New AI Model Spots Ships Hiding in Radar Images Despite Noise and Clutter

Ships can slip past even the sharpest human eye when the only view comes from a radar satellite. Synthetic aperture radar, or SAR, images the ocean surface with microwave pulses that work through clouds, darkness and storms, which makes the technology a backbone of maritime surveillance. Yet the same images that reveal vessels in bad weather also bury them in speckle noise, sea clutter and background interference, while ship sizes vary from tiny fishing boats to large cargo vessels. A research team now reports an improved detection model, called MFF-YOLO, that raises detection accuracy on two widely used SAR ship benchmarks by upgrading the popular YOLOv8n architecture with three targeted innovations.

The work, published as an open access article in July 2025, tackles a task that matters well beyond the laboratory. Ship detection in SAR images supports maritime search and rescue, traffic monitoring and the detection of illegal fishing, activities that depend on reliably spotting vessels in imagery that is often noisy and cluttered. The authors note that traditional detection methods, such as CFAR and its Gaussian model-based and wavelet-based variants, rely on manually designed classifiers. These approaches are computationally efficient but struggle with complex scenarios like cluttered backgrounds and varying target sizes, and their robustness and scalability are limited as SAR data volumes grow.

Deep learning has changed that picture. Convolutional neural networks and models in the YOLO family learn features automatically from data, handle complex backgrounds and varying ship sizes, and offer real-time processing with direct predictions. The new study builds directly on YOLOv8n, a lightweight member of that family whose architecture splits into a backbone that extracts feature maps, a neck that refines multi-scale representations and a head that predicts bounding boxes, class labels and confidence scores. The researchers modified all three stages of the pipeline rather than treating the detector as a fixed black box.

The first innovation is the Multi-Scale Dual Attention module, or MDSA, which applies joint attention across channel and spatial dimensions. Its ConvSE submodule, inspired by the Squeeze-and-Excitation block, recalibrates channel-wise feature responses using global context, so that channels most relevant to ship detection are emphasized. Its Multi-Scale Spatial Attention submodule computes average-pooled and max-pooled feature maps, concatenates them, and processes the result with convolutional kernels of 3 by 3, 5 by 5 and 7 by 7 pixels to build attention maps at multiple receptive fields, which are then averaged. The authors state that this joint application of channel-wise recalibration and multi-scale spatial emphasis improves feature focus and target localization under challenging SAR imaging conditions, providing support for detecting small and low-contrast ships. In the network, MDSA is integrated with the C2f module and is also added between each C2f module and the detection head in the neck, helping the model focus on key features while reducing background clutter.

The second innovation, the Frequency-Aware Feature Fusion module or FAFF, replaces some conventional convolutions in the backbone. It runs two branches in parallel: a local branch that applies a 3 by 3 convolution to capture fine spatial structures, and a global frequency branch that performs a two-dimensional Fast Fourier Transform on each channel. A soft thresholding mechanism based on a normalized sigmoid function assigns continuous weights to frequency components, emphasizing salient frequencies while attenuating less informative ones, and the masked spectrum is transformed back to the spatial domain with an inverse FFT before being fused with the local features through a 1 by 1 convolution. According to the paper, the processed feature maps show reduced background clutter and better preservation of high-frequency object details such as edges and contours, and the frequency-aware enhancement also sharpens the attention maps produced downstream by MDSA.

The third change concerns training rather than architecture. Standard IoU-based loss functions such as DIoU, EIoU, GIoU, SIoU and CIoU can suffer from gradient vanishing when predicted and ground-truth boxes overlap poorly, and they do not adjust their optimization focus based on sample quality. The team adopted Wise IoU v3, a loss function that dynamically adjusts gradient contributions based on the outlier degree of anchor boxes, prioritizing medium-quality predictions and reducing the influence of low-quality samples. A quality parameter derived from the ratio of an exponential moving average of the IoU loss to the current IoU loss distinguishes high-quality from low-quality anchors, with hyperparameters shaping the focusing function. The authors report that WIoU v3 improves localization accuracy and robustness, particularly for small or hard-to-detect targets in noisy environments, and that its benefits are amplified when applied to feature maps already enhanced by FAFF and MDSA.

The performance numbers come from experiments on two benchmark datasets. The SSDD dataset, created for ship detection tasks, contains 1,160 images and 2,456 annotated ship targets, with all images resized to 640 by 640 pixels for the experiments. The LS-SSDD dataset focuses on offshore environments and presents small-scale, low-contrast targets under complex sea conditions, being a synthesized dataset that simulates lower contrast, simplified object shapes and spatial ambiguity between targets and background. Both datasets carry noise-related challenges: real SAR imagery shows granular speckle from coherent signal interference and clutter from sea surface scattering, while the synthetic data adds compression artifacts and resolution limitations that degrade feature clarity.

On SSDD, the improved model reached a Precision of 97.11 percent, a Recall of 96.80 percent, an mAP at an IoU threshold of 0.5 of 98.33 percent, and an mAP averaged over IoU thresholds from 0.5 to 0.95 of 73.10 percent. Relative to the baseline YOLOv8n, which scored 94.79 percent Precision, 93.38 percent Recall, 95.28 percent mAP0.5 and 67.78 percent mAP0.5:0.95, those figures represent gains of 2.32, 3.42, 3.55 and 5.32 percentage points respectively. On LS-SSDD, the model achieved a Precision of 86.91 percent, Recall of 80.79 percent, mAP0.5 of 78.84 percent and mAP0.5:0.95 of 28.17 percent, gains of 2.66, 2.94, 5.73 and 1.12 percentage points over the baseline. The model also compared favorably with other detectors on SSDD, including RT-DETR, Faster R-CNN, HR-SDNet and several YOLO variants, although YOLOv10n posted a slightly higher mAP0.5 of 98.59 percent while trailing on the stricter averaged metric.

Ablation experiments traced the gains to each component. Adding MDSA alone to YOLOv8n raised Precision by 0.29 percentage points, Recall by 0.86, mAP0.5 by 1.28 and mAP0.5:0.95 by 0.34, while FAFF alone lifted mAP0.5 to 96.33 percent and mAP0.5:0.95 to 68.33 percent. WIoU v3 produced the largest single-module improvement, raising Precision to 96.67 percent, Recall to 94.43 percent and mAP0.5 to 96.97 percent without modifying feature extraction. Combining FAFF and MDSA yielded stronger results than either alone, and the full three-module configuration achieved the best overall scores, which the authors describe as a synergistic effect in which frequency enhancement, attention refinement and regression optimization complement each other. The gains came at a computational cost: MFF-YOLO requires 12.6 GFLOPs and 4.611 million parameters, more than YOLOv8n’s 8.2 GFLOPs and 3.211 million parameters, though far less than Faster R-CNN at 152.28 GFLOPs or Grid R-CNN at 253.17 GFLOPs.

Visualization with Grad-CAM attention heatmaps offered a qualitative view of the changes. On SSDD, the baseline model produced more missed detections and false positives in cluttered scenes, while MFF-YOLO concentrated attention on ship structures such as masts, decks and hull edges; in one example the baseline detected a target with a confidence of 0.70 and dispersed attention, whereas MFF-YOLO reached 0.83 with attention focused on the ship’s superstructure. On LS-SSDD, both models missed some near-shore small targets, and in a high-noise scene the attention maps of both became scattered, with MFF-YOLO maintaining only a slightly higher confidence of 0.29 versus 0.26. The authors acknowledge the limits of their approach, noting that challenges remain in extremely cluttered environments and for small targets where noise may still affect accuracy, and they plan to explore more effective denoising techniques to further improve robustness.

Subject of Research: An improved YOLOv8n deep learning model for detecting ships in synthetic aperture radar images

Article Title: MFF-YOLO: enhanced ship detection in noisy synthetic aperture radar images with dynamic loss adjustment

Article References: Yang, Y., & He, Y. (2025). MFF-YOLO: enhanced ship detection in noisy synthetic aperture radar images with dynamic loss adjustment. Discover Imaging, 2(1), Article 10. https://doi.org/10.1007/s44352-025-00016-z

Image Credits: AI Generated

DOI: 10.1007/s44352-025-00016-z

Keywords: synthetic aperture radar, ship detection, deep learning, YOLOv8n, attention mechanism, frequency domain, loss function, maritime surveillance, computer vision, object detection, MFF-YOLO, enhanced