Submit Manuscript
Peer Review
Editor WorkMost Downloaded
1
2026, 15(2): 387-408.
Abstract
4613KB
Deep learning is primarily used for target detection in Synthetic Aperture Radar (SAR) images; however, its performance heavily relies on large-scale labeled datasets. The detection performance of deep learning models degrades when applied to SAR data with varying distributions, hindering their real-world applicability. In addition, manual labeling of SAR data is costly. Hence, cross-domain learning strategies based on multisource information are being explored to address these challenges. These strategies can assist detection models in realizing cross-domain knowledge migration by integrating prior information from optical remote sensing images or heterogeneous SAR images acquired from different sensors. This paper focuses on cross-domain learning technologies within the deep learning framework. In addition, it provides a systematic overview of the latest research progress in this field and analyzes the core issues, advantages, and applicable scenarios of existing technologies from a methodological perspective. It outlines future research directions based on the law of technological evolution, aiming to offer theoretical support and methodological references to enhance the generalizability of target detection in SAR images.
Deep learning is primarily used for target detection in Synthetic Aperture Radar (SAR) images; however, its performance heavily relies on large-scale labeled datasets. The detection performance of deep learning models degrades when applied to SAR data with varying distributions, hindering their real-world applicability. In addition, manual labeling of SAR data is costly. Hence, cross-domain learning strategies based on multisource information are being explored to address these challenges. These strategies can assist detection models in realizing cross-domain knowledge migration by integrating prior information from optical remote sensing images or heterogeneous SAR images acquired from different sensors. This paper focuses on cross-domain learning technologies within the deep learning framework. In addition, it provides a systematic overview of the latest research progress in this field and analyzes the core issues, advantages, and applicable scenarios of existing technologies from a methodological perspective. It outlines future research directions based on the law of technological evolution, aiming to offer theoretical support and methodological references to enhance the generalizability of target detection in SAR images.
2
2026, 15(2): 409-440.
Synthetic Aperture Radar (SAR) is widely used in military and civilian applications, with intelligent target interpretation of SAR images being a crucial component of SAR applications. Vision-Language Models (VLMs) play an important role in SAR target interpretation. By incorporating natural language understanding, VLMs effectively address the challenges posed by large intraclass variability in target characteristics and the scarcity of high-quality labeled samples, thereby advancing the field from purely visual interpretation toward semantic understanding of targets. Drawing upon our team’s extensive research experience in SAR target interpretation theory, algorithms, and applications, this paper provides a comprehensive review of intelligent SAR target interpretation based on VLMs. We provide an in-depth analysis of existing challenges and tasks, summarize the current state of research, and compile available open-source datasets. Furthermore, we systematically outline the evolution, ranging from task-specific VLMs to contrastive-, conversational-, and generative-based VLMs and foundational models. Finally, we discuss the latest challenges and future outlooks in SAR target interpretation by VLMs.
Synthetic Aperture Radar (SAR) is widely used in military and civilian applications, with intelligent target interpretation of SAR images being a crucial component of SAR applications. Vision-Language Models (VLMs) play an important role in SAR target interpretation. By incorporating natural language understanding, VLMs effectively address the challenges posed by large intraclass variability in target characteristics and the scarcity of high-quality labeled samples, thereby advancing the field from purely visual interpretation toward semantic understanding of targets. Drawing upon our team’s extensive research experience in SAR target interpretation theory, algorithms, and applications, this paper provides a comprehensive review of intelligent SAR target interpretation based on VLMs. We provide an in-depth analysis of existing challenges and tasks, summarize the current state of research, and compile available open-source datasets. Furthermore, we systematically outline the evolution, ranging from task-specific VLMs to contrastive-, conversational-, and generative-based VLMs and foundational models. Finally, we discuss the latest challenges and future outlooks in SAR target interpretation by VLMs.
3
2026, 15(2): 583-604.
Research on target recognition using radar High-Resolution Range Profiles (HRRPs) is extensive and diverse in methodology. In particular, the application and development of deep learning to radar HRRP target recognition have enabled efficient, precise target perception directly from radar echoes. However, deep learning-based recognition networks rely on large amounts of training data. For non-cooperative targets, due to limited radar system parameters and rapid target attitude variations, acquiring adequate HRRP training samples that comprehensively cover target attitudes in advance is difficult in practice. Consequently, deep recognition networks are prone to overfitting and exhibit considerably degraded generalization capability. To address these issues, and given the ease of obtaining full-attitude electromagnetic simulation data for the target, this paper leverages simulated data as auxiliary information to mitigate the small-sample-size problem through data augmentation and cross-domain knowledge-transfer learning. For data augmentation, based on the analysis of differences in mean and variance between simulated and measured HRRPs within a given attitude-angle range, a linear transformation is applied to a set of simulated HRRPs spanning the same angular domain as a small set of measured HRRPs. This adjustment ensures that the simulated data’s mean and variance match the characteristics of the measured HRRPs, thereby achieving data augmentation that approximates the true distributional properties of HRRPs. Meanwhile, for cross-domain knowledge transfer learning, the proposed method introduces a domain alignment strategy based on generative adversarial constraints and a class alignment strategy based on contrastive learning constraints. These approaches draw the domain features of full-attitude simulation—strong discriminability and generalizability—closer to the measured domain features on a class-by-class basis, thereby further aiding learning from the measured domain data and leading to substantial improvements in few-shot recognition performance. Experimental results based on electromagnetic simulated and measured HRRP data for three and ten types of aircraft and ground vehicle targets, respectively, demonstrate that the proposed method yields superior recognition robustness compared with existing few-shot recognition methods.
Research on target recognition using radar High-Resolution Range Profiles (HRRPs) is extensive and diverse in methodology. In particular, the application and development of deep learning to radar HRRP target recognition have enabled efficient, precise target perception directly from radar echoes. However, deep learning-based recognition networks rely on large amounts of training data. For non-cooperative targets, due to limited radar system parameters and rapid target attitude variations, acquiring adequate HRRP training samples that comprehensively cover target attitudes in advance is difficult in practice. Consequently, deep recognition networks are prone to overfitting and exhibit considerably degraded generalization capability. To address these issues, and given the ease of obtaining full-attitude electromagnetic simulation data for the target, this paper leverages simulated data as auxiliary information to mitigate the small-sample-size problem through data augmentation and cross-domain knowledge-transfer learning. For data augmentation, based on the analysis of differences in mean and variance between simulated and measured HRRPs within a given attitude-angle range, a linear transformation is applied to a set of simulated HRRPs spanning the same angular domain as a small set of measured HRRPs. This adjustment ensures that the simulated data’s mean and variance match the characteristics of the measured HRRPs, thereby achieving data augmentation that approximates the true distributional properties of HRRPs. Meanwhile, for cross-domain knowledge transfer learning, the proposed method introduces a domain alignment strategy based on generative adversarial constraints and a class alignment strategy based on contrastive learning constraints. These approaches draw the domain features of full-attitude simulation—strong discriminability and generalizability—closer to the measured domain features on a class-by-class basis, thereby further aiding learning from the measured domain data and leading to substantial improvements in few-shot recognition performance. Experimental results based on electromagnetic simulated and measured HRRP data for three and ten types of aircraft and ground vehicle targets, respectively, demonstrate that the proposed method yields superior recognition robustness compared with existing few-shot recognition methods.
4
,
The threat from low-altitude targets to airspace security such as airport is increasing, making accurate detection and recognition essential for radar systems. High-quality measured radar datasets are crucial for advancing low-altitude target recognition. However, most existing public radar datasets for these targets consist of simulation data or short-range collected data, which have difficulty accurately reflecting and verifying radar target recognition performance in long-range scenarios. To overcome these limitations, this study creates a low-altitude target detection and recognition dataset based on Holographic Staring Radar (HSR), including measured data collection and recognition validation for typical low-altitude targets in outdoor environments. The dataset includes common targets such as multirotor unmanned aerial vehicles, sparrows, and large migratory birds, along with representative motion scenarios like hovering, circling, and radial flight. It also offers synchronized target micro-Doppler waterfall plots and radar-measured track information (including azimuth and elevation angles, radial velocity, and normalized signal-to-noise ratio), providing a data foundation for exploring the intrinsic link between target detailed features and motion states. Building on this, a multimodal adaptive feature fusion network is developed to extract and combine Doppler and kinematic features from different targets, demonstrating the dataset’s effectiveness in distinguishing various low-altitude targets.
The threat from low-altitude targets to airspace security such as airport is increasing, making accurate detection and recognition essential for radar systems. High-quality measured radar datasets are crucial for advancing low-altitude target recognition. However, most existing public radar datasets for these targets consist of simulation data or short-range collected data, which have difficulty accurately reflecting and verifying radar target recognition performance in long-range scenarios. To overcome these limitations, this study creates a low-altitude target detection and recognition dataset based on Holographic Staring Radar (HSR), including measured data collection and recognition validation for typical low-altitude targets in outdoor environments. The dataset includes common targets such as multirotor unmanned aerial vehicles, sparrows, and large migratory birds, along with representative motion scenarios like hovering, circling, and radial flight. It also offers synchronized target micro-Doppler waterfall plots and radar-measured track information (including azimuth and elevation angles, radial velocity, and normalized signal-to-noise ratio), providing a data foundation for exploring the intrinsic link between target detailed features and motion states. Building on this, a multimodal adaptive feature fusion network is developed to extract and combine Doppler and kinematic features from different targets, demonstrating the dataset’s effectiveness in distinguishing various low-altitude targets.
5
2023, 12(2): 456-469.
Marine target detection and recognition depend on the characteristics of marine targets and sea clutter. Therefore, understanding the essential features of marine targets based on the measured data is crucial for advancing target detection and recognition technology. To address the issue of insufficient data on the scattering characteristics of marine targets, the Sea-Detecting Radar Data-Sharing Program (SDRDSP) was upgraded to obtain data on marine targets and their environment under different polarizations and sea states. This upgrade expanded the physical dimension of radar target observation and improved radar and auxiliary data acquisition capabilities. Furthermore, a dual-polarized multistate scattering characteristic dataset of marine targets was constructed, and the statistical distribution characteristics, time and space correlation, and Doppler spectrum were analyzed, supporting the data usage. In the future, the types and quantities of maritime targets will continue to accumulate, providing data support for improving marine target detection and recognition performance and intelligence.
Marine target detection and recognition depend on the characteristics of marine targets and sea clutter. Therefore, understanding the essential features of marine targets based on the measured data is crucial for advancing target detection and recognition technology. To address the issue of insufficient data on the scattering characteristics of marine targets, the Sea-Detecting Radar Data-Sharing Program (SDRDSP) was upgraded to obtain data on marine targets and their environment under different polarizations and sea states. This upgrade expanded the physical dimension of radar target observation and improved radar and auxiliary data acquisition capabilities. Furthermore, a dual-polarized multistate scattering characteristic dataset of marine targets was constructed, and the statistical distribution characteristics, time and space correlation, and Doppler spectrum were analyzed, supporting the data usage. In the future, the types and quantities of maritime targets will continue to accumulate, providing data support for improving marine target detection and recognition performance and intelligence.
6
,
A large-scale Vision-Language Model (VLM) pre-trained on massive image-text datasets performs well when processing natural images. However, there are two major challenges in applying it to Synthetic Aperture Radar (SAR) images: (1) The high cost of high-quality text annotation limits the construction of SAR image-text paired datasets, and (2) The considerable differences in image features between SAR images and optical natural images increase the difficulty of cross-domain knowledge transfer. To address these problems, this study developed a knowledge transfer method for VLM tailored to SAR images. First, this study leveraged paired SAR and optical remote sensing images and employed a generative VLM to automatically produce textual descriptions of the optical images, thereby indirectly constructing a low-cost SAR-text paired dataset. Second, a two-stage transfer strategy was designed to address the large domain discrepancy between natural and SAR images, reducing the difficulty of each transfer stage. Finally, experimental validation was conducted through the zero-shot scene classification, image retrieval, and object recognition of SAR images. The results demonstrated that the proposed method enables effective knowledge transfer from a large-scale VLM to the SAR image domain.
A large-scale Vision-Language Model (VLM) pre-trained on massive image-text datasets performs well when processing natural images. However, there are two major challenges in applying it to Synthetic Aperture Radar (SAR) images: (1) The high cost of high-quality text annotation limits the construction of SAR image-text paired datasets, and (2) The considerable differences in image features between SAR images and optical natural images increase the difficulty of cross-domain knowledge transfer. To address these problems, this study developed a knowledge transfer method for VLM tailored to SAR images. First, this study leveraged paired SAR and optical remote sensing images and employed a generative VLM to automatically produce textual descriptions of the optical images, thereby indirectly constructing a low-cost SAR-text paired dataset. Second, a two-stage transfer strategy was designed to address the large domain discrepancy between natural and SAR images, reducing the difficulty of each transfer stage. Finally, experimental validation was conducted through the zero-shot scene classification, image retrieval, and object recognition of SAR images. The results demonstrated that the proposed method enables effective knowledge transfer from a large-scale VLM to the SAR image domain.
7
2026, 15(2): 503-522.
Synthetic Aperture Radar (SAR) offers all-weather, all-day maritime surveillance capabilities. Direct ship detection in the Range Compressed Domain (RCD) eliminates computationally intensive imaging steps—such as range cell migration correction and azimuth compression—thereby considerably improving processing efficiency for near-real-time or real-time applications. However, current detection methods face inherent limitations; traditional constant false alarm rate detectors rely on fixed statistical models and often underperform in complex sea clutter environments. In addition, deep learning approaches heavily rely on annotated data and do not fully leverage phase information; moreover, they exhibit weak interpretability. To address these issues, this paper proposes a self-supervised reinforcement learning framework for ship target detection in the SAR RCD. This framework effectively integrates the physical principles of radar signals with deep reinforcement learning, achieving enhanced detection performance while improving model interpretability and generalization. The framework has the following characteristics: (1) It introduces a reward signal-generation mechanism constrained by statistical scattering models, achieving self-supervised learning without the need for manual annotation; (2) It designs a dual-modal feature-fusion module that can jointly represent amplitude and phase information, effectively retaining the Doppler characteristics of ships; and (3) It adopts a lightweight agent module that integrates a lightweight Q-network, an adaptive feature enhancement module, and a discriminator network; this module reduces computational complexity, meets real-time processing requirements, and enhances the robustness of the model through adversarial training. Experimental results demonstrate that the proposed method achieves an average inference time of only 31.75 s on a large-scale SAR RCD dataset of 20 k×20 k pixels, with a computational load of only 23.81% compared with a two-dimensional convolutional neural network. On a complex-valued RCD dataset, the method attains F1 and recall scores of 50.72% and 54.28%, respectively, outperforming mainstream self-supervised methods by 8.76% and 10.45%, respectively. This study pioneers the application of reinforcement learning to ship detection using SAR RCD, offering a novel approach to robust maritime surveillance by integrating signal modeling and data-driven learning.
Synthetic Aperture Radar (SAR) offers all-weather, all-day maritime surveillance capabilities. Direct ship detection in the Range Compressed Domain (RCD) eliminates computationally intensive imaging steps—such as range cell migration correction and azimuth compression—thereby considerably improving processing efficiency for near-real-time or real-time applications. However, current detection methods face inherent limitations; traditional constant false alarm rate detectors rely on fixed statistical models and often underperform in complex sea clutter environments. In addition, deep learning approaches heavily rely on annotated data and do not fully leverage phase information; moreover, they exhibit weak interpretability. To address these issues, this paper proposes a self-supervised reinforcement learning framework for ship target detection in the SAR RCD. This framework effectively integrates the physical principles of radar signals with deep reinforcement learning, achieving enhanced detection performance while improving model interpretability and generalization. The framework has the following characteristics: (1) It introduces a reward signal-generation mechanism constrained by statistical scattering models, achieving self-supervised learning without the need for manual annotation; (2) It designs a dual-modal feature-fusion module that can jointly represent amplitude and phase information, effectively retaining the Doppler characteristics of ships; and (3) It adopts a lightweight agent module that integrates a lightweight Q-network, an adaptive feature enhancement module, and a discriminator network; this module reduces computational complexity, meets real-time processing requirements, and enhances the robustness of the model through adversarial training. Experimental results demonstrate that the proposed method achieves an average inference time of only 31.75 s on a large-scale SAR RCD dataset of 20 k×20 k pixels, with a computational load of only 23.81% compared with a two-dimensional convolutional neural network. On a complex-valued RCD dataset, the method attains F1 and recall scores of 50.72% and 54.28%, respectively, outperforming mainstream self-supervised methods by 8.76% and 10.45%, respectively. This study pioneers the application of reinforcement learning to ship detection using SAR RCD, offering a novel approach to robust maritime surveillance by integrating signal modeling and data-driven learning.
8
2025, 14(5): 1276-1293.
This study addresses the issue of fine-grained feature extraction and classification for Low-Slow-Small (LSS) targets, such as birds and drones, by proposing a multi-band multi-angle feature fusion classification method. First, data from five types of rotorcraft drones and bird models were collected at multiple angles using K-band and L-band frequency-modulated continuous-wave radars, forming a dataset for LSS target detection. Second, to capture the periodic vibration characteristics of the L-band target signals, empirical mode decomposition was applied to extract high-frequency features and reduce noise interference. For the K-band echo signals, short-time Fourier transform was applied to obtain high-resolution micro-Doppler features from various angles. Based on these features, a Multi-band Multi-angle Feature Fusion Network (MMFFNet) was designed, incorporating an improved convolutional long short-term memory network for temporal feature extraction, along with an attention fusion module and a multiscale feature fusion module. The proposed architecture improves target classification accuracy by integrating features from both bands and angles. Validation using a real-world dataset showed that compared with methods relying on single radar features, the proposed approach improved the classification accuracy for seven types of LSS targets by 3.1% under a high Signal-to-Noise Ratio (SNR) of 5 dB and by 12.3% under a low SNR of −3 dB.
This study addresses the issue of fine-grained feature extraction and classification for Low-Slow-Small (LSS) targets, such as birds and drones, by proposing a multi-band multi-angle feature fusion classification method. First, data from five types of rotorcraft drones and bird models were collected at multiple angles using K-band and L-band frequency-modulated continuous-wave radars, forming a dataset for LSS target detection. Second, to capture the periodic vibration characteristics of the L-band target signals, empirical mode decomposition was applied to extract high-frequency features and reduce noise interference. For the K-band echo signals, short-time Fourier transform was applied to obtain high-resolution micro-Doppler features from various angles. Based on these features, a Multi-band Multi-angle Feature Fusion Network (MMFFNet) was designed, incorporating an improved convolutional long short-term memory network for temporal feature extraction, along with an attention fusion module and a multiscale feature fusion module. The proposed architecture improves target classification accuracy by integrating features from both bands and angles. Validation using a real-world dataset showed that compared with methods relying on single radar features, the proposed approach improved the classification accuracy for seven types of LSS targets by 3.1% under a high Signal-to-Noise Ratio (SNR) of 5 dB and by 12.3% under a low SNR of −3 dB.
9
2026, 15(2): 441-462.
Synthetic Aperture Radar (SAR), as an active microwave remote sensing system, offers all-weather, all-day observation capabilities and has considerable application value in disaster monitoring, urban management, and military reconnaissance. Although deep learning techniques have achieved remarkable progress in interpreting SAR images, existing methods for target recognition and detection primarily focus on local feature extraction and single-target discrimination. They struggle to comprehensively characterize the global semantic structure and multitarget relationships in complex scenes, and the interpretation process remains highly dependent on human expertise with limited automation. SAR image captioning aims to translate visual information into natural language, serving as a key technology to bridge the gap between “perceiving targets” and “cognizing scenes,” which is of great importance for enhancing the automation and intelligence of SAR image interpretation. However, the inherent speckle noise, the scarcity of textural details, and the substantial semantic gap in SAR images further exacerbate the difficulty of cross-modal understanding. To address these challenges, this paper proposes a spatial-frequency aware model for SAR image captioning. First, a spatial-frequency aware module is constructed. It employs a Discrete Cosine Transform (DCT) mask attention mechanism to reweight spectral components for noise suppression and structure enhancement, combined with a Gabor multiscale texture enhancement submodule to improve sensitivity to directional and edge details. Second, a cross-modal semantic enhancement loss function is designed to bridge the semantic gap between visual features and natural language through bidirectional image-text alignment and mutual information maximization. Furthermore, a large-scale fine-grained SAR image captioning dataset, FSAR-Cap, containing 72400 high-quality image-text pairs, is constructed. The experimental results demonstrate that the proposed method achieves CIDEr scores of 151.00 and 95.14 on the SARLANG and FSAR-Cap datasets, respectively. Qualitatively, the model effectively suppresses hallucinations and accurately captures fine-grained spatial-textural details, considerably outperforming mainstream methods.
Synthetic Aperture Radar (SAR), as an active microwave remote sensing system, offers all-weather, all-day observation capabilities and has considerable application value in disaster monitoring, urban management, and military reconnaissance. Although deep learning techniques have achieved remarkable progress in interpreting SAR images, existing methods for target recognition and detection primarily focus on local feature extraction and single-target discrimination. They struggle to comprehensively characterize the global semantic structure and multitarget relationships in complex scenes, and the interpretation process remains highly dependent on human expertise with limited automation. SAR image captioning aims to translate visual information into natural language, serving as a key technology to bridge the gap between “perceiving targets” and “cognizing scenes,” which is of great importance for enhancing the automation and intelligence of SAR image interpretation. However, the inherent speckle noise, the scarcity of textural details, and the substantial semantic gap in SAR images further exacerbate the difficulty of cross-modal understanding. To address these challenges, this paper proposes a spatial-frequency aware model for SAR image captioning. First, a spatial-frequency aware module is constructed. It employs a Discrete Cosine Transform (DCT) mask attention mechanism to reweight spectral components for noise suppression and structure enhancement, combined with a Gabor multiscale texture enhancement submodule to improve sensitivity to directional and edge details. Second, a cross-modal semantic enhancement loss function is designed to bridge the semantic gap between visual features and natural language through bidirectional image-text alignment and mutual information maximization. Furthermore, a large-scale fine-grained SAR image captioning dataset, FSAR-Cap, containing 72400 high-quality image-text pairs, is constructed. The experimental results demonstrate that the proposed method achieves CIDEr scores of 151.00 and 95.14 on the SARLANG and FSAR-Cap datasets, respectively. Qualitatively, the model effectively suppresses hallucinations and accurately captures fine-grained spatial-textural details, considerably outperforming mainstream methods.
10
To address issues such as insufficient feature extraction, limited spatiotemporal correlation modeling, and poor classification performance in radar classification of Low, Slow, and Small (LSS) targets, this paper investigates on graph network-based feature extraction and classification methods. First, focusing on digital array ubiquitous radar, a radar detection dataset for LSS targets, named LSS-DAUR-1.0, is constructed; it contains Doppler and track data for six types of targets: Passenger ships, speedboats, helicopters, rotor drones, birds, and fixed-wing drones. Second, based on this dataset, the multidomain and multidimensional characteristics of the targets are analyzed, and the complementarity between Doppler and physical motion features is verified through correlation and cosine similarity analyses. On this basis, a Graph Convolutional Network with Dynamic Graph Construction (DG-GCN) classification method fusing dual features is proposed. An adaptive window adjustment, a hybrid attenuation function, and a dynamic threshold mechanism are designed to construct an adaptive dynamic graph based on spatiotemporal correlation. Combined with graph convolution-based feature learning and classification modules, this approach achieves refined classification of low, slow, and small targets. Validation on the LSS-DAUR-1.0 dataset shows that the DG-GCN achieves 99.66% classification accuracy, which is 6.78% and 17.97% higher than that of ResNet and Transformer models, respectively. The total processing time is only 4.98 ms, which is more than 80% lower than that of the aforementioned comparison models. Hence, the DG-GCN achieves both high accuracy and efficiency. In addition, noise environment tests show good robustness. Ablation experiments verify that the dynamic edge weight mechanism compensates for the lack of spatial feature correlation in purely temporal connections and improves the model’s generalizability.
To address issues such as insufficient feature extraction, limited spatiotemporal correlation modeling, and poor classification performance in radar classification of Low, Slow, and Small (LSS) targets, this paper investigates on graph network-based feature extraction and classification methods. First, focusing on digital array ubiquitous radar, a radar detection dataset for LSS targets, named LSS-DAUR-1.0, is constructed; it contains Doppler and track data for six types of targets: Passenger ships, speedboats, helicopters, rotor drones, birds, and fixed-wing drones. Second, based on this dataset, the multidomain and multidimensional characteristics of the targets are analyzed, and the complementarity between Doppler and physical motion features is verified through correlation and cosine similarity analyses. On this basis, a Graph Convolutional Network with Dynamic Graph Construction (DG-GCN) classification method fusing dual features is proposed. An adaptive window adjustment, a hybrid attenuation function, and a dynamic threshold mechanism are designed to construct an adaptive dynamic graph based on spatiotemporal correlation. Combined with graph convolution-based feature learning and classification modules, this approach achieves refined classification of low, slow, and small targets. Validation on the LSS-DAUR-1.0 dataset shows that the DG-GCN achieves 99.66% classification accuracy, which is 6.78% and 17.97% higher than that of ResNet and Transformer models, respectively. The total processing time is only 4.98 ms, which is more than 80% lower than that of the aforementioned comparison models. Hence, the DG-GCN achieves both high accuracy and efficiency. In addition, noise environment tests show good robustness. Ablation experiments verify that the dynamic edge weight mechanism compensates for the lack of spatial feature correlation in purely temporal connections and improves the model’s generalizability.
11
2026, 15(2): 543-562.
The effective utilization of Synthetic Aperture Radar (SAR) adversarial examples enables specific targets to achieve remote sensing stealth against intelligent detection systems, thereby evading detection and recognition by adversaries. Digital domain SAR adversarial methods, which operate exclusively in the image domain, produce adversarial images that are not physically realizable and therefore cannot generated by real SAR imaging systems. Existing physical domain approaches typically involve deploying corner reflectors or electromagnetic metasurfaces around targets and simulating adversarial examples using via computational electromagnetics. However, the limited accuracy of scattering estimation often constrains the practical protective efficacy of these methods. To overcome these limitations, this paper proposes an active jammer-based adversarial attack method that integrates SAR active jamming technology with adversarial attack methods to generate adversarial examples by perturbing the target’s echo signals in the signal domain. First, a multiple-phase sectionalized modulation jamming method based on cosine amplitude weighting is selected, enabling parameterized control of the adversarial jamming signal through the design of perturbation components. Next, the adversarial jamming signal generated by the active jammer is fused with the target’s echo signal according to the principles and actual processes of SAR imaging and is then subjected to imaging processing to produce physically realizable SAR adversarial examples. Finally, the differential evolution algorithm is employed to dynamically adjust parameters, such as the energy distribution and jamming range of the adversarial jamming signal, thereby optimizing the SAR adversarial examples to achieve optimal attack success rates even with minimal interference intensity. Experimental results on the MSTAR dataset, a widely used benchmark in the field of SAR Automatic Target Recognition (ATR), show that the proposed method achieves an average fooling rate of 90.88% and demonstrates superior transferability across five different SAR ATR models, with the highest transfer fooling rate reaching 75.57%. Overall, the proposed method generates more physically realizable adversarial examples compared with existing digital domain methods, effectively protecting specific targets in remote sensing detection and providing guidance for the practical application of active jamming signals in real-world scenarios.
The effective utilization of Synthetic Aperture Radar (SAR) adversarial examples enables specific targets to achieve remote sensing stealth against intelligent detection systems, thereby evading detection and recognition by adversaries. Digital domain SAR adversarial methods, which operate exclusively in the image domain, produce adversarial images that are not physically realizable and therefore cannot generated by real SAR imaging systems. Existing physical domain approaches typically involve deploying corner reflectors or electromagnetic metasurfaces around targets and simulating adversarial examples using via computational electromagnetics. However, the limited accuracy of scattering estimation often constrains the practical protective efficacy of these methods. To overcome these limitations, this paper proposes an active jammer-based adversarial attack method that integrates SAR active jamming technology with adversarial attack methods to generate adversarial examples by perturbing the target’s echo signals in the signal domain. First, a multiple-phase sectionalized modulation jamming method based on cosine amplitude weighting is selected, enabling parameterized control of the adversarial jamming signal through the design of perturbation components. Next, the adversarial jamming signal generated by the active jammer is fused with the target’s echo signal according to the principles and actual processes of SAR imaging and is then subjected to imaging processing to produce physically realizable SAR adversarial examples. Finally, the differential evolution algorithm is employed to dynamically adjust parameters, such as the energy distribution and jamming range of the adversarial jamming signal, thereby optimizing the SAR adversarial examples to achieve optimal attack success rates even with minimal interference intensity. Experimental results on the MSTAR dataset, a widely used benchmark in the field of SAR Automatic Target Recognition (ATR), show that the proposed method achieves an average fooling rate of 90.88% and demonstrates superior transferability across five different SAR ATR models, with the highest transfer fooling rate reaching 75.57%. Overall, the proposed method generates more physically realizable adversarial examples compared with existing digital domain methods, effectively protecting specific targets in remote sensing detection and providing guidance for the practical application of active jamming signals in real-world scenarios.
12
2026, 15(2): 605-619.
Low-altitude targets, represented by rotor unmanned aerial vehicles, can typically adopt a slow-cruise mode. As a result, their echoes fall within the Doppler Blind Zone (DBZ) and evade radar detection and tracking. The cluttered low-altitude environment adds to further complexity. To address this issue, this study proposes a method grounded in the framework of random finite set and designed for tracking slow-moving targets with a low-altitude surveillance radar. Inspired by the Bayesian occupancy filter, the proposed method initially models the radar Field of View (FoV) as a grid map. It is uniformly partitioned along the angle-range axis, ensuring that each cell captures a specific segment of the FoV. Then, adaptive filtering parameter modules are meticulously designed by leveraging the distinct dynamic characteristics of slow-moving targets and ground clutter. Subsequently, a probability hypothesis density filter is deployed to conduct unified filtering on the grid map situated within the DBZ. The final step involves the use of clustering methods to extract information about the target of interest. Simulation results validate the effectiveness, robustness, and superior performance of the proposed method across typical surveillance scenarios involving multiple slow-moving targets, noise, and clutter.
Low-altitude targets, represented by rotor unmanned aerial vehicles, can typically adopt a slow-cruise mode. As a result, their echoes fall within the Doppler Blind Zone (DBZ) and evade radar detection and tracking. The cluttered low-altitude environment adds to further complexity. To address this issue, this study proposes a method grounded in the framework of random finite set and designed for tracking slow-moving targets with a low-altitude surveillance radar. Inspired by the Bayesian occupancy filter, the proposed method initially models the radar Field of View (FoV) as a grid map. It is uniformly partitioned along the angle-range axis, ensuring that each cell captures a specific segment of the FoV. Then, adaptive filtering parameter modules are meticulously designed by leveraging the distinct dynamic characteristics of slow-moving targets and ground clutter. Subsequently, a probability hypothesis density filter is deployed to conduct unified filtering on the grid map situated within the DBZ. The final step involves the use of clustering methods to extract information about the target of interest. Simulation results validate the effectiveness, robustness, and superior performance of the proposed method across typical surveillance scenarios involving multiple slow-moving targets, noise, and clutter.
13
Driven by complex electromagnetic environments and multi-target collaborative detection needs, enhancing the overall effectiveness of radar networks through autonomous coordination technology has become a key research area in radar collaborative surveillance. Extensive research has been conducted worldwide, yielding substantial advances in theoretical development, technical validation, and equipment application. This paper systematically discusses the foundational concepts and main features of autonomous coordination in radar networks, examining the primary technical challenges faced during implementation and performance optimization. It also reviews recent notable research findings and technological strategies, focusing on collaborative architecture design, sensing, intelligent decision-making and control, and autonomous evolution. Finally, this paper offers an outlook on future trends in the field and provides references for related theoretical research and practical applications.
Driven by complex electromagnetic environments and multi-target collaborative detection needs, enhancing the overall effectiveness of radar networks through autonomous coordination technology has become a key research area in radar collaborative surveillance. Extensive research has been conducted worldwide, yielding substantial advances in theoretical development, technical validation, and equipment application. This paper systematically discusses the foundational concepts and main features of autonomous coordination in radar networks, examining the primary technical challenges faced during implementation and performance optimization. It also reviews recent notable research findings and technological strategies, focusing on collaborative architecture design, sensing, intelligent decision-making and control, and autonomous evolution. Finally, this paper offers an outlook on future trends in the field and provides references for related theoretical research and practical applications.
14
2025, 14(3): 754-780.
Maritime target detection and identification technology are developed using large-scale, high-quality multi-sensor measurement data. Therefore, the Sea Detection Radar Data Sharing Program (SDRDSP) was upgraded to the Maritime Target Data Sharing Program (MTDSP), integrating multiple observation modalities, such as HH-polarized radar, VV-polarized radar, electro-optical devices, and Automatic Identification System (AIS) equipment to conduct multisource observation experiments on maritime vessel targets. The program collects various data types, including radar intermediate frequency/video echo slice data, visible and infrared imagery, AIS static and dynamic messages, and meteorological and hydrological data, covering representative sea conditions and multiple vessel types. A comprehensive multisource observation dataset was constructed, enabling the matching and annotation of multimodal data for the same target. Moreover, an automated data management system was implemented to support data storage, conditional retrieval, and batch export, providing a solid foundation for the automated acquisition, long-term accumulation, and efficient use of maritime target characteristic data. Based on this system and measured data, the time/frequency domain features of the same and different vessel targets under different sea states, attitudes, polarization conditions are compared and analyzed, and the statistical conclusion of the change in target features is obtained.
Maritime target detection and identification technology are developed using large-scale, high-quality multi-sensor measurement data. Therefore, the Sea Detection Radar Data Sharing Program (SDRDSP) was upgraded to the Maritime Target Data Sharing Program (MTDSP), integrating multiple observation modalities, such as HH-polarized radar, VV-polarized radar, electro-optical devices, and Automatic Identification System (AIS) equipment to conduct multisource observation experiments on maritime vessel targets. The program collects various data types, including radar intermediate frequency/video echo slice data, visible and infrared imagery, AIS static and dynamic messages, and meteorological and hydrological data, covering representative sea conditions and multiple vessel types. A comprehensive multisource observation dataset was constructed, enabling the matching and annotation of multimodal data for the same target. Moreover, an automated data management system was implemented to support data storage, conditional retrieval, and batch export, providing a solid foundation for the automated acquisition, long-term accumulation, and efficient use of maritime target characteristic data. Based on this system and measured data, the time/frequency domain features of the same and different vessel targets under different sea states, attitudes, polarization conditions are compared and analyzed, and the statistical conclusion of the change in target features is obtained.
15
2026, 15(2): 463-478.
Variations in imaging geometry are the main cause of relative feature distortion in Synthetic Aperture Radar (SAR) images, greatly increasing the difficulty of image matching. Using simulated SAR images as references can remove the feature distortions caused by geometric differences. However, significant differences in scattering characteristics and noise patterns between measured and simulated images still exist. Additionally, since most existing matching algorithms mainly rely on symmetric keypoint detection and descriptor matching, the number and precision of matched points are not optimal. To solve these problems, this paper introduces an asymmetric Local Fitting Consistency (LFC) similarity metric based on the local statistical features of both measured and simulated SAR images. Using this metric, a coarse-to-fine matching framework for airborne and simulated SAR images is designed. Furthermore, terrain features are added to improve keypoint detection diversity, leading to more robust matching between airborne and simulated SAR images. Experimental results show that the proposed LFC-based matching method offers better robustness and accuracy compared to other approaches, significantly surpassing current state-of-the-art algorithms in terms of matching precision and other key metrics.
Variations in imaging geometry are the main cause of relative feature distortion in Synthetic Aperture Radar (SAR) images, greatly increasing the difficulty of image matching. Using simulated SAR images as references can remove the feature distortions caused by geometric differences. However, significant differences in scattering characteristics and noise patterns between measured and simulated images still exist. Additionally, since most existing matching algorithms mainly rely on symmetric keypoint detection and descriptor matching, the number and precision of matched points are not optimal. To solve these problems, this paper introduces an asymmetric Local Fitting Consistency (LFC) similarity metric based on the local statistical features of both measured and simulated SAR images. Using this metric, a coarse-to-fine matching framework for airborne and simulated SAR images is designed. Furthermore, terrain features are added to improve keypoint detection diversity, leading to more robust matching between airborne and simulated SAR images. Experimental results show that the proposed LFC-based matching method offers better robustness and accuracy compared to other approaches, significantly surpassing current state-of-the-art algorithms in terms of matching precision and other key metrics.
16
2026, 15(2): 759-778.
Ocean currents play a critical role in global climate regulation. Synthetic Aperture Radar (SAR) provides high-resolution observational support for ocean current detection by measuring Doppler shifts; however, SAR Doppler shifts contain multiple contributing components. To accurately retrieve ocean currents from these data, nongeophysical contributions must be precisely corrected, and wind- and wave-induced Doppler shifts must be accurately estimated. This paper proposes a machine learning-based method for modeling such shifts and retrieving ocean currents from Sentinel-1 SAR data. First, nongeophysical contributions in the SAR Doppler shift are precisely corrected to remove the effects unrelated to ocean motion. Second, BackproPagation Neural Network (BPNN) and eXtreme Gradient Boosting (XGBoost) models, optimized using the particle swarm optimization algorithm, are developed to describe the nonlinear relationship between the wind-wave Doppler shift and sea-surface wind-wave parameters derived from SAR data. Finally, the corrected Doppler shift is utilized to retrieve ocean surface current velocities. This paper comparatively evaluates the estimation accuracies of the wind-wave Doppler shifts obtained using the BPNN and XGBoost models, as well as the respective influence of each model's performance on the effectiveness of ocean current retrieval. Results indicate that the XGBoost model achieves superior estimation accuracy compared with the BPNN model. The Root Mean Square Error (RMSE) of the Doppler shift estimated by the XGBoost model is approximately 4.043 Hz, which is 2.898 Hz lower than that of the BPNN model. Compared with those of the HYCOM current data, the RMSE of the currents retrieved by the XGBoost model is about 0.202 m/s; this value is reduced by 0.122 m/s compared with that of the BPNN model. Validations against the current velocities detected by HF radar show that the RMSE of currents retrieved by the XGBoost model is 0.21 m/s, representing a 16% reduction compared with that of the BPNN model. These findings indicate that the proposed technical approach for ocean current retrieval using spaceborne SAR is highly accurate.
Ocean currents play a critical role in global climate regulation. Synthetic Aperture Radar (SAR) provides high-resolution observational support for ocean current detection by measuring Doppler shifts; however, SAR Doppler shifts contain multiple contributing components. To accurately retrieve ocean currents from these data, nongeophysical contributions must be precisely corrected, and wind- and wave-induced Doppler shifts must be accurately estimated. This paper proposes a machine learning-based method for modeling such shifts and retrieving ocean currents from Sentinel-1 SAR data. First, nongeophysical contributions in the SAR Doppler shift are precisely corrected to remove the effects unrelated to ocean motion. Second, BackproPagation Neural Network (BPNN) and eXtreme Gradient Boosting (XGBoost) models, optimized using the particle swarm optimization algorithm, are developed to describe the nonlinear relationship between the wind-wave Doppler shift and sea-surface wind-wave parameters derived from SAR data. Finally, the corrected Doppler shift is utilized to retrieve ocean surface current velocities. This paper comparatively evaluates the estimation accuracies of the wind-wave Doppler shifts obtained using the BPNN and XGBoost models, as well as the respective influence of each model's performance on the effectiveness of ocean current retrieval. Results indicate that the XGBoost model achieves superior estimation accuracy compared with the BPNN model. The Root Mean Square Error (RMSE) of the Doppler shift estimated by the XGBoost model is approximately 4.043 Hz, which is 2.898 Hz lower than that of the BPNN model. Compared with those of the HYCOM current data, the RMSE of the currents retrieved by the XGBoost model is about 0.202 m/s; this value is reduced by 0.122 m/s compared with that of the BPNN model. Validations against the current velocities detected by HF radar show that the RMSE of currents retrieved by the XGBoost model is 0.21 m/s, representing a 16% reduction compared with that of the BPNN model. These findings indicate that the proposed technical approach for ocean current retrieval using spaceborne SAR is highly accurate.
17
2026, 15(2): 687-709.
Geosynchronous Orbit (GEO) Synthetic Aperture Radar (SAR) detection ensures persistent, wide-area surveillance. However, this ship-detection method faces significant technical challenges, such as imaging defocusing, low Signal-to-Clutter Ratio (SCR), and large position offsets, due to the long detection distance, long synthetic aperture time, clutter accumulation within a large field of view, and nonplanar observation geometry. To address these challenges, this paper proposes a novel integrated detection-tracking-localization framework for moving-ship targets in GEO SAR. First, a GEO SAR observation signal model is established for moving ships, after which their echo characteristics within the ultra-long synthetic aperture time are analyzed in depth. On this basis, the model realizes target-image detection and long-term tracking localization via optimal subaperture processing. Using an improved back-projection imaging algorithm tailored for moving ships, effective energy accumulation and focusing of noncooperative ships under low SCR are achieved within the aperture. In addition, the relationship between the offset position of moving targets and the Range-Doppler (RD) parameters under GEO SAR nonplanar geometric observation is obtained. Second, under the assumption of short-term uniform ship motion, a bidirectional smoothing filter is applied to track the multisubaperture detection results. The velocity estimation of moving ships is obtained from the long-term tracking results, and the relocation of moving ships is realized using the RD relationship between the offset position and the actual position. Finally, the proposed framework is validated using simulation data and on-orbit GEO SAR satellite test data.
Geosynchronous Orbit (GEO) Synthetic Aperture Radar (SAR) detection ensures persistent, wide-area surveillance. However, this ship-detection method faces significant technical challenges, such as imaging defocusing, low Signal-to-Clutter Ratio (SCR), and large position offsets, due to the long detection distance, long synthetic aperture time, clutter accumulation within a large field of view, and nonplanar observation geometry. To address these challenges, this paper proposes a novel integrated detection-tracking-localization framework for moving-ship targets in GEO SAR. First, a GEO SAR observation signal model is established for moving ships, after which their echo characteristics within the ultra-long synthetic aperture time are analyzed in depth. On this basis, the model realizes target-image detection and long-term tracking localization via optimal subaperture processing. Using an improved back-projection imaging algorithm tailored for moving ships, effective energy accumulation and focusing of noncooperative ships under low SCR are achieved within the aperture. In addition, the relationship between the offset position of moving targets and the Range-Doppler (RD) parameters under GEO SAR nonplanar geometric observation is obtained. Second, under the assumption of short-term uniform ship motion, a bidirectional smoothing filter is applied to track the multisubaperture detection results. The velocity estimation of moving ships is obtained from the long-term tracking results, and the relocation of moving ships is realized using the RD relationship between the offset position and the actual position. Finally, the proposed framework is validated using simulation data and on-orbit GEO SAR satellite test data.
18
Research Advances and Applications of Microwave Photonic Broadband Vortex Electromagnetic Wave Radar
,
Vortex Electromagnetic (EM) wave radars utilize EM waves carrying orbital angular momentum to enrich target scattering information, thereby providing intrinsic in-beam azimuth resolution. Hence, this technology holds significant potential for advanced target detection and imaging. However, as sensing scenarios become more complex, conventional electronic vortex EM wave radars are increasingly limited by device bandwidth. Specifically, they encounter substantial challenges in broadband signal generation and control, making it difficult to achieve high range and azimuth resolutions simultaneously. Microwave photonics technology, with its inherent advantages of wide bandwidth, low transmission loss, and robustness against electromagnetic interference, is an effectivesolution to overcome these limitations. This paper reviews recent progress in microwave photonic broadband vortex EM wave radars, addressing the requirements for forward-looking imaging. The fundamental system architectures and imaging mechanisms are elucidated, followed by a critical analysis of the frequency-dependent characteristics of broadband vortex waves and their implications for imaging performance. Key microwave photonic enabling technologies, including broadband phase shifting, optical beamforming, and broadband signal generation, are summarized, and their advantages over traditional electronic schemes in terms of performance are highlighted. Based on these insights, typical system implementation schemes are described, and their high-resolution forward-looking imaging capabilities are demonstrated through proof-of-concept experiments. Finally, future development trends and research directions are discussed.
Vortex Electromagnetic (EM) wave radars utilize EM waves carrying orbital angular momentum to enrich target scattering information, thereby providing intrinsic in-beam azimuth resolution. Hence, this technology holds significant potential for advanced target detection and imaging. However, as sensing scenarios become more complex, conventional electronic vortex EM wave radars are increasingly limited by device bandwidth. Specifically, they encounter substantial challenges in broadband signal generation and control, making it difficult to achieve high range and azimuth resolutions simultaneously. Microwave photonics technology, with its inherent advantages of wide bandwidth, low transmission loss, and robustness against electromagnetic interference, is an effectivesolution to overcome these limitations. This paper reviews recent progress in microwave photonic broadband vortex EM wave radars, addressing the requirements for forward-looking imaging. The fundamental system architectures and imaging mechanisms are elucidated, followed by a critical analysis of the frequency-dependent characteristics of broadband vortex waves and their implications for imaging performance. Key microwave photonic enabling technologies, including broadband phase shifting, optical beamforming, and broadband signal generation, are summarized, and their advantages over traditional electronic schemes in terms of performance are highlighted. Based on these insights, typical system implementation schemes are described, and their high-resolution forward-looking imaging capabilities are demonstrated through proof-of-concept experiments. Finally, future development trends and research directions are discussed.
19
2026, 15(2): 746-758.
Sea surface elevation is crucial for characterizing individual waves, wave groups, and freak waves, offering an accurate representation of inhomogeneous sea states. This study presents a quasi-linear inversion strategy for retrieving sea surface elevation from GF-3 Synthetic Aperture Radar (SAR) images. The algorithm enables rapid inversion within 10 s per scene without the need for external data and effectively resolves range-traveling waves. Case studies conducted under three distinct sea states demonstrate its ability to extract maximum wave heights and identify wave groups. Additionally, inversion results from 2405 GF-3 wave mode SAR images following quality control are compared with ERA5 reanalysis spectra and altimeter data. The comparisons reveal that the retrieved Significant Wave Height (SWH) has a root mean square error of 0.48 m compared with ERA5 data. In low-to-moderate sea states, with significant wave heights below 3 m, the retrieved SWH shows strong consistency with ERA5 spectra and altimeter measurements. This algorithm serves as an effective tool for rapid monitoring and analysis of sea states using GF-3 SAR.
Sea surface elevation is crucial for characterizing individual waves, wave groups, and freak waves, offering an accurate representation of inhomogeneous sea states. This study presents a quasi-linear inversion strategy for retrieving sea surface elevation from GF-3 Synthetic Aperture Radar (SAR) images. The algorithm enables rapid inversion within 10 s per scene without the need for external data and effectively resolves range-traveling waves. Case studies conducted under three distinct sea states demonstrate its ability to extract maximum wave heights and identify wave groups. Additionally, inversion results from 2405 GF-3 wave mode SAR images following quality control are compared with ERA5 reanalysis spectra and altimeter data. The comparisons reveal that the retrieved Significant Wave Height (SWH) has a root mean square error of 0.48 m compared with ERA5 data. In low-to-moderate sea states, with significant wave heights below 3 m, the retrieved SWH shows strong consistency with ERA5 spectra and altimeter measurements. This algorithm serves as an effective tool for rapid monitoring and analysis of sea states using GF-3 SAR.
20
2026, 15(2): 563-582.
Adversarial sample generation is a key research direction for uncovering the vulnerabilities of deep neural networks and improving the robustness of Synthetic Aperture Radar Automatic Target Recognition (SAR ATR) systems. This study proposes an optimal adversarial sample generation method for SAR ATR that jointly optimizes misleading effectiveness and fidelity, aiming to resolve the core contradiction between adversarial effectiveness and visual concealment. The generation process is modeled as a joint optimization problem with the goals of balancing “misleading” and “fidelity”. First, an integrated composite transform attack strategy is designed to enhance attack effectiveness, and a joint measurement model is developed that combines the classification accuracy of the target model with the Learned Perceptual Image Patch Similarity (LPIPS) to quantify the two optimization goals. Next, an improved uniformity-guided multiobjective RIME algorithm is proposed. By integrating the Tent chaotic map, hybrid dynamic weighting, and golden sine guidance, the model is efficiently solved, yielding a set of Pareto-optimal solutions that represent various tradeoff degrees. Finally, the YOLOv10 object detection network is employed to identify perturbations in the samples within the solution set, thereby locating the critical points where disturbances occur and enabling the quantification of optimal parameters. Experiments on MSTAR and MiniSAR datasets show that the proposed ensemble compound transform attack method achieves an average target model recognition accuracy of 8.96% across different ensemble models and classification networks, improving the overall misleading effect by an average of 2.25% compared to other methods. Among them, the complex model increases by an average of 5.56%, while the proposed uniformity-guided multiobjective RIME algorithm improves the solution set diversity and convergence speed by over 25% compared with the comparison method. Using this method, the learned perceptual image patch similarity is maintained at 0.407 and the perturbation factor at 0.031, while classification accuracy decreases to 28.81%, demonstrating a tradeoff between misleading effectiveness and visual fidelity. This parameter maintains effective misleading performance under six different defense strategies, demonstrating strong robustness and providing a new approach and quantitative benchmark for adversarial attack research in SAR ATR.
Adversarial sample generation is a key research direction for uncovering the vulnerabilities of deep neural networks and improving the robustness of Synthetic Aperture Radar Automatic Target Recognition (SAR ATR) systems. This study proposes an optimal adversarial sample generation method for SAR ATR that jointly optimizes misleading effectiveness and fidelity, aiming to resolve the core contradiction between adversarial effectiveness and visual concealment. The generation process is modeled as a joint optimization problem with the goals of balancing “misleading” and “fidelity”. First, an integrated composite transform attack strategy is designed to enhance attack effectiveness, and a joint measurement model is developed that combines the classification accuracy of the target model with the Learned Perceptual Image Patch Similarity (LPIPS) to quantify the two optimization goals. Next, an improved uniformity-guided multiobjective RIME algorithm is proposed. By integrating the Tent chaotic map, hybrid dynamic weighting, and golden sine guidance, the model is efficiently solved, yielding a set of Pareto-optimal solutions that represent various tradeoff degrees. Finally, the YOLOv10 object detection network is employed to identify perturbations in the samples within the solution set, thereby locating the critical points where disturbances occur and enabling the quantification of optimal parameters. Experiments on MSTAR and MiniSAR datasets show that the proposed ensemble compound transform attack method achieves an average target model recognition accuracy of 8.96% across different ensemble models and classification networks, improving the overall misleading effect by an average of 2.25% compared to other methods. Among them, the complex model increases by an average of 5.56%, while the proposed uniformity-guided multiobjective RIME algorithm improves the solution set diversity and convergence speed by over 25% compared with the comparison method. Using this method, the learned perceptual image patch similarity is maintained at 0.407 and the perturbation factor at 0.031, while classification accuracy decreases to 28.81%, demonstrating a tradeoff between misleading effectiveness and visual fidelity. This parameter maintains effective misleading performance under six different defense strategies, demonstrating strong robustness and providing a new approach and quantitative benchmark for adversarial attack research in SAR ATR.
- First
- Prev
- 1
- 2
- 3
- 4
- 5
- Next
- Last
- Total:5
- To
- Go
微信 | 公众平台 
