep neural networks is the largest difficulty in applying deep learning to hyperspectral target detection. When calculating the average output energy of the background, the calculation of the correlation matrix of all samples is used. Therefore, the target pixel also participates in the calculation, causing a certain degree of damage to the target spectrum. The background is used as a training sample to train the entire network to solve the above problems, and the reconstructed background is utilized for constrained energy minimization detection to reduce the impact on the target spectrum during the detection process.This paper proposes a hyperspectral target detection based on constrained energy minimization variational autoencoder. First, the image is roughly detected to obtain the training background sample. The background sample then is inputted into the variational autoencoder for training. The network introduces a constraint energy minimization regularization to remove the characteristics of the target sample and help the reconstructed sample retain only the background information. The 3D residual is acquired by calculating the difference between the original image and the reconstructed background. Thus, the constraint energy minimization is used to detect the residual. The background correlation matrix is employed in the detection process to replace the residual correlation matrix. Finally, the final detection result is obtained by weighting.Compared with other comparative experiments, the proposed method achieved good detection results. The AUC table shows that most of the six hyperspectral datasets performed better than the comparison experiments, and most of the AUC values reached more than 99%. The detection map reveals that the target part is well detected. A close ROC curve to the upper left corner yields satisfactory effects. The curve of the proposed method performs well compared with other methods.Overall, a hyperspectral target detection based on constrained energy minimization variational autoencoder is proposed. The algorithm utilizes the characteristics of the large distribution of background pixels in hyperspectral images. First, the coarse detection is used to obtain the background samples for training. The variational autoencoder then trains the background self-representation model and reconstructs its background. A constraint energy minimization regularization is introduced to help the reconstructed samples retain only the background information. Simultaneously, when using the constraint energy minimization to detect the residual image, the background correlation matrix contributes to the calculation to prevent the participation of non-background pixels in the calculation and lose the target signal output. Experimental results on real hyperspectral datasets show that the algorithm outperforms other comparative experimental results. However, the model is highly dependent on the effect of background reconstruction. If the effect of background reconstruction is superior, then the detection rate will be high. Therefore, improving the stability of background reconstruction in the future is necessary.  
      关键词:hyperspectral;target detection;background reconstruction;constrained energy minimization;correlation matrix   
      70
      |
      47
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 48419212 false
      发布时间:2024-02-29

      Hyperspectral Information Extraction

    • LIU Qinsen,SUN Bangyong
      Vol. 28, Issue 1, Pages: 88-104(2024) DOI: 10.11834/jrs.20233067
      Optical-signal token guided change detection network for heterogeneous remote sensing image
      摘要:Change Detection (CD) is a vital technique for identifying and analyzing changes over time in a specific area using optical signals from remote sensing images. This technique has been extensively utilized in various fields, including national defense security, environmental monitoring, and urban construction. However, some challenges in achieving accurate and reliable CD are still encountered due to inherent disparities in imaging mechanisms, spectral ranges, and spatial resolutions among heterogeneous images. These challenges lead to issues such as inadequate accuracy, missed detections, and false detections. Heterogeneous remote sensing images can be regarded as sequences of different optical signals from the channel perspective. For example, RGB and infrared images can be regarded as sequences of spectral signals from different ranges. Transformers employ a multi-head attention mechanism that can effectively handle and analyze sequence information to achieve accurate heterogeneous CD. Thus, the paper proposes an optical signal token guided CD network for heterogeneous remote sensing images.This paper presents a novel heterogeneous CD network, primarily comprising the optical-signal token transformer (OT-Former) and the cross-temporal transformer (CT-Former). The proposed method demonstrates the capacity to effectively handle diverse remote sensing images of distinct categories and attain precise CD results. Specifically, OT-Former can encode diverse heterogeneous images in channel-wise for adaptively generating the optical-signal tokens. Meanwhile, CT-Former can use the optical-signal tokens as a guide to interact with the patch token for the learning of change rules. Moreover, a Difference Amplification Module (DAM) is embedded into the network to enhance the extraction of difference information. This module utilizes a 1×2 convolutional kernel to effectively fuse difference information. Finally, the differential token is predicted by multilayer perceptron to output the CD results.Experiments were conducted on three heterogeneous datasets and one homogeneous dataset to evaluate the performance of the proposed method. Furthermore, the proposed method was compared with six typical CD methods and evaluated the performance using overall accuracy (OA), Kappa coefficient, and F1-score, among other evaluation metrics, to validate the effectiveness of the proposed network in this study. A limited number of samples were utilized for training during the experiment. Under identical experimental conditions, the proposed method demonstrated exceptional performance in homogeneous and heterogeneous CD. The results show that the proposed approach surpasses existing state-of-the-art methods in terms of qualitative and visual performance. Additionally, ablation experiments and parameter analyses were conducted to validate the effectiveness of the proposed methods, including the OT-Former, CT-Former, and DAM modules, and to assess the impact of various parameters within the network.Overall, the current study presents a novel heterogeneous CD network based on the transformer framework. Within this network, OT-Former is proposed to achieve the adaptive generation of optical-signal tokens from diverse remote sensing images. Moreover, the CT-Former utilizes these optical-signal tokens as a guide to facilitate interaction with patch tokens for the learning of change rules. Additionally, DAM modules were embedded into the network to effectively extract the difference information. An extremely limited number of samples were utilized only for training in the experiments. Remarkably, the proposed method outperformed the existing state-of-the-art methods, achieving a significantly advanced performance in heterogeneous CD.  
      关键词:remote sensing;heterogeneous images;change detection;multimodal analysis;deep learning;Transformer   
      97
      |
      58
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 48420468 false
      发布时间:2024-02-29
    • ZHOU Chengle,SHI Qian,LI Jun,ZHANG Xinchang
      Vol. 28, Issue 1, Pages: 105-120(2024) DOI: 10.11834/jrs.20232600
      Spectral-frequency domain attribute pattern fusion for hyperspectral image change detection
      摘要:HyperSpectral Imagery (HSI) is a three-dimensional cube data that combines spatial imagery and spectral information, which introduces increased conveniences to the accurate interpretation of observation information of ground coverings. However, high-dimensional nonlinear data processing for the HSI Change Detection (HSI-CD) task encounters challenges. Therefore, an HSI-CD method based on Spectral-Frequency Domain Attribute Pattern Fusion (SFDAPF) is introduced to gradually quantify the spectral representation of pixel attribute patterns. Specifically, a Saliency Enhancement (SE) strategy for pixel attribute patterns based on Fourier transform theory is developed to improve the separability between pixel attribute patterns in the current work. The proposed SFDAPF method comprises four components as follows.First, a gradient correlation-based spectral absolute distance (GCASD) is designed in this paper. Therefore, the attribute patterns of pixel pairs in bitemporal HSI can be gradually quantified from the aspect of spectral information representation. Then, an SE strategy of attribute patterns of pixel pairs is proposed in accordance with Fourier transform theory, which improves the separability of attribute patterns of changing and non-changing pixel pairs in terms of global spatial information utilization. Next, the saliency level and GCASD per pixel are fused to obtain the comprehensive discrimination value of change detection. Finally, the binarization results of the bitemporal HSI-CD are obtained in accordance with the false alarm threshold.The proposed SFDAPF method is applied to two open-source bitemporal HSI datasets (i.e., River and Farmland datasets). Experimental results show that the proposed SFDAPF method can outperform the traditional and state-of-the-art HSI-CD methods. For the River dataset, compared with the traditional methods, the SFDAPF method in this paper introduces the local context information of the pixel in the calculation stage of the GCASD and adopts the global SE strategy, which is effective in reducing false alarms. Compared with the state-of-the-art methods, the SFDAPF method in this paper achieves the highest accuracy for most of the performance evaluation indicators. For the Farmland dataset, the AA, Kappa, F1, IoU, and OA indicators of the SFDAPF method in this paper have reached the highest accuracy, which is 0.01985, 0.05653, 0.01474, 0.02798, and 0.02187 higher than the second highest accuracy. In addition, the OAu (0.97500) and OAc (0.96766) indicators of the SFDAPF method did not achieve the highest accuracy. However, they were only 0.00673 and 0.01237 lower than the highest accuracy, which can be called slightly lower than the highest accuracy. Therefore, the experiments verified the effectiveness of the proposed SFDAPF method in the HSI-CD task.The proposed SFDAPF method generally considers the representation of spectral information and the utilization of neighborhood spatial information, thus promoting the overall accuracy of HSI-CD. However, the proposed SFDAPF method only considers the single-window eight-connected neighborhood in the spectral characterization stage and the magnitude features represented in the frequency domain. Therefore, future research work should further explore the contribution of dual-window spectral information representation and phase information of frequency domain representation to HSI-CD task.  
      关键词:hyperspectral image;change detection;image fusion;feature extraction;saliency analysis;fourier transform   
      184
      |
      222
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 41049879 false
      发布时间:2024-02-29
    • TU Kun,XIONG Fengchao,HOU Xueqiang
      Vol. 28, Issue 1, Pages: 121-131(2024) DOI: 10.11834/jrs.20243119
      Low-rank tensor embedded deep neural network for hyperspectral image denoising
      摘要:Spectral imagery has emerged as a powerful tool with widespread applications across various fields. This tool’s unique ability to identify materials continuous narrow bands has made it invaluable for tasks such as environmental monitoring, resource management, and agricultural early warning. However, the hyperspectral imagery utility is often compromised by the presence of noise from factors such as equipment errors and atmospheric conditions. This noise poses a significant challenge to the accuracy of subsequent analytical tasks, requiring the development of effective hyperspectral image denoising techniques. The spatial, spectral, and spatial-spectral joint correlations observed in hyperspectral images indicate that clean hyperspectral images occupy a low-dimensional subspace. This characteristic can be effectively characterized by low-rank and sparse representations. Consequently, a considerable body of research has been dedicated to exploring denoising methods based on such representations to enhance hyperspectral image quality. On the one hand, while deep learning methods offer the advantage of directly extracting prior information from data, they often exhibit low efficiency in the utilization of physical knowledge specific to hyperspectral images, such as their inherent low-rank nature. On the other hand, model-based techniques require the manual setting of priors and intricate parameter tuning, presenting a challenge in terms of practicality and adaptability. The objective of this study is to address the existing challenges in hyperspectral image denoising by proposing a novel approach that combines the strengths of deep learning- and model-based methods. The proposed methodology leverages the spatial-spectral low-rank characteristics inherent in hyperspectral images and embeds the low-rank tensor decomposition module into the U-Net for enhanced denoising. The low-rank tensor decomposition module is based on CP decomposition, generates rank-one vectors, and reconstructs low-rank tensors through operations like global pooling and convolution. The low-rank tensor decomposition module is integrated with the U-net architecture to represent shallow features as low-rank tensors. This strategy enables the model to capture spatial-spectral low-rank characteristics comprehensively, thereby significantly enhancing its denoising capabilities. Experimental evaluations, encompassing both simulated and real data, validate the efficacy of the proposed low-rank deep neural network method. Across a noise standard deviation range of [0—95], the algorithm achieves a peak signal-to-noise ratio of 41.02 dB and a structural similarity index of 0.9888. Empirical results underscore the superiority of the proposed low-rank deep neural network method over alternative approaches in terms of denoising performance for hyperspectral images. By effectively leveraging the spatial-spectral low-rank characteristics intrinsic to hyperspectral images, this methodology presents a robust solution for enhancing the accuracy of hyperspectral imagery in diverse applications. The amalgamation of low-rank tensor decomposition with deep learning techniques not only addresses existing challenges but also opens up promising avenues for future research in hyperspectral image processing, paving the way for improved methodologies and innovative solutions. The comprehensive exploration of this combined approach provides valuable insights and contributes to the evolving landscape of hyperspectral image analysis and enhancement.  
      关键词:hyperspectral image denoising;deep neural network;low-rank tensor representation;knowledge-driven deep learning;CP decomposition;U-Net   
      77
      |
      125
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 48419653 false
      发布时间:2024-02-29
    • HE Ke,SUN Weiwei,HUANG Ke,CHEN Binjie,YANG Gang
      Vol. 28, Issue 1, Pages: 132-141(2024) DOI: 10.11834/jrs.20232505
      Multifeature deep subspace clustering for hyperspectral band selection
      摘要:Hyperspectral Images (HSIs) contain abundant spectral information of ground objects through dozens or even hundreds of contiguous narrow bands with a high spatial resolution. However, HSIs are plagued by strong correlation between high-dimensional bands, which increases the difficulty in processing and applications of HSIs. Therefore, dimensionality reduction is one of the important steps of hyperspectral preprocessing. Band selection can effectively preserve the spectral significance of HSIs; thus, it is broadly used for dimensionality reduction. Unfortunately, the existing hyperspectral band selection methods typically consider inter-band relationships in a linear perspective while only partially focusing on multiscale information and demonstrating susceptibility to noise, resulting in poor performance of the subset of bands selected by existing methods. This paper proposes a multifeature deep subspace clustering for HSI band selection to overcome the above problems.MFDSC embeds the self-expression layer into the autoencoder to learn subspace self-expression coefficients, which considers the interaction of spatial and spectral information and explores the inter-band relationship with a nonlinear perspective. In addition, this paper couples the spatial-spectral attention and multifeature extraction modules with DSC to further reduce the interference of outliers and improve the learning capability of latent representation, thus enhancing the accuracy of self-expression coefficients. MFDSC starts with a spatial–spectral attention module to reweight the HSI to suppress useless information such as noise. Afterward, MFDSC uses convolution kernels of different sizes to extract features at different scales for encoding. Then, MFDSC learns the subspace coefficient matrix through the self-expression layer and reconstructs the original HSI using the decoder. Finally, the subspace coefficient matrix is partitioned using spectral clustering, and the bands closest to the cluster center in each class are calculated. These bands are the final results of band selection.In this paper, the proposed method is compared with the five state-of-the-art methods in a variety of experiments on three hyperspectral datasets (i.e., Indian Pines, PaviaU, and YRD datasets). The support vector machine, which adopts radial basis function as kernels, is employed as the classifier. Experimental results demonstrate that the proposed method can attain better performances than the comparison methods. Superior results are obtained when the number of bands reaches a certain number instead of using all bands. In addition, the computational efficiency of MFDSC is acceptable and significantly faster than that of DARecNet, which is a deep learning-based method.MFDSC considers the interference of noise and outliers on the self-expression performance of subspace clustering. Meanwhile, MFDSC nonlinearly learns the latent representation of data at different scales without deepening the network depth based on multiscale autoencoders. Thus, MFDSC can select a representative subset of bands and reduce the difficulty of subsequent applications.  
      关键词:hyperspectral remote sensing;dimensionality reduction;band selection;multi-feature;deep subspace clustering   
      77
      |
      77
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 48419651 false
      发布时间:2024-02-29
    • XIE Jinfeng,CHEN Tao
      Vol. 28, Issue 1, Pages: 142-153(2024) DOI: 10.11834/jrs.20232587
      Hyperspectral unmixing network considering spectral information and superpixel segmentation
      摘要:The pixel is the basic unit of remote sensing images. If a single pixel contains multiple types of covering ground objects, then it is called a mixed pixel. Hyperspectral unmixing aims to decompose the mixed pixels into several basic component units (endmembers) and obtain the proportion (abundance) of each endmember, which can improve the accuracy of remote sensing image classification and subpixel level target detection. Thus, research on this method promotes the development of hyperspectral remote sensing technology. Studies show that considering the spatial information in the process of hyperspectral unmixing can effectively improve the unmixing accuracy. However, most of the nonlinear unmixing networks based on deep learning only use the spectral information of images.A hyperspectral unmixing network considering spectral information and superpixel segmentation (SSUNet) is proposed on the basis of the supervised unmixing idea and one-dimensional convolutional neural network to maximize the spectral and spatial information of images. First, the original hyperspectral data should be processed using superpixel segmentation to obtain the superpixel segmentation data with spatial characteristics. Then, SSUNet is used to train and unmix the original hyperspectral and superpixel segmentation data. The loss function adds regularization constraint term based on the root mean square error to promote the sparsity of the unmixing abundance and generate closer unmixing results to the real value. The activation function of the network output layer is softmax, which yields output values of each output node within the range of [0, 1] and constrains their sum to 1, thus satisfying the two constraints of unmixing: the abundance nonnegative constraint and abundance sum-to-one constraint.Experiments on simulated datasets generated by the linear and nonlinear mixed models and the two real datasets show that the proposed network has higher unmixing accuracy and better robustness than the unmixing results of SUnSAL, SUnSAL-TV, SCLRSU, MTAEU, EGU-Net-pw, and 1DCNN. Three Gaussian noises with different SNR levels (20, 30, and 40 dB) are added to the simulated dataset. The proposed network can achieve the best unmixing results at all SNR levels, and the network also achieves high unmixing accuracy with the increase in SNR. In addition, the influence of the change of w value on the unmixing result of the simulated datasets under different SNR is verified. The experimental results show that when the value range of w is [3, 13], the RMSE value does not change substantially, and the best value of w is 5. Experiments on real datasets show that SSUNet can still achieve the best unmixing results in complex real scenes.The SSUNet network uses the dual-branch structure to mine the features of the original image data and the superpixel segmentation data with spatial features. This network also utilizes the fusion layer to fuse the features and improve the unmixing accuracy of the model. Experiments on simulated and real hyperspectral datasets show that the proposed network has high accuracy.  
      关键词:hyperspectral images;hyperspectral unmixing;spectral and spatial information;superpixel segmentation;deep learning;convolutional neural network   
      198
      |
      79
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 39019500 false
      发布时间:2024-02-29

      Hyperspectral Object Identification

    • HANG Renlong,SUN Yu,LIU Qingshan
      Vol. 28, Issue 1, Pages: 154-167(2024) DOI: 10.11834/jrs.20232635
      Joint classification of hyperspectral and LiDAR data based on inter-modality match learning
      摘要:Several excellent models for joint classification of hyperspectral image and LiDAR data, which were designed on the basis of supervised learning methods such as convolutional neural networks, have been developed in recent years. Their classification performance depends largely on the quantity and quality of training samples. However, when the distribution of ground objects becomes increasingly complex and the resolutions of remote sensing images grow increasingly high, obtaining high-quality labels only with limited cost and manpower is difficult. Therefore, numerous scholars have made efforts to learn features directly from unlabeled samples. For instance, the theory of autoencoder was applied to multimodal joint classification, achieving satisfactory performance. Methods based on the reconstruction idea reduce the dependence on labeled information to a certain extent, but several problems that must be settled still exist. For example, these methods pay more attention to data reconstruction but fail to guarantee that the extracted features have sufficient discriminant capability, thus affecting the performance of joint classification.This paper proposes an effective model named Joint Classification of Hyperspectral and LiDAR Data Based on Inter-Modality Match Learning to address the aforementioned issue. Different from feature extraction models based on reconstruction idea, the proposed model tends to compare the matching relationship between samples from different modalities, thereby enhancing the discriminative capability of features. Specifically, this model comprises inter-modality matching learning and multimodal joint classification networks. The former is prone to identify matching of the input patch pairs of hyperspectral image and LiDAR data; therefore, reasonable construction of matching labels is essential. Thus, spatial positions of center pixels in cropped patches and KMeans clustering methods are employed. These constructed labels and patch pairs are combined to train the network. Notably, this process does not use manual labeled information and can directly extract features from abundant unlabeled samples. Furthermore, in the joint classification stage, the structure and trained parameters of matching learning network are transferred, and a small number of manually labeled training samples are then used to finetune the model parameters.Extensive experiments were conducted on two widely used datasets, namely Houston and MUUFL, to verify the effectiveness of the proposed model. These experiments include comparison experiments with several state-of-the-art models, hyperparameter analysis experiments, and ablation studies. Take the first experiment as an example. Compared with other models, such as CNN, EMFNet, AE_H, AE_HL, CAE_H, CAE_HL, IP-CNN, and PToP CNN, the proposed model can achieve higher performance on both datasets with OAs of 88.39% and 81.46%, respectively. Overall, the proposed model reduces the dependence on manually labeled data and improves the joint classification accuracy in the case of limited training samples. A superior model structure and additional testing datasets will be explored in the future to make further improvements.  
      关键词:remote sensing image;hyperspectral image;LiDAR data;deep learning;Match Learning;Joint Classification   
      304
      |
      172
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 37558643 false
      发布时间:2024-02-29
    • LIU Qian,WU Zebin,XU Yang,ZHENG Peng,ZHENG Shangdong,WEI Zhihui
      Vol. 28, Issue 1, Pages: 168-186(2024) DOI: 10.11834/jrs.20243292
      Hyperspectral remote sensing image classification based on multidirectional adaptive aware network
      摘要:Hyperspectral remote sensing can realize the simultaneous acquisition of spectral data and spatial images for the observation scene to satisfy the detection requirements for the composition and morphology of the target object, which is widely used in the fields of precision agriculture, geological survey, environmental monitoring, biology, and medicine. In recent years, benefiting from the powerful representation capabilities of spectral-spatial information, the HyperSpectral Image (HSI) classification methods based on convolutional neural networks have demonstrated superior classification performances to traditional methods. However, the convolutional operation, performed within fixed square windows following certain rules, encounters difficulties in adapting spectral-spatial feature extraction for different objects and spatial distributions and introduces irrelevant information from other categories, leading to information diffusion between different categories and misclassification of edge pixels. This paper proposes a hyperspectral image classification method based on multidirectional adaptive awareness to address this problem. This method integrates the spatial contextual information of different neighborhood ranges to improve the local spectral-spatial modeling capability of the model. First, the full-window filter of the regular convolution is split into side-window filters with different orientations to design the Side-Window Convolution (SWC) acquiring the directional spectral-spatial features. Therefore, multiple side-window filter kernels of different directions are integrated into a unified convolution architecture to construct the spectral-spatial separable multidirectional convolution (S3MDC). The direction-adaptive aware module (DAAM) is designed to assist S3MDC and build a multidirectional adaptive aware network (MDAAN) with dense connections for HSI classification. MDAAN can adaptively learn multidirectional spatial-spectral features, improve the representation of complex spectral-spatial structures, and compensate for the deficiency that SWC only captures spectral-spatial relationships in a single direction. Experiments were conducted on three public datasets, including Indian Pines, University of Pavia, and Kennedy Space Center. MDAAN achieves the overall highest classification accuracy of 97.67% (Indian pines, 5%/class), 99.40% (University of Pavia, 1%/class), and 99.64% (Kennedy Space Center, 5%/class), which is superior to other deep learning methods, verifying the effectiveness of the proposed model. First, compared with other deep learning methods, MDAAN can provide better classification performance. Second, MDAAN generalizes better with the small number of training samples to prove the stability of its classification performance. Finally, the ablation analysis of different convolutional models and DAAM demonstrates the effectiveness and necessity of S3MDC and the adaptive aware mechanisms for multidirectional features.  
      关键词:remote sensing;hyperspectral image;deep learning;multi-directional adaptive awareness;spectral-spatial structure modeling;spectral-spatial classification   
      101
      |
      96
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 48419185 false
      发布时间:2024-02-29
    • YU Yao,SU Hongjun,TAO Yang
      Vol. 28, Issue 1, Pages: 187-202(2024) DOI: 10.11834/jrs.20221704
      Dynamic selection algorithm for collaborative representation of hyperspectral remote sensing based on joint spatial information
      摘要:Ensemble learning has recently attracted considerable attention for hyperspectral image analysis. This model integrates multiple base classifiers for joint decision making, which is better than using a base classifier. Ensemble learning includes static and dynamic classifier ensembles. In the static ensemble method, the same classifier combination scheme is selected for testing sample. However, this method ignores the difference in classifier performance for each testing sample. Considering the features of testing sample, the best classifier is selected adaptively in dynamic ensemble methods. Therefore, this classifier can generally achieve better performance than static ensemble methods for hyperspectral image classification. However, numerous dynamic ensemble methods only consider the spectral information of the validation and training samples, ignoring the rich spatial information of hyperspectral images.A Variable K-neighborhood and Spatial information algorithm (VKS) is proposed in this paper to further improve the accuracy and reliability of hyperspectral image classification. Firstly, the VKS algorithm comprehensively considers the accuracy and similarity of the classifier to adaptively adjust the K-neighborhood of the testing sample, increasing the reliability and flexibility of the region of competence setting. Thus, the testing samples with good spectral discrimination performance are preferentially classified. The label information of spatial neighborhood samples is used for predicting the testing samples with poor spectral discrimination performance. A fixed window is designed to provide local spatial information in hyperspectral images. However, fixed windows cannot reveal the complex and changeable morphological characteristics of ground objects. An adaptive window that can effectively reflect complex spatial information is proposed to capture the complex and changeable spatial structure in a hyperspectral image, and a variable K-neighborhood with a shape-adaptive (VKSA) algorithm is further designed.The Purdue Campus, Indian Pines, and Salinas hyperspectral remote sensing data sets are used to design experiments and verify the performance of the proposed VKS and VKSA algorithms. Four state-of-the-art methods, namely, majority voting, overall local accuracy, modified local accuracy, and multiple classifier behavior, are used to quantify the classification accuracy. Experimental results demonstrate that the VKS and VKSA algorithms outperform static ensemble methods and three classic dynamic ensemble methods in overall classification accuracy. Moreover, the VKSA algorithm with an adaptive window perform better than the VKS algorithm with a fixed window.  
      关键词:hyperspectral remote sensing;dynamic selection;shape-adaptive neighborhood;collaborative representation;image classification   
      658
      |
      149
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 29322043 false
      发布时间:2024-02-29
    • WANG Le,PENG Jiangtao,CHEN Na,SUN Weiwei
      Vol. 28, Issue 1, Pages: 203-218(2024) DOI: 10.11834/jrs.20242326
      Hyperspectral classification algorithm based on covariance pooling and cross scale feature extraction
      摘要:The deep convolution neural network algorithm has achieved excellent performance in hyperspectral image classification. However, these deep learning algorithms generally use first-order pooling operation, which ignores the correlation between different spectral bands. Thus, obtaining high-order statistical discriminant features is difficult. In addition, using these algorithms to choose the optimal window size and capture different receptive field information is complicated. This paper proposes a hyperspectral classification method combining covariance pooling and cross-scale feature extraction to solve the aforementioned problems. This method aims to automatically extract the complementary and discriminative information of different scales and exploit the first- and second-order pooling features to improve the classification performance.A covariance pooling and cross-scale feature extraction method is proposed for hyperspectral image classification. In this method, a cross-scale adaptive feature extraction module is designed. This module can automatically combine multiscale feature information and obtain complementary information of different visual fields, avoiding the scale selection problem. Furthermore, the first- and second-order statistics combined with spatial-spectral information are obtained using the joint pooling operation of average and fast covariance pooling. Finally, the first- and second-order pooled features are fused for classification.A total of 5%, 5%, and 1% labeled samples were randomly selected from three public hyperspectral datasets, namely, Indian pines, Houston University, and Pavia University, respectively. The overall classification accuracy of the proposed algorithm reached 97.63%, 98.48%, and 98.21%, and the classification performance was better than the state-of-the-art deep learning methods.Cross-scale feature extraction considers the complementary spatial-spectral information between different scales to obtain additional adaptive feature information. Combining fast covariance and average pooling, the discriminant features are obtained by pooling feature fusion to obtain superior classification results.  
      关键词:hyperspectral image classification;covariance pooling;Multiscale;Feature fusion;convolution neural networks   
      76
      |
      105
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 48420063 false
      发布时间:2024-02-29
    • JIN Yao,DONG Yanni,DU Bo
      Vol. 28, Issue 1, Pages: 219-230(2024) DOI: 10.11834/jrs.20232286
      Fast self-training based on spatial-spectral information for hyperspectral image classification
      摘要:Hyperspectral image classification has been a popular issue in the field of hyperspectral image interpretation. The prominent problem in hyperspectral image classification currently involves the considerably time-consuming and expensive manual acquisition of labeled samples for hyperspectral images in practical applications. This problem leads to a sparse number of training samples and increases the difficulty of obtaining good classification results. Self-training methods are widely used in hyperspectral image classification to solve the difficulty in labeled sample acquisition in hyperspectral image classification. Traditional self-training methods mostly use spectral information to classify unlabeled samples and then utilize the expanded labeled data set to iteratively train the classifier to complete the classification task. In this model, the spatial information provided by the hyperspectral images is ignored, resulting in poor classification accuracy. Simultaneously, the classification of unlabeled data must be completed once during each iteration, resulting in a significant time cost. Therefore, a fast self-training method based on spatial–spectral information is proposed in this paper for hyperspectral image classification to address the above problems.FST-SS (Fast Self-training based on Spatial-Spectral information) supplements the spatial information in hyperspectral images by exploiting the consistency of the spatial distribution of features in the hyperspectral images. Instead of using the classifier to classify unlabeled samples, this approach uses the spatial–spectral information to filter unlabeled data and extend the labeled samples during the iterative process. The spatial nearest neighbors are first selected using a spatial domain patch for the initial labeled sample. The spatial nearest neighbors are then filtered using an adaptive threshold to obtain the spatial–spectral nearest neighbors to be labeled. Finally, the classifier is trained to complete the classification task based on the expanded labeled samples.This paper compares FST-SS with the supervised classification algorithm 1NN, the semi-supervised classification algorithm Star-SVM, Tri-Training, ST-DP, and LeMA on two real hyperspectral datasets to demonstrate the effectiveness of FST-SS. Experimental results show that the overall classification accuracy reaches 93.17% and 95.43% when 2 and 10 training samples, respectively, are selected for each class in the Washington DC Mall subimage dataset. The overall classification accuracy in Indian Pines dataset reaches 59.75% and 86.13%, which is a significant improvement compared with the comparison algorithm.The FST-SS algorithm uses the spatial–spectral information provided by hyperspectral images to label unlabeled samples by combining the ideas of self-training methods. Compared with the conventional self-training methods, instead of using a classifier for classification, FST-SS uses the spatial–spectral information to filter the unlabeled samples directly, which markedly improves the computational efficiency of the algorithm.  
      关键词:hyperspectral image classification;semi-supervised classification;spatial-spectral information;self-training   
      83
      |
      99
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 48419186 false
      发布时间:2024-02-29
    • YU Chunyan,XU Mingyang,SONG Meiping,HU Yabin,CHANG Chein-I
      Vol. 28, Issue 1, Pages: 231-246(2024) DOI: 10.11834/jrs.20232580
      Unsupervised domain adaptive classification for hyperspectral remote sensing by adversary coupled with distillation
      摘要:Unsupervised Domain Adaptive (UDA) classification aims to categorize the target domain scenes without labeled samples using knowledge from the source domain data with the labeled samples. Thus, UDA classification is one of the important cross-scene classification methods in the field of Hyperspectral Image Classification (HSIC). The existing domain adaptive classification methods for hyperspectral remote sensing data mainly utilize the adversarial training mode to achieve the feature alignment between the target and source domains. The popular UDA approach with the local alignment condition of dual domains generates acceptable classification accuracy. However, the key issue of the sufficient transfer of the source domain knowledge to the target domain is not considered. An unsupervised domain adaptation classification by adversary coupled with distillation is proposed in this paper for the unsupervised HSIC to effectively extract and transfer source domain knowledge. In the proposed framework, the dense-base network with convolutional block attention module is presented to extract abundant features for the representation of the source and target domain categories. In the source domain training, a self-distillation learning schema is adopted to reduce the class-wised difference by matching the predictive distribution of the same class samples. The self-distillation regularization constraint is increased between the samples of the same category in the source domain to reduce the intraclass difference of the classification subspace and improve the knowledge expression accuracy of the source domain classification model. Thus, the capability of the adaptive classification model to refine the source domain supervision knowledge is improved. In addition, a novel mechanism of adversarial training coupled with distillation knowledge is presented to guarantee the complete transfer of source domain knowledge to the target domain scene with feature alignment. Moreover, dual classifiers are employed in the adversarial training process to eliminate the prediction effect of the confused samples. The maximum and minimum discrepancies of the dual classifiers during the adversarial training rapidly promote the feature alignment without confusion. Thus, knowledge distillation is conducted to improve the recognition capability of the network in the domain while ensuring the complete transfer of hyperspectral source domain knowledge in the feature alignment process to improve the knowledge acquisition capability of the model in the target domain. Finally, the unsupervised classification of HSIs in the target domain is completed after the knowledge transfer. The experiments for HSI cross-scene image classification are conducted on four hyperspectral remote sensing scene datasets, including Pavia University, Pavia Center, Houston 2013, and Houston 2018. Results demonstrate that the proposed model is superior to other hyperspectral domain adaptive methods. Under the same sample conditions, the classification accuracy achieves 91.75% (Pavia University to Pavia Center), 74.41% (Pavia Center to Pavia University), 70.68% (Houston 2013 to Houston 2018), and 67.76% (Houston 2018 to Houston 2013). In addition, the ablation study illustrates that the final classification accuracy of the unsupervised HSIC is improved with the self-distillation and the distillation loss in the adversarial training model. The parameters with different weights and temperatures are analyzed in the experiments with variations of values. The validity of the method is verified by all the mentioned experimental results and analyses.  
      关键词:hyperspectral remote sensing;image classification;domain adaptation;knowledge distillation;generative adversarial networks   
      421
      |
      117
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 39688458 false
      发布时间:2024-02-29
    • SU Han,CHEN Na,PENG Jiangtao,SUN Weiwei
      Vol. 28, Issue 1, Pages: 247-265(2024) DOI: 10.11834/jrs.20232492
      Hyperspectral image classification based on three branch network with grouped spatial-spectral attention
      摘要:Hyperspectral images have abundant spatial and spectral information. Numerous hyperspectral classification algorithms focus on the extraction and maximization of spatial and spectral information. Deep feature extraction networks generally extract spectral-spatial features using single-branch serial or double-branch parallel structures. However, single-branch structures may lead to mutual interference between features of spectral and spatial dimensions, and double-branch parallel structures tend to ignore the correlation between spatial and spectral features. This paper proposes a three-branch grouped spatial-spectral attention network (TGSSAN) to consider the differences and correlations between spatial and spectral features. TGSSAN can extract independent spectral-spatial features while preserving their correlation.This paper proposes the TGSSAN, which has three parallel branches (i.e., spectral, spatial, and spectral-spatial branches). These branches can separately extract spectral, spatial, and spatial-spectral features. Different attention blocks are designed in three branches to enhance the discriminative capability of features. In particular, a grouped spatial-spectral attention mechanism is proposed in the spectral-spatial branch to obtain spatial and spectral attention simultaneously. Finally, three branch features are fused for classification.In the experiment, the proposed TGSSAN algorithm is compared with some advanced deep learning algorithms, such as SSRN, FDSSC, DBMA, DBDA, HResNetAM, and A2S2KResNet. The performance of different algorithms is evaluated on five hyperspectral data sets. Experimental results show that the proposed algorithm achieved superior classification performance on IP, PU, SA, HU, and HHK datasets. In particular, the proposed algorithm achieves higher classification accuracy despite limited training samples compared with the existing advanced algorithms.The TGSSAN method proposed in this paper improves the shortcomings of the single-branch serial and double-branch parallel structures for continuous extraction of spectral-spatial features, which can effectively extract image spectral-spatial feature information. The three attention blocks designed in this paper namely, spectral, grouped spatial-spectral, and spatial attention modules, can effectively enhance the feature discrimination capability and further improve the classification performance.  
      关键词:hyperspectral image classification;attention mechanism;three-branch network;deep network   
      59
      |
      85
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 48420059 false
      发布时间:2024-02-29

      Hyperspectral Remote Sensing Applications

    • LOU Anjun,HE Zhi,XIAO Man,LI Xinyuan
      Vol. 28, Issue 1, Pages: 266-279(2024) DOI: 10.11834/jrs.20232246
      Wetland classification based on Sentinel-2 and 3D multisource domain self-attention model
      摘要:Accurate wetland classification methods can quickly grasp the spatial-temporal variation characteristics of wetlands and play an important role in wetland research. Considering the limitation of the existing wetland classification method based on few-shot learning to the use of target or single-source domain dataset, this paper proposes a 3D multisource domain self-attention few-shot learning (3D-MDAFSL) model. First, combining the advantages of convolution and attention mechanism, a 3D feature extractor based on self-attention mechanism and deep residual convolution is designed. Then, the conditional adversarial domain adaptation strategy is used to achieve multisource domain distribution alignment, and few-shot learning is performed separately in each domain. Finally, the features extracted by the trained model are imputed to the K-nearest neighbor classifier to obtain classification results. Results show that compared with the framework without feature extraction, the 3D feature extractor improves the overall accuracy by approximately 6.79%. When using multisource domain datasets, the overall accuracy of the 3D-MDAFSL model for the Sentinel-2 wetland dataset in Zhongshan City can reach 93.52%, which is a significant improvement compared with the existing algorithms. The 3D-MDAFSL model proposed in this paper has good application value in the high-precision extraction and classification of wetland features.  
      关键词:remote sensing;Few-shot;wetland classification;Multi-source domain adaption;Self-attention   
      196
      |
      131
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 40070315 false
      发布时间:2024-02-29
    • HU Shunshi,YANG Bin,HUANG Ying,CEN Yi,QI Wenchao
      Vol. 28, Issue 1, Pages: 280-292(2024) DOI: 10.11834/jrs.20222054
      Fine classification of vegetable crops covered with different planting facilities using UAV hyperspectral image
      摘要:With large-scale and high-output values, the vegetable industry of China is a pillar industry to promote the income increase of farmers and the development of rural agricultural economy. Rapidly and accurately obtaining the structural information of vegetable crop planting is of considerable importance for agricultural modernization, automation, and precision. With the advantages of fast mobility, flexibility, and image-spectrum merging, Unarmed Aerial Vehicle (UAV) hyperspectral remote sensing has wide prospects in fine classification of crops. However, vegetable crop planting scales and modes have considerable variations, and the fragmentation of agricultural landscape is high in China. The vegetable crops are also affected by the coverage of plastic film, greenhouse, and bird proof net, which easily produced the mixed spectral effect in UAV hyperspectral images and also introduced considerable challenges to the fine classification of vegetable crops.Hyperspectral images of Gaoqiao scientific research base of Hunan Academy of Agricultural Sciences were obtained by UAV. The field survey revealed that the area contains 14 ground feature categories, including eggplant, towel gourd, rice, pepper, and tomato. Support Vector Machine (SVM) is widely used in crop classification due to low requirements for data and excellent generalization capability. Meanwhile, deep convolution neural network structures can automatically learn the abstract features of images and obtain high-level and rich semantic information of samples to successfully complete the classification task. Therefore, SVM and Deep Learning (DL) methods were applied to the classification of vegetable crops in this study. Unlike other hyperspectral classification verification experiments that randomly select training sets, training and test samples were manually selected in this study to reduce the spatial correlation between training and test sets, and the performance of different classification methods was evaluated using confusion matrix.The results showed that based on hyperspectral images obtained by UAVs, the average overall accuracy of vegetable crop classification using SVM and DL methods is 78.03% and 90.75%, respectively, and the average Kappa coefficients are 0.7359 and 0.8887, respectively. Compared with the SVM methods, the fine classification effects obtained by the DL methods are more ideal. This finding is attributed to the effective extraction of spectral and spatial feature information from the image using the three-dimensional convolutional neural network and the convolutional neural network with attention mechanism, thus demonstrating a superior performance in the classification of vegetable crops. The spatial texture characteristics of vegetable crops are observed on large-scale plots, while they are various on small-scale plots. Thus, using different DL methods for the classification of vegetable crops on different scale plots is appropriate.Vegetable crops under different planting facilities were classified in this study using UAV hyperspectral images. Under the influence of complex backgrounds such as plastic films, bird nets, and greenhouses, satisfactory performance was still achieved using SVM and DL methods, which can provide technological support for the modernization, automation, and refinement of regional vegetable crop management.  
      关键词:fine classification;vegetable crops;Unmanned Aerial Vehicle (UAV);hyperspectral;greenhouses;mulch film   
      880
      |
      149
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 25632567 false
      发布时间:2024-02-29
    • LIU Qian,WANG Mengdi,GUO Long,WANG Ran,JIA Zhongfu,HU Xianjun,TANG Qiankun,SHI Tiezhu
      Vol. 28, Issue 1, Pages: 293-305(2024) DOI: 10.11834/jrs.20221805
      Mapping of soil organic carbon density at farmland scale based on airborne hyperspectral images
      摘要:Accurate monitoring of Soil Organic Carbon Density (SOCD) is important for regulating soil carbon sinks and rationally using soil resources. Airborne hyperspectral images provide important data sources for SOCD mapping. The noise in the spectrum affects the accuracy of SOCD estimation because airborne hyperspectral images are easily affected by external factors during data collection. A set of technical processes that are suitable for airborne hyperspectral data processing is still lacking. Therefore, this study aims to investigate the technical process of SOCD estimation based on airborne hyperspectral images. The original spectra are preprocessed by First Derivative (FD) and Continuum Removal (CR) transform. Genetic Algorithm (GA) was used to select the feature bands. Different regression methods, such as Partial Least-Squares Regression (PLSR), Multiple Linear Regression (MLR), Support Vector Machine (SVM), and Artificial Neural Network (ANN), were used to estimate SOCD. Results showed that the accuracy of SOCD prediction for original, FD, and CR spectra was improved after feature band selection by GA. With the feature bands of original spectra, the R² of SOCD predicted by PLSR, MLR, SVM, and ANN are 0.672, 0.621, 0.551, and 0.678, respectively. The range of R² are 0.452—0.593 and 0.332—0.602 with FD and CR feature bands, respectively, which demonstrate large errors. The feature bands of the original spectrum were used in this study for SOCD mapping. The SOCD predicted by four regression models has a highly similar trend in space and is similar to the SOCD measured value. The points with large absolute errors mostly occur near the edges of the sampling points.  
      关键词:soil organic carbon density;airborne hyperspectral images;genetic algorithm;digital soil mapping   
      823
      |
      144
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 28025500 false
      发布时间:2024-02-29

      Data Paper

    • LIU Yuan,ZHENG Xiangtao,LU Xiaoqiang
      Vol. 28, Issue 1, Pages: 306-319(2024) DOI: 10.11834/jrs.20233283
      Hyperspectral scene classification dataset based on Zhuhai-1 images
      摘要:Hyperspectral remote sensing is a key technology for remotely obtaining the physical parameters of ground objects and realizing fine identification. It can not only get geometrical properties of the target scenes but also obtain radiance that reflects the characteristics of ground objects. With the development of hyperspectral remote sensing data to unprecedented spatial, spectral, temporal resolution and large data volume, how to adapt to the requirements of massive data and achieve efficient and rapid processing of hyperspectral remote sensing data has become the current research focus. Researchers are introducing scene classification into hyperspectral image classification, integrating the spatial and spectral information to obtain semantic information oriented to larger observation units. However, almost all existing multispectral/hyperspectral scene classification datasets have a number of limitations, including inconsistent spectral and spatial resolutions or spatial resolutions too large to meet the needs of fine-grained classification. Based on the hyperspectral images of Xi’an taken by the “Zhuhai-1” constellation, we combine the result of unsupervised spectral clustering and Google Earth to establish a hyperspectral satellite image scene classification dataset named HSCD-ZH (Hyperspectral Scene Classification Dataset from Zhuhai-1). It consists of 737 images divided into six categories: urban, agriculture, rural, forest, water, and unused land. Each image with a size of 64×64 pixels consists of 32 bands covering the wavelength in the range of 400—1000 nm. In addition, we conduct spatial-based and spectral-based experiments to analyze the performance of existing datasets, and the benchmark results are reported as a valuable baseline for subsequent research. We choose false-color image for the spatial-based experiments and then use popular deep and non-deep learning scene classification techniques. In the experiments based on spectral, the spectral vectors at the pixel are directly used as local spectral features, and BoVW, IFK, and LLC are used to encode them to generate global representations for the scene. Using SVM as the classifier, the optimal overall classification achieved by the two experiments on the proposed dataset is 92.34% and 88.96%, respectively. Considering that those methods have a large amount of information loss, we cascade the features extracted by the two approaches to generate spatial-spectral features. The highest overall accuracy obtained reaches 94.64%, which is the highest improvement in overall accuracy compared to the other datasets. We construct HSCD-ZH by effectively exploiting both spectral and spatial features of hyperspectral images, selecting various scenes that either have representative spectral compositions, clear spatial textures, or both. It has the advantages of big intraclass diversity, strong scalability, and adapting to satellite hyperspectral intelligent information extraction requirements. Both dataset and experiments can provide effective data support for remote sensing scene classification research in the hyperspectral field. Meanwhile, experiments can indicate that extracting features based on spatial or spectral misses a large amount of available information, and integrating the features extracted by the two methods can compensate for this deficiency. In our future work, we aim to expand the number of categories and images of HSCD-ZH and continue to explore algorithms for integrating spatial and spectral information that can accelerate the interpretation and efficient exploitation of hyperspectral scene cubes.  
      关键词:hyperspectral remote sensing;Zhuhai-1;scene classification;dataset;feature extraction   
      177
      |
      120
      |
      0
      <HTML>
      <L-PDF><Enhanced-PDF><Meta-XML>
      <引用本文> <批量引用> 48420466 false
      发布时间:2024-02-29
    0