SAR Image Object Detection and Information Extraction: Methods and Applications

1. Introduction

Synthetic aperture radar (SAR) is now recognized as a critical source of observational data in domains such as military reconnaissance, maritime monitoring, and disaster response, owing to its ability to deliver fine spatial resolution and broad-area imaging irrespective of weather or daylight conditions [1]. Unlike conventional optical sensors, SAR systems are unaffected by environmental factors such as cloud cover or varying illumination, which allows them to provide uninterrupted and reliable observations. However, the unique imaging mechanisms and high noise levels of SAR images continue to present major obstacles to object detection and information extraction. Recently, accelerated progress in deep learning and AI has stimulated the development of a wide range of SAR-oriented algorithms for target detection and information mining, fostering major advances in automated scene understanding and smart processing in challenging environments.

Despite the substantial boost in SAR target detection accuracy achieved by deep learning, traditional network designs often entail substantial complexity and computational demands, which constrain their suitability for embedded platforms and practical operational settings. In response, researchers have developed a variety of streamlined detection frameworks, leveraging novel backbone structures, efficient feature integration, and automated pruning, to achieve competitive accuracy with lower memory footprint and reduced computation. Given the scarcity of labeled SAR data and restricted dataset scales, weakly supervised and unsupervised domain adaptation techniques have become critical approaches for improving the generalization capability of SAR object detection [2]. Meanwhile, to cope with special challenges such as arbitrary direction, large scale changes, and the complex backgrounds of targets such as ships, researchers have developed arbitrary-direction detection algorithms [3] and alleviated the domain offset problem under different sensors and imaging conditions through cross-modal and cross-sensor feature adaptation. Within SAR target recognition, advances in deep learning have led to the increasingly accurate and rapid automated identification of maritime and aerial objects, broadening practical application prospects in real-time scenarios. To address challenges including continual category expansion, sample scarcity, and the need for robustness in complex scenarios, strategies such as incremental learning, scattering center graph structure modeling, transformer integration, and unsupervised domain adaptation have been proposed, greatly enhancing model adaptability and recognition accuracy in dynamic environments. The area of SAR target segmentation is also advancing rapidly. Multi-scale feature enhancement networks, exemplified by MFE-Net, employ direction-aware segmentation heads and multi-scale feature extraction to achieve high-precision boundary segmentation in complex near-shore scenes. In addition, to tackle the challenge that SAR ship wake detection is easily affected by ocean clutter, speckle interference, and insufficient annotated data, researchers have proposed a feature-guided self-supervised denoising method, Wake2Wake [4], which markedly improves both the reliability and practical effectiveness of wake detection. In the field of multi-target tracking, integrating lightweight detection with multi-target tracking techniques has enabled efficient tracking of short-time sequence targets in SAR images [5].

This Special Issue covers a variety of cutting-edge technologies for SAR image object detection and information extraction, with 11 articles covering multiple directions, such as detection, recognition, tracking, and interference analysis. A brief overview of these contributions is provided, as outlined below.

2. An Overview of Published Articles

The research by Xu et al. (Contribution 1) investigates how to achieve an optimal trade-off between detection speed and accuracy for moving target shadow identification in video SAR imagery, proposing a lightweight detector named MambaShadowDet. By leveraging the complementary strengths of the state-space model (SSM) and the YOLOv8 architecture, this approach introduces the Mamba-Backbone for long-range context modeling and streamlines complexity with the Slim-PAFPN module, while preserving robust multi-scale feature extraction. Validation experiments demonstrate that the proposed method delivers significant improvements in both F1-score and processing efficiency on the SNL video SAR dataset, striking an effective balance between accuracy and real-time performance. This study highlights the considerable potential of state-space modeling for the intelligent interpretation of video SAR data.

The article by Chen et al. (Contribution 2) focuses on the challenges of imaging maritime targets in the presence of corner reflector interference. Unlike traditional imaging methods that mainly address a single type of target, the study addresses complex scenarios where ships and corner reflectors coexist, establishes motion models for both types of targets, and proposes a time-frequency processing approach that integrates CR-QCRD estimation with an improved Clean algorithm to tackle Doppler aliasing and defocusing issues. It is worth noting that the authors introduce a novel termination criterion based on spectral kurtosis, which improves both iteration efficiency and robustness. This method not only enhances the accuracy of scattering point extraction in interference environments but also advances the practical application of unified SAR imaging in complex interference backgrounds.

The third article in this Special Issue, written by Zhang et al. (Contribution 3), addresses the coding efficiency problem in rotating object detection for SAR images and proposes a two-stage detection scheme called BurgsVO. This study introduces the Burgs equation to construct a feature heuristic module that optimizes the feature extraction stage and proposes a diagonal vertex offset (ADVO) encoding method, which simplifies the representation of rotated bounding boxes into a six-dimensional parameter space, thereby significantly reducing the model’s computational burden. Compared with most rotation detection algorithms that focus primarily on high accuracy while overlooking model lightweight design, BurgsVO achieves both efficient feature encoding and accelerated inference speed, offering a practical solution for rotating ship identification in complex SAR scenarios.

The fourth article, written by Zhao et al. (Contribution 4), addresses the challenge of detecting new target classes in SAR images under small-sample conditions. The study proposes the SCGF detection framework, which integrates a context enhancement module, robust Gaussian flow representation, and a category difference aggregation mechanism to build a small-sample detection model with strong generalization capability. The study particularly emphasizes stable modeling of the distribution of new classes under sparse sampling conditions and guides high-dimensional semantic alignment through prototype matching strategies. Experiments demonstrate that this approach achieves strong target discrimination in cross-domain tasks, providing a generalized detection paradigm well suited to resource-scarce SAR imagery.

The research by Hu et al. (Contribution 5) systematically analyzes the insufficient exploitation of ship scattering characteristics in polarimetric SAR images and proposes guiding a context aggregation network to effectively distinguish ships from backgrounds by introducing two novel feature components: local spiral scattering (HSE) and edge intensity difference (MSID). Unlike previous methods that rely solely on pseudo-color images or simple polarization channel stacking, this approach emphasizes the complementarity of structural and boundary information in its design. By introducing a component for scattering feature aggregation and extraction, the study substantially improves the extraction of salient target features and the suppression of complicated scene interference, thereby advancing the deep interpretation of polarimetric SAR data.

Guan et al.’s study (Contribution 6) focuses on the persistent challenge of identifying small-scale objects within SAR imagery and proposes an enhanced network based on the YOLO architecture. Its key improvements include the combination of channel rearrangement and convolution reparameterization to form an SR module for efficient feature extraction, the use of an SPD structure to optimize the information down-sampling path, a hybrid attention mechanism to enhance responsiveness to small ships, and the incorporation of shape-NWD loss to improve boundary sensitivity during small-target detection. The proposed approach yields a substantial enhancement in recognizing small vessels over several SAR image collections, thereby verifying the synergistic effect of structural lightweight design and enhanced target perception strategies.

The seventh study, written by Zhu et al. (Contribution 7), motivated by practical requirements for intelligent suppression interference evaluation in SAR images, constructs a detection-centric interference effectiveness evaluation system. For the first time, the authors divide the evaluation process into two stages, determining whether the target is detectable or undetectable, and propose two quantitative indicators, target exposure and interference concealment, to characterize the impact of different interference levels on object detection performance. Experimental results show that each 0.1 increase in suppression leads to a 35% to 45% reduction in the detection rate. The study redefines the logic of SAR interference evaluation from the perspective of task outcomes and provides important guidance for SAR image quality assessment in electronic countermeasure applications.

In the eighth article, Wang et al. (Contribution 8) propose an angle-controllable image generation approach utilizing a GAN-based image synthesis framework to address the challenges of uneven data and insufficient target azimuth distribution in SAR images. The method integrates a local feature alignment mechanism with a sparse azimuth synthesis model and introduces an angle discriminator to supervise the perspective consistency of the generated images. By enabling the generation of adjustable azimuth images across multiple target categories, the study not only enriches the diversity of training data but also enhances the robustness of subsequent recognition tasks to perspective variations. These results are of practical significance for SAR target data augmentation and incremental learning.

The article written by Guan et al. (Contribution 9) focuses on the challenges of detecting multi-directional ships in SAR images caused by contour discontinuities and noise interference and proposes an object detection method that combines edge deformable convolution with point set geometric modeling. The innovation lies in the use of deformable convolution to adaptively capture discontinuous target regions, while the mapping between point sets and rotated bounding boxes is established via radial basis functions, enabling high-precision contour reconstruction and pose prediction. This approach demonstrates strong detection robustness in densely populated ship regions and provides a novel methodology for modeling complex structured targets.

The tenth work, written by Ren et al. (Contribution 10), develops LPFFNet, an efficient SAR ship detection approach tailored for practical deployment requirements. The model is customized at multiple levels, including the backbone network, feature fusion, attention mechanism, and convolutional structure. It introduces modules such as multi-channel enhancement, channel prior fusion, and residual attention and incorporates the SWF strategy into the loss function to more accurately focus on hard-to-classify samples. This approach achieves excellent detection performance while maintaining computational efficiency, demonstrating the feasibility and effectiveness of a multi-scale lightweight detection framework guided by prior knowledge.

The final article, written by Qian et al. (Contribution 11), proposes a cross-layer adaptive feature aggregation network, CLAFANet, to address the challenges of detecting ships in SAR images with arbitrary orientation, multi-scale variations, and diverse distributions. Its core innovations include the CLAFA module, driven by feature similarity, which alleviates the representational inconsistency present within feature pyramids, and the frequency selective phase shift encoder (FSPSC), which resolves the periodic ambiguity issue in angle regression. Experimental results demonstrate that this method outperforms existing rotation detection approaches on multiple public datasets and exhibits outstanding adaptability and accuracy in complex maritime scenarios.

3. Conclusions

This Special Issue includes 11 high-quality research studies, which systematically present the latest advances in methodological innovation and real-world implementation within the field of SAR image target detection and information extraction. These studies encompass a series of key tasks in SAR images, including object detection, recognition, tracking, and interference effectiveness evaluation, fully reflecting the breadth of research topics and the diversity of technical approaches within this domain. From lightweight network architecture design and polarization feature mining to intelligent small-sample detection, cross-modal domain adaptation, and generative sample augmentation, the studies in this Special Issue highlight innovative solutions proposed by various scholars to longstanding challenges such as complex backgrounds, small-sample scarcity, and efficient model deployment in SAR imagery. Collectively, these contributions greatly enhance the applicability and scalability of intelligent SAR image processing technologies.

Notably, several studies in this Special Issue have actively expanded into non-traditional application scenarios, including video SAR, polarimetric SAR, complex interference environments, and angle-controllable image generation. These explorations underscore the ongoing evolution of SAR intelligent interpretation toward frontier domains such as dynamic scene understanding, high-complexity environment analysis, and high-precision recognition under limited sample conditions. Furthermore, research themes such as multi-task collaboration, lightweight modeling, and small-target detection have continued to attract substantial attention, fully illustrating the concerted efforts of the research community to bridge SAR intelligent interpretation methodologies with practical, real-world applications.

In the future, with the continuous evolution of artificial intelligence technology, especially the continuous breakthroughs in emerging paradigms such as generative large models (such as multimodal basic models) and embodied intelligence, SAR object detection and information extraction are expected to usher in new leapfrog development opportunities. A key research frontier will be leveraging the semantic modeling capabilities of large-scale models and seamlessly integrating structural perception and temporal reasoning into SAR information flow processing, thereby driving future innovation in this field.

As a final note, this Special Issue offers a strong theoretical underpinning and valuable practical reference for the advancement of intelligent SAR image analysis. We warmly invite the research community to remain engaged with the latest developments in this field and to actively contribute to future Special Issues. By fostering collaboration and promoting the deep integration of SAR intelligent interpretation with autonomous cognitive technologies, we can collectively advance the capabilities of SAR remote sensing systems for broader and more intelligent applications across diverse platforms and scenarios.