AquaPile-YOLO: Pioneering Underwater Pile Foundation Detection with Forward-Looking Sonar Image Processing

Zhongwei Xu, Rui Wang, Tianyu Cao, Wenbo Guo, Bo Shi, Qiqi GeRemote 2025 · DOI

TL;DR

This paper introduces AquaPile-YOLO, a novel deep learning model for detecting underwater pile foundations using forward-looking sonar images. By integrating multi-scale feature fusion, advanced attention mechanisms, and robust data augmentation, AquaPile-YOLO significantly improves detection accuracy and robustness, achieving a remarkable 96.89% accuracy rate. This advancement is crucial for efficient and safe marine engineering and environmental monitoring.

What Problem Does This Paper Address?

Accurately identifying underwater pile foundations is vital for marine engineering projects, such as harbor construction and infrastructure maintenance. Traditional methods, relying on human divers, are inefficient, costly, and pose safety risks. While sonar imaging offers a promising alternative, sonar images present unique challenges: high noise, poor contrast, and structural distortions, making precise detection difficult. Existing deep learning approaches, though advanced, still struggle with small, densely packed targets and varying environmental conditions. This paper aims to bridge this gap by developing a more robust and accurate deep learning model specifically for underwater pile foundation detection using forward-looking sonar.

Methods

The core of this work is the AquaPile-YOLO algorithm, an enhanced version of the YOLOv5 object detection framework. Key innovations include:

  • Multi-Scale Feature Fusion: To address the variability in target sizes, the model integrates feature maps from different network resolutions. This is achieved using a Feature Pyramid Network (FPN) structure, which combines deep semantic information with shallow spatial details to improve the detection of small targets.
  • Enhanced Attention Mechanisms: The model incorporates the Convolutional Block Attention Module (CBAM) into its C3 module (C3-CBAM). This module adaptively adjusts feature map weights, allowing the model to focus on salient regions and improving its response to targets in noisy and occluded underwater environments.
  • MPConv Module: This novel component processes input data through multiple parallel convolutional paths with varying kernel sizes and aspect ratios. This allows for the capture of a rich set of features and a comprehensive representation of underwater structures, aiding in the detection of targets with diverse scales and orientations.
  • C3N Module: Building on the C3 module, this component utilizes depth-separable convolutions and an inverted Bottleneck design inspired by ConvNeXt. This design enhances parameter utilization and feature correlation capture, proving beneficial for detecting small, densely packed targets.
  • Soft-NMS: To improve detection accuracy and robustness, especially for overlapping targets, the traditional Non-Maximum Suppression (NMS) is replaced with Soft-NMS. This method adjusts detection box scores using a Gaussian function instead of simply discarding them, reducing missed and false detections.
  • Data Augmentation: Techniques like random cropping, rotation, and noise addition are employed to increase the diversity of the training data, enhancing the model's generalization and robustness to varying underwater conditions.
  • Transfer Learning: Given the scarcity of sonar image data, the model is pre-trained on a large dataset (ImageNet) and then fine-tuned on the specific sonar image dataset, accelerating learning and improving performance.

Datasets

The study utilizes a custom dataset of 4000 forward-looking sonar images collected from field experiments. These images were generated from original acoustic signal data (.hca, .son formats) captured by the HY1645 imaging sonar, with a resolution of 848 × 600 pixels. The dataset was preprocessed, including grayscaling, noise removal, and contrast enhancement. For training, the dataset was split into 3000 images for training and 1000 for validation.

The paper also references the Underwater Acoustic Target Detection (UATD) dataset for comparative analysis, which includes identification results for various underwater objects like balls, cubes, and tires.

Key Insights and Ideas

  • Sonar Image Challenges are Unique: The paper highlights that sonar images, unlike optical images, lack true color and often present significant noise and distortion, requiring specialized approaches beyond standard computer vision techniques.
  • Attention is Key for Underwater Clarity: The integration of attention mechanisms (CBAM) proved crucial for the model to effectively "see" through the noise and occlusions inherent in sonar data, focusing on the subtle features of pile foundations.
  • Multi-Scale Fusion is Essential for Small Targets: The success in detecting pile foundations, which can be relatively small in sonar images, is strongly linked to the multi-scale feature fusion strategy, enabling the model to capture details at various resolutions.
  • Soft-NMS Significantly Improves Robustness: Moving beyond hard suppression in NMS to Soft-NMS was vital for handling overlapping detections, a common issue in dense underwater environments, leading to more reliable results.
  • A Large, Real-World Dataset Matters: The creation and use of a 4000-image dataset from field experiments significantly contribute to overcoming the data scarcity problem in sonar image analysis, providing a solid foundation for deep learning models.

Key Results

  • High Accuracy: The AquaPile-YOLO model achieved an impressive 96.89% accuracy rate for underwater target recognition in similar experimental scenarios.
  • Superior Performance: Compared to other object detection algorithms like SSD300, YOLOv3, Faster R-CNN, and Cascade R-CNN, AquaPile-YOLO demonstrated superior performance across key metrics:
  • Precision: 0.888
    • Recall: 0.798
    • mAP@50: 0.821
  • Fast Inference: The model achieved a high FPS (Frames Per Second) of 111.1, indicating its suitability for real-time applications.
  • Effective Ablation Studies: Experiments showed that combining multi-scale feature fusion, CBAM, Sonar Loss, and Soft-NMS led to significant improvements in mAP and recall, validating the contribution of each component.

Limitations and Future Work

  • Controlled Environment Testing: The current study primarily tested AquaPile-YOLO in controlled environments. Its performance in more variable natural settings with diverse water qualities and lighting conditions needs further exploration.
  • Computational Demands: The model's computational requirements might be a challenge for real-time applications in resource-constrained environments.
  • Future Directions:
    • Algorithm Optimization: Further refining the model structure to reduce computational load and improve detection speed.
  • Multimodal Data Fusion: Integrating sonar data with other sensors (e.g., optical, LiDAR) for a more comprehensive understanding of the underwater environment.
  • Broader Environmental Adaptability: Testing and adapting the model to a wider range of underwater conditions.
  • Automation and Intelligence: Developing automated sonar image collection systems integrated with underwater robots for fully autonomous detection.
  • Engineering Application Deployment: Integrating AquaPile-YOLO into existing underwater monitoring systems for long-term deployment and evaluation.

Summary updated

Abstract

Underwater pile foundation detection is crucial for environmental monitoring and marine engineering. Traditional methods for detecting underwater pile foundations are labor-intensive and inefficient. Deep learning-based image processing has revolutionized detection, enabling identification through sonar imagery analysis. This study proposes an innovative methodology, named the AquaPile-YOLO algorithm, for underwater pile foundation detection. Our approach significantly enhances detection accuracy and robustness by integrating multi-scale feature fusion, improved attention mechanisms, and advanced data augmentation techniques. Trained on 4000 sonar images, the model excels in delineating pile structures and effectively identifying underwater targets. Experimental data show that the model can achieve good target identification results in similar experimental scenarios, with a 96.89% accuracy rate for underwater target recognition.
BibTeX
@inproceedings{xu2025aquapile,
  author = {Zhongwei Xu and Rui Wang and Tianyu Cao and Wenbo Guo and Bo Shi and Qiqi Ge},
  title = {{AquaPile-YOLO: Pioneering Underwater Pile Foundation Detection with Forward-Looking Sonar Image Processing}},
  booktitle = {Remote},
  year = {2025},
  doi = {10.3390/rs17030360},
}
Figure