2 min readfrom Frontiers in Marine Science | New and Recent Articles

Performance evaluation of YOLO models for target detection from ocean sidescan sonar imagery

Our take

Sidescan sonar imagery presents unique challenges for target detection, characterized by seabed clutter, low contrast, and frequently missed small targets. This study comparatively evaluates six YOLO models—YOLOv4, YOLOv6, YOLOv7, YOLOv9, YOLOv13n, and YOLO26n—across two distinct datasets (D1 and D2), utilizing metrics like mAP and FPS. Findings reveal a trade-off between model size and generalization, with lightweight models demonstrating superior noise resilience. Notably, YOLO26n achieves exceptional speed (526.32 FPS) and suitability for deployment on low-computing devices, as explored further in our Ocean Data
Performance evaluation of YOLO models for target detection from ocean sidescan sonar imagery

The challenges of underwater target detection using sidescan sonar imagery are significant, and this recent study provides a valuable, empirically-driven comparison of popular YOLO models to address them. Sidescan sonar, while crucial for mapping the seafloor and identifying submerged objects, produces data plagued by seabed clutter, low contrast, and the frequent loss of small targets – all factors that severely hinder automated detection. Current detection models often struggle to reconcile the competing demands of accuracy, speed, and the ability to be deployed on resource-constrained embedded systems. This research, evaluating YOLOv4, YOLOv6, YOLOv7, YOLOv9, YOLOv13n, and YOLO26n across two distinct datasets (D1 and D2), offers a much-needed clarity on which model best suits specific operational needs. It’s a vital contribution, especially when considered alongside efforts to improve ocean data accessibility and analysis, such as the Welcome to the Ocean Data Platform Academy | Demo, which aims to democratize access to sophisticated oceanographic tools. The findings underscore the importance of understanding how imaging quality directly impacts detection performance, a reality that necessitates careful dataset design and model selection.

The study’s emphasis on the trade-offs between model size, accuracy, and computational cost is particularly insightful. The authors correctly identify that larger, more complex models like YOLOv4, while potentially offering higher accuracy in ideal conditions, are often hampered by overfitting – a tendency to memorize noise textures in the training data, leading to poor generalization across different sea conditions. Conversely, the lightweight networks, YOLOv13n and YOLO26n, demonstrate remarkable robustness against noise, achieving stable detection results even in challenging environments. This aligns with broader trends in machine learning, where simpler models often outperform complex ones when data is limited or noisy. Furthermore, the specific recommendations for each model – YOLOv9 for high-precision underwater archaeology, YOLOv13n for large-area scanning with smaller vehicles, and YOLO26n for real-time deployment on low-computing devices – provide practical guidance for engineers and researchers. The application of these techniques will be essential for monitoring critical maritime areas, as demonstrated by recent events like the attacks on Saudi oil tankers in the Red Sea and Gulf of Aden, where robust and timely underwater surveillance capabilities are paramount Houthis Take Responsibility For Attacks On 2 Saudi Oil Tankers In Red Sea & Gulf Of Aden.

The quantitative evaluation using seven metrics – Precision, Recall, mAP@0.5, mAP50–95, FPS, FLOPs, and parameter count – provides a rigorous assessment of each model’s performance. The finding that YOLO26n achieves an impressive 526.32 FPS with remarkably low computational overhead (2.375M parameters and 5.2G FLOPs) is particularly noteworthy, highlighting its potential for real-time, embedded applications. This is a critical development, as it moves beyond purely laboratory-based analysis and towards practical deployment in underwater vehicles and portable sonar systems. The authors’ acknowledgement of the need for expanded datasets, encompassing diverse sea areas and incorporating noise augmentation strategies, is also crucial for further improving generalization and small target detection performance – areas where significant challenges remain, particularly in complex deep-sea environments. Understanding these trends also builds upon broader efforts to understand oceanographic data, such as the analysis of salinity trends over decades Linear trends in salinity for the World Ocean, 1955–1998 - Boyer - 2005 - Geophysical Research Letters - AGU Publications.

Looking ahead, the integration of these advanced detection models with increasingly sophisticated ocean data platforms represents a significant opportunity. The ability to automatically identify and classify underwater objects in real-time, with minimal computational resources, will unlock new possibilities for marine resource management, environmental monitoring, and underwater exploration. However, a key question remains: how can we ensure that these powerful tools are used responsibly and ethically, particularly in sensitive marine environments and with respect to potential impacts on marine life? As we continue to push the boundaries of ocean intelligence, it is imperative that we prioritize not only technological innovation but also the long-term health and sustainability of our oceans.

Sidescan sonar images suffer from strong seabed clutter, low gray contrast of underwater targets and frequent missed small targets. Current detection models struggle to balance detection accuracy, inference speed and embedded deployment. To solve these problems, this paper selects six mainstream one-stage detection models, YOLOv4, YOLOv6, YOLOv7, YOLOv9, YOLOv13n and YOLO26n, for comparative experiments. Two sidescan sonar datasets D1 and D2 with distinct imaging features are built. Seven metrics are used to quantitatively evaluate overall model performance: Precision, Recall, mAP@0.5, mAP50–95, FPS, FLOPs and parameter count. Experimental results show that imaging quality directly affects detection performance. The traditional heavy model YOLOv4 has numerous parameters and high computational cost. It produces many missed and false detections under heavy noise and fails to support real-time detection on underwater equipment. Lightweight YOLOv13n and YOLO26n show stronger anti-interference ability on noisy sonar images. Among all models, YOLOv6 on D1 achieves the highest overall mAP@0.5 of 0.944, YOLOv9 on D2 delivers the best localization accuracy across multiple IoU thresholds, and YOLO26n reaches an ultra-high inference speed of 526.32 FPS with only 2.375M parameters and 5.2G FLOPs. By comparing performance gaps between the two datasets, we find that parameter size is positively correlated with overfitting risk. Large models easily memorize unique noise textures from training data and have weak generalization. Restricted by limited parameter capacity, lightweight networks naturally suppress overfitting and achieve more stable detection results across different sea conditions. This paper clarify application scenarios for each model based on accuracy, real-time performance, lightweight degree and generalization. YOLOv4 is only suitable for offline analysis in laboratories. YOLOv9 fits high-precision contour mapping tasks such as underwater archaeology. YOLOv13n works for large-area scanning with medium and small underwater vehicles. YOLO26n has the best overall engineering performance. It can be deployed on low-computing devices including portable sonars and miniature underwater vehicles for long-term continuous real-time underwater detection. This study clarifies the performance trade-offs of various YOLO models for underwater target detection on side-scan sonar images. It provides references for model selection of underwater detection systems with different hardware and task requirements. Future work will expand sonar datasets covering multiple sea areas and adopt noise augmentation strategies to further boost generalization and small target detection performance in complex deep-sea environments.

Read on the original site

Open the publisher's page for the full experience

View original article