Gas Plume Target Detection in Multibeam WCI: Comparison
View Latest Version
Please note this is a comparison between Version 2 by Sirius Huang and Version 1 by Xiao WANG.

A multibeam water column image (WCI) can provide detailed seabed information and is an important means of underwater target detection. However, gas plume targets in an image have no obvious contour information and are susceptible to the influence of underwater environments, equipment noises, and other factors, resulting in varied shapes and sizes. 

  • multibeam water column image
  • gas plume
  • target detection
  • YOLOv7
  • deep residual aggregation structure
  • SimAM block

1. Introduction

Multibeam echo sounding (MBES) is a high-precision underwater data measurement technique [1]. Compared with the traditional single-beam echo sounding technique, MBES uses multiple transmitters and receivers to collect echo signals in multiple directions simultaneously, obtaining more precise water depth and water body data. This technique has been widely applied in marine resource exploration [2], underwater pipeline laying [3], and underwater environmental monitoring [4]. The water column image (WCI) formed by water body data is an important means of underwater target detection, and the gas plumes in the water may be an indication of the presence of gas hydrates in the nearby seabed sediment layers [5]. Development and excavation of these resources will play a crucial role in alleviating current global issues such as energy scarcity and environmental degradation. Therefore, how to detect and locate gas plumes quickly and accurately has become an important research topic.

2. Gas Plume Target Detection Methods

In traditional WCI target detection, the processing steps are image denoising, feature extraction, and target classification, in that order. Due to the influence of the working principle of MBES, some sidelobe beams are produced around the main lobe beam during transmission. When receiving the echo signal, the echo information of these sidelobe beams is mistaken for real signals, causing significant arc noise in the image [6], which is the main factor affecting image quality. In addition, the image also includes multisector noise and environmental noise. To make the collected raw data more accurate, [7] used weighted least squares to estimate the optimal beam incidence angle and corrected the difference in echo intensity under different water depths, thus obtaining normalized echo data. To effectively remove noise in the image, the image masking method was used to eliminate static noise interference in [8], and artificial thresholding was used to remove environmental noise. In [9], the data that fell within the minimum slant range were considered valid and used in the analysis, and the noise was removed by using the average echo intensity as the threshold. The second step in WCI target detection is to extract features from the denoised image. Feature extraction aims to obtain distinctive and representative image features, such as edges, morphology, and texture, to reduce data dimensions for subsequent classification. In [10], the authors used a clustering algorithm to extract information about regions containing gas plumes and then used feature histograms for feature matching to identify them. In [11], gas plume features were extracted using intensity and morphological characteristics to distinguish them from the surrounding environment. In [12], multiple features, such as color, gradient, and direction, were used for feature extraction, obtaining a set of feature vectors that can effectively distinguish between textures and nontextures. The final step is target classification. Based on the principle of feature invariance, the extracted features are input into a classifier for training. Commonly used classifiers include SVM (Support Vector Machine), Adaboost (Adaptive Boosting), and Random Forest. Then the training results are evaluated and optimized to achieve the high-precision detection of targets in the image. Traditional target detection methods for a WCI are complex, with human factors having a significant impact on image denoising and feature extraction algorithms being unable to extract representative features in complex images. The classifiers used in target classification are strongly influenced by lighting, angle, and noise, making it easy to miss or misidentify targets. Overall, target detection using traditional methods in a WCI is highly limited.
Convolutional neural networks (CNN) have proven effective for solving various visual tasks [13,14,15][13][14][15] in recent years, thus providing a new solution for multibeam WCI target detection. Compared to various machine learning classifiers, CNN not only automatically extracts image features, reducing human intervention, but can also learn more complex features, improving the model’s robustness. In addition, the end-to-end advantage and the introduction of transfer learning [16] have increased the efficiency and accuracy of the model, and it has been applied to different scenarios and tasks. CNN detectors can be divided into one-stage and two-stage, which differ mainly in the order of detection and classification. One-stage detectors refer to the prediction of target category and position directly from feature maps, such as YOLO (You Only Look Once) [17[17][18][19][20][21],18,19,20,21], SSD (Single-Shot Multi-Box Detector) [22], RetinaNet [23], and EfficientDet [24]. Among them, YOLO adopts a multiscale feature map and anchor mechanism, which can detect multiple targets simultaneously. SSD adopts feature maps of different scales and multiple detection heads, which can detect targets of different sizes. The focal loss function is used in RetinaNet to reduce the effect of target class imbalance and has achieved excellent detection results. EfficientDet adopts a scalable convolutional neural network structure based on EfficientNet [25] and performs well in speed and accuracy. In [26], to adapt to the particularity of the underwater environment, the author introduced the CBAM (Convolutional Block Attention Module) based on YOLOv4 to help find attention regions in object-dense scenes. In [27], the author established direct connections between different levels of feature pyramid networks to better utilize low-level feature information, thereby increasing the capability and accuracy of feature representation. To make the model smaller, in [28], by pruning and fine-tuning the EfficientDet-d2 model, the author achieved a 50% reduction in model size without sacrificing detection accuracy. Two-stage detector initially generates candidate frames and then uses them for classification and position regression, such as Faster R-CNN (Region-based Convolutional Neural Network) [29], Mask R-CNN [30], and Cascade R-CNN [31]. Among them, Faster R-CNN uses the RPN (Region Proposal Network) to generate candidate boxes and then uses RoI (Region of Interest) pooling to extract features for classification and regression. Mask R-CNN extends Faster R-CNN by incorporating segmentation tasks, enabling it to perform target detection and instance segmentation simultaneously. Cascade R-CNN enhances the robustness and accuracy of the detector through a cascaded classifier. Wang [32] enhanced the Faster R-CNN algorithm by implementing automatic selection of difficult samples for training, thereby improving the model’s ability to perform well and generalize on difficult samples. In [33], Song proposed Boosting R-CNN, a two-stage underwater detector featuring a new region proposal network for generating high-quality candidate boxes.
These methods have shown good improvements in their respective tasks but may not be applicable to gas plume object detection. This is because a gas plume generally consists of numerous bubbles, which are close together and interfere with each other. Compared to other objects, the reflection intensity of the gas plume is weaker, and the edges of the bubbles are not easily distinguishable. In addition, gas plumes in water often experience fracturing as they rise, which makes them difficult to accurately locate.

References

  1. Schimel, A.C.G.; Brown, C.J.; Ierodiaconou, D. Automated Filtering of Multibeam Water-Column Data to Detect Relative Abundance of Giant Kelp (Macrocystis pyrifera). Remote Sens. 2020, 12, 1371.
  2. Czechowska, K.; Feldens, P.; Tuya, F.; Cosme de Esteban, M.; Espino, F.; Haroun, R.; Schönke, M.; Otero-Ferrer, F. Testing Side-Scan Sonar and Multibeam Echosounder to Study Black Coral Gardens: A Case Study from Macaronesia. Remote Sens. 2020, 12, 3244.
  3. Guan, M.; Cheng, Y.; Li, Q.; Wang, C.; Fang, X.; Yu, J. An Effective Method for Submarine Buried Pipeline Detection via Multi-Sensor Data Fusion. IEEE Access. 2019, 7, 125300–125309.
  4. Zhu, G.; Shen, Z.; Liu, L.; Zhao, S.; Ji, F.; Ju, Z.; Sun, J. AUV Dynamic Obstacle Avoidance Method Based on Improved PPO Algorithm. IEEE Access. 2022, 10, 121340–121351.
  5. Logan, G.A.; Jones, A.T.; Kennard, J.M.; Ryan, G.J.; Rollet, N. Australian offshore natural hydrocarbon seepage studies, a review and re-evaluation. Mar. Pet. Geol. 2010, 27, 26–45.
  6. Liu, H.; Yang, F.; Zheng, S.; Li, Q.; Li, D.; Zhu, H. A method of sidelobe effect suppression for multibeam water column images based on an adaptive soft threshold. Appl. Acoust. 2019, 148, 467–475.
  7. Hou, T.; Huff, L.C. Seabed characterization using normalized backscatter data by best estimated grazing angles. In Proceedings of the International Symposium on Underwater Technology (UT04), Koto Ward, Tokyo, Japan, 7–9 April 2004; pp. 153–160.
  8. Urban, P.; Köser, K.; Greinert, J. Processing of multibeam water column image data for automated bubble/seep detection and repeated mapping. Limnol. Oceanogr. Methods 2017, 15, 1–21.
  9. Church, I. Multibeam sonar water column data processing tools to support coastal ecosystem science. J. Acoust. Soc. Am. 2017, 141, 3949.
  10. Ren, X.; Ding, D.; Qin, H.; Ma, L.; Li, G. Extraction of Submarine Gas Plume Based on Multibeam Water Column Point Cloud Model. Remote Sens. 2022, 14, 4387.
  11. Hughes, J.B.; Hightower, J.E. Combining split-beam and dual-frequency identification sonars to estimate abundance of anadromous fishes in the Roanoke River, North Carolina. N. Am. J. Fish. Manag. 2015, 35, 229–240.
  12. Fatan, M.; Daliri, M.R.; Shahri, A.M. Underwater cable detection in the images using edge classification based on texture information. Measurement 2016, 91, 309–317.
  13. Lu, S.; Liu, X.; He, Z.; Zhang, X.; Liu, W.; Karkee, M. Swin-Transformer-YOLOv5 for Real-Time Wine Grape Bunch Detection. Remote Sens. 2022, 14, 5853.
  14. Li, Z.; Zeng, Z.; Xiong, H.; Lu, Q.; An, B.; Yan, J.; Li, R.; Xia, L.; Wang, H.; Liu, K. Study on Rapid Inversion of Soil Water Content from Ground-Penetrating Radar Data Based on Deep Learning. Remote Sens. 2023, 15, 1906.
  15. Wu, J.; Xie, C.; Zhang, Z.; Zhu, Y. A Deeply Supervised Attentive High-Resolution Network for Change Detection in Remote Sensing Images. Remote Sens. 2023, 15, 45.
  16. Yosinski, J.; Clune, J.; Bengio, Y.; Lipson, H. How transferable are features in deep neural networks? NIPS 2014, 27, 3320–3328.
  17. YOLOv5 Models. Available online: https://Github.com/Ultralytics/Yolov5 (accessed on 13 January 2023).
  18. Ge, Z.; Liu, S.; Wang, F.; Li, Z.; Sun, J. Yolox: Exceeding yolo series in 2021. arXiv 2021, arXiv:2107.08430.
  19. Li, C.; Li, L.; Jiang, H.; Weng, K.; Geng, Y.; Li, L.; Wei, X. YOLOv6: A single-stage object detection framework for industrial applications. arXiv 2022, arXiv:2209.02976.
  20. Wang, C.Y.; Bochkovskiy, A.; Liao, H.Y.M. YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. arXiv 2022, arXiv:2207.02696.
  21. Redmon, J.; Divvala, S.; Girshick, R.; Farhadi, A. You only look once: Unified, real-time object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, 26 June–1 July 2016; pp. 779–788.
  22. Liu, W.; Anguelov, D.; Erhan, D.; Szegedy, C.; Reed, S.; Fu, C.Y.; Berg, A.C. Ssd: Single shot multibox detector. In Proceedings of the Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, 11–14 October 2016; pp. 21–37.
  23. Lin, T.Y.; Goyal, P.; Girshick, R.; He, K.; Dollár, P. Focal loss for dense object detection. In Proceedings of the IEEE international Conference on Computer Vision, Venice, Italy, 22–29 October 2017; pp. 2980–2988.
  24. Tan, M.; Pang, R.; Le, Q.V. Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, WA, USA, 14–19 June 2020; pp. 10781–10790.
  25. Tan, M.; Le, Q. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the International Conference on Machine Learning, Long Beach, CA, USA, 9–15 June 2019; pp. 6105–6114.
  26. Yu, K.; Cheng, Y.; Tian, Z.; Zhang, K. High Speed and Precision Underwater Biological Detection Based on the Improved YOLOV4-Tiny Algorithm. J. Mar. Sci. Eng. 2022, 10, 1821.
  27. Peng, F.; Miao, Z.; Li, F.; Li, Z. S-FPN: A shortcut feature pyramid network for sea cucumber detection in underwater images. ESWA 2021, 182, 115306.
  28. Zocco, F.; Huang, C.I.; Wang, H.C.; Khyam, M.O.; Van, M. Towards More Efficient EfficientDets and Low-Light Real-Time Marine Debris Detection. arXiv 2022, arXiv:2203.07155.
  29. Ren, S.; He, K.; Girshick, R.; Sun, J. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE TPAMI 2015, 28, 1137–1149.
  30. He, K.; Gkioxari, G.; Dollár, P.; Girshick, R. Mask r-cnn. In Proceedings of the IEEE International Conference on Computer Vision, Venice, Italy, 22–29 October 2017; pp. 2961–2969.
  31. Cai, Z.; Vasconcelos, N. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, 18–22 June 2018; pp. 6154–6162.
  32. Wang, H.; Xiao, N. Underwater Object Detection Method Based on Improved Faster RCNN. Appl. Sci. 2023, 13, 2746.
  33. Song, P.; Li, P.; Dai, L.; Wang, T.; Chen, Z. Boosting R-CNN: Reweighting R-CNN samples by RPN’s error for underwater object detection. NEUCOM 2023, 530, 150–164.
More
© Text is available under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/)
Video Production Service
Feedback