Automatic single fish detection with a commercial echosounder using YOLO v5 and its application for echosounder calibration

Frontiers in Marine Science

Marine Science

Minghua Xue¹  Zhenhong Zhu¹  Weiqi Wang¹  Jianfeng Tong²  Siqian Tian³  Jun Han⁴ 

[1] College of Marine Sciences, Shanghai Ocean University, Shanghai, China;College of Marine Sciences, Shanghai Ocean University, Shanghai, China;Key Laboratory of Marine Ecological Monitoring and Restoration Technologies, Ministry of Natural Resources (MNR), Shanghai, China;National Engineering Research Center for Oceanic Fisheries, Shanghai Ocean University, Shanghai, China;College of Marine Sciences, Shanghai Ocean University, Shanghai, China;National Engineering Research Center for Oceanic Fisheries, Shanghai Ocean University, Shanghai, China;Key Laboratory of Sustainable Exploitation of Oceanic Fisheries Resources, Ministry of Education, Shanghai, China;School of Electrical Engineering and Telecommunications, The University of New South Wales, Sydney, NSW, Australia;

关键词: fishing vessel;    automatic detection;    commercial echosounder calibration;    Cololabis saira;    deep learning;    single fish detection;   

DOI  :  10.3389/fmars.2023.1162064

 received in 2023-02-09, accepted in 2023-05-22,  发布年份 2023

来源: Frontiers

【 摘 要 】

Nowadays, most fishing vessels are equipped with high-resolution commercial echo sounders. However, many instruments cannot be calibrated and missing data occur frequently. These problems impede the collection of acoustic data by commercial fishing vessels, which are necessary for species classification and stock assessment. In this study, an automatic detection and classification model for echo traces of the Pacific saury (Cololabis saira) was trained based on the algorithm YOLO v5m. The in situ measurement value of the Pacific saury was measured using single fish echo trace. Rapid calibration of the commercial echo sounder was achieved based on the living fish calibration method. According to the results, the maximum precision, recall, and average precision values of the trained model were 0.79, 0.68, and 0.71, respectively. The maximum F1 score of the model was 0.66 at a confidence level of 0.454. The living fish calibration offset values obtained at two sites in the field were 116.30 dB and 118.19 dB. The sphere calibration offset value obtained in the laboratory using the standard sphere method was 117.65 dB. The differences between in situ and laboratory calibrations were 1.35 dB and 0.54 dB, both of which were within the normal range.

【 授权许可】

Unknown   
Copyright © 2023 Tong, Wang, Xue, Zhu, Han and Tian