Aug 2026· Remote Sensing· Vol 18, pp. 2687· 0 citations· 33 references
Abstract
Synthetic-aperture radar (SAR) ship detection is a fundamental task in maritime remote sensing, supporting wide-area surveillance, traffic monitoring, and emergency response under all-weather imaging conditions. Existing deep detectors mainly rely on spatial cues such as intensity, shape and context, but structured sea clutter and near-shore interference can still produce ship-like responses, while fine scattering details are weakened by deep downsampling. We address two practical representation limitations: incomplete preservation of shallow high-resolution details, and limited explicit modeling of local directional variation. To this end, we propose HMF-RTMDet, a shallow-neck spatial–frequency fusion detector. A P2 high-resolution path combines C2 features with upsampled P3 semantics. HybridMFBlock then processes the fused feature through a morphology branch and a trainable depthwise branch initialized by fractional Gabor templates, followed by channel-wise fusion. In the reported main HRSID run, HMF-RTMDet improves RTMDet-s from 67.9% to 72.6% in AP50:95, from 90.2% to 94.2% in AP50, and from 68.2% to 73.4% in APs. Across three runs, however, its AP50:95 is 72.17 ± 0.38%, comparable to the SFS-Conv and MCU-only controls. The evidence therefore identifies the P2 path as the main gain source but does not establish a stable advantage for HybridMFBlock over these controls. On SSDD, overall AP50:95 remains nearly unchanged and large-target performance decreases, defining an important boundary of the current design.
Synthetic aperture radar (SAR) ship detection in complex maritime scenes is challenged by speckle noise, sea clutter, coastal interference, weak small-target responses, and large-scale variations. To address these problems while maintaining low computational complexity, this article proposes CHL-YOLO, a lightweight detector based on YOLOv11n. The name CHL-YOLO corresponds to its three principal improvements: the Convolutional Gated Linear Unit (CGLU), the Hybrid-Scale Feature Pyramid Network (HSFPN), and Localization Quality Estimation (LQE). First, a C2PSA_CGLU module is introduced into the backbone to enhance spatial–channel feature representation and dynamically suppress redundant background responses. Second, HSFPN performs channel-wise feature selection before cross-scale aggregation and uses high-level semantic features to generate input-dependent gates for low-level spatial details, thereby reducing the repeated propagation of sea clutter. Third, the LQE branch estimates bounding-box localization reliability from discrete boundary distributions and combines the quality score with classification confidence during non-maximum suppression. Experiments on SSDD and HRSID demonstrate that CHL-YOLO contains only 1.83 M parameters and requires 5.1 GFLOPs, corresponding to reductions of 33.0% and 22.7%, respectively, compared with YOLOv11n. On SSDD, mAP@0.5:0.95 increases from 62.6% to 67.6%, while on HRSID it increases from 62.7% to 64.7%. The proposed model achieves inference speeds of 68.82 FPS and 67.58 FPS on the two datasets, respectively. These results demonstrate that CHL-YOLO achieves a favorable balance among detection accuracy, model complexity, and real-time inference for complex SAR ship detection.
Ship detection in synthetic aperture radar (SAR) imagery remains challenging because near-shore clutter, coherent speckle noise, dense scattering responses, and large target-scale variations often obscure vessel boundaries and weaken small-ship signatures. Although single-stage detectors provide efficient inference, their predominantly local convolutional modeling and fixed multiscale fusion strategies are insufficient for capturing long-range sea-surface context and adaptively emphasizing discriminative ship responses. To address these limitations, this paper proposes SeaMamba, a frequency-stabilized selective state-space multiscale detector for SAR ship detection in complex maritime scenes. Specifically, a frequency-domain speckle prior is introduced to stabilize SAR inputs while preserving target localization cues. A bidirectional selective state-space modeling module is then used to propagate long-range contextual information with input-adaptive scanning. Furthermore, a gated pyramid reassembly module is designed to refine multiscale features before dense prediction. The proposed method is evaluated on the SAR Ship Detection Dataset (SSDD) and High-Resolution SAR Images Dataset (HRSID) under a unified five-fold cross-validation protocol. SeaMamba achieved mean average precision at an intersection-over-union threshold of 0.5 (mAP@0.5) values of 99.16 ± 0.11% on SSDD and 93.74 ± 0.15% on HRSID. Per-category evaluation, ablation studies, efficiency analysis, and Grad-CAM-based interpretability visualization further demonstrate that SeaMamba improves small-vessel detection, suppresses near-shore false responses, and maintains a practical accuracy-efficiency trade-off.
Xiao-Peng Song, Zhong-Biao Sheng, Shi-Wei Li et al.· Electronics· 0 citations
Wide-area maritime surveillance typically requires a tradeoff between processing efficiency and imaging resolution. The 2-D asymmetric resolution mode (2-D-ARM) produces azimuth low-resolution imagery by compressing the effective azimuth Doppler bandwidth, thereby enabling efficient coarse search. However, this mode also introduces strong anisotropy: ship targets occupy only a few pixels along the azimuth direction and appear as range-elongated ultrathin stripe-like responses with scarce visual details, and high-IoU localization becomes unusually sensitive to pixel-level boundary deviations. For this problem, we construct ALR-SHIP with 367 offshore scenes and four paired azimuth-resolution subsets, each containing 2978 patches and 4032 ship instances. Furthermore, we propose directional statistics-guided input encoding (DSGIE), which incorporates azimuth-gradient and structural anisotropy cues while preserving main magnitude-channel statistics to enhance ultrathin target learnability. We also design an anisotropic boundary-aware loss (ABAL), which adaptively constrains azimuth-boundary regression to enhance high-IoU localization. Experiments show consistent high-IoU gains across azimuth-resolution settings.
Fan-Long Meng, Feng-Li Xue, Xiang-Yang Qi et al.· IEEE Geoscience and Remote S...· 0 citations
Multi-modal synthetic aperture radar (SAR)–optical object detectors raise detection accuracy by fusing complementary physical responses, but require both modalities to be simultaneously available at inference. When the optical stream becomes unavailable—under heavy cloud cover, night-time conditions, or downlink disruption—the detector reverts to SAR-only operation and accuracy degrades sharply. A natural remedy is to distil a multi-modal teacher into a SAR-only student via privileged-information knowledge distillation. However, we observe that the leading channel-wise feature-level method (CWD) reduces the student’s accuracy below the non-distilled baseline, with its smallest-target AP collapsing to near zero, because SAR speckle and target high-frequency edges share the same band and the alignment loss is dominated by broadband speckle energy. We refer to this failure mode as the speckle-fitting trap, formalize it as a gradient-pollution effect, and validate it through spectral and feature-manifold diagnostics. To counter the trap, we propose M-FSAD-KD, a full-link distillation framework whose neck-stage Fourier-gated alignment transfers low-frequency structural content while preserving target-edge high-frequency content; a joint spatial–channel attention mask, a shallow backbone adapter, and a response-level knowledge distillation (KD) term complete the chain. With a MAIENet teacher on OGSOD-1.0, the advantage of M-FSAD-KD over the strongest response-level KD baseline scales with student capacity: it matches KD on a 2.39 M-parameter student (both ≈48% mean average precision at an intersection-over-union (IoU) threshold of 0.5 (mAP50), averaged over multiple seeds) and exceeds it by 2.0 absolute points on a 19.98 M-parameter student (+8 over the non-distilled baseline), where it is the best of all distillation methods; at full convergence the 19.98 M-parameter student reaches 81.9% mAP50, within 8.9 absolute points of the multi-modal teacher. A frozen-feature transfer test to an out-of-domain SAR benchmark (SSDD ship detection) further shows that distilling from the multi-modal teacher yields substantially more transferable SAR features—about ten absolute points above the non-distilled backbone—with M-FSAD-KD transferring best. Cross-architecture validation with a dual-stream DEYOLO teacher yields 48.6% mAP50 at the student—1.1 absolute points below the MAIENet result—indicating that the framework transfers across the two representative teacher architectures tested (single-stream and dual-stream).
Yu-Ming Tong, Kai-Na Xiong, Jun Liu et al.· Remote Sensing· 0 citations
Ship detection is one of the important application directions of synthetic aperture radar (SAR) technology. Although significant progress has been made in recent years, it still faces challenges, such as the diverse sizes of ship targets, high aspect ratios, and the cross-shaped sidelobe artifacts, caused by strong scattering in SAR imaging. To address these challenges, this article proposed a multiscale rotation ship detection network for SAR images based on scale-aware Gaussian loss and directionally decoupled attention (SGDDNet). First, a global–local perception module was designed, which employs a dual-branch mechanism in the frequency and spatial domains to acquire global context while preserving local fine-grained features. This structure enables effective multiscale feature extraction for the network. Second, a direction decoupled attention module was proposed, which performs average and maximum pooling operations along the horizontal and vertical directions, respectively. This design suppresses the feature aliasing caused by side lobe interference from targets, thereby enhancing the feature discriminability of ship targets. Furthermore, an adaptive cross-layer fusion module was proposed, employing dynamic channel weighting strategies to achieve adaptive alignment and fusion of multilevel features. In addition, a scale-aware Gaussian loss function was designed. Through the collaborative modeling of Gaussian domain measurement and parameter space constraints, it effectively alleviated the interference of the scale difference of slender ships on the regression of rotation angles. On the HRSID and SSDD+ datasets, the proposed method attained average accuracies of 91.86% and 96.89%, respectively, demonstrating superior performance over existing SAR ship detection approaches.
Ke-Hong Liu, Ming Zhang, D. Yu et al.· IEEE Journal of Selected Top...· 0 citations
Synthetic aperture radar (SAR) is indispensable for maritime monitoring due to its all-weather, high-resolution imaging capabilities. However, inherent speckle noise, complex inshore sea clutter, and limited computational resources of edge devices prevent existing algorithms from balancing high accuracy with lightweight deployment. To address these challenges, an efficient, lightweight model termed FSP-YOLO was developed for complex backgrounds. First, a Frequency-Spatial Attention Module (FSAM) was designed; it incorporates a two-dimensional discrete cosine transform (2D-DCT) to decouple high-frequency target features from low-frequency redundant noise, precisely focusing on strong scattering centers. Second, a lightweight multi-scale feature fusion neck network utilizing Partial Convolution (PConv), designated as C3-PConv, was constructed to eliminate computational redundancy while preserving cross-scale representational capacity. Finally, the Minimum Point Distance Intersection over Union (MPDIoU) loss function optimized bounding box regression by minimizing corner geometric distances, significantly improving localization accuracy for densely clustered and weak targets. Extensive evaluations on the standard SAR Ship Detection Dataset (SSDD) and the High-Resolution SAR Images Dataset (HRSID) demonstrated that FSP-YOLO effectively alleviated missed detections under complex interference conditions. Compared with the YOLOv11n baseline, FSP-YOLO reduced the parameter count and computational complexity by 15.4% and 7.6%, respectively, resulting in only 2.2M parameters and 6.1 GFLOPs, while achieving mAP@50 scores of 98.3% on SSDD and 92.3% on HRSID. In addition, the proposed method achieved an inference speed of 118.2 FPS on an RTX 3090 GPU. These results indicate that FSP-YOLO achieves a favorable balance among detection accuracy, model complexity, and real-time inference capability, making it a promising lightweight solution for practical maritime surveillance and edge deployment on resource-constrained platforms.
Hongyang Wang, Duoqiang Li, Chao Wang et al.· IEEE Access· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
