Abstract
Multibeam imaging sonars have application to monitoring fish and marine mammal presence and behaviours in the near field of tidal turbine installations, including evaluating avoidance, evasion, and potential blade strikes. Previous work in the Pathway Program has recommended use of the Tritech Gemini 720is, which demonstrated a high level of utility for visually detecting and tracking targets from vessel and bottom-mounted orientations in tidal flows up to approximately 2.5 m/s in Grand Passage, Bay of Fundy, Nova Scotia.
This project focuses on a comparison of two approaches for automated analysis of Tritech Gemini 720is sonar data: (1) an optical-based deep learning detection approach led by Dr. James Joslin, and (2) an approach based on spatial and temporal filtering for target detection and tracking led by Dr. Benjamin Williamson. The scope of this project was developed based on a practical need to increase efficiency in sonar data assessment, working toward methods that can incorporate reliable automation. The project goal is to advance the development of automated methods for detecting, tracking, and classifying acoustic targets in high energy tidal flows. The results will help inform the Department of Fisheries and Oceans Canada, tidal energy developers, and other stakeholders in the design and implementation of effective monitoring systems for tidal energy projects in the Bay of Fundy and beyond.
The primary datasets for analysis include data from an upward-oriented, seafloor-mounted Gemini collected in Grand Passage in October 2020, and from a downward-oriented, vessel-mounted Gemini from Minas Passage collected by SOAR on September 1, 2021. Both of these datasets focus on artificial targets: a V-Wing glider from Dartmouth Ocean Technologies, a ca. 10 cm diameter basalt rock, and a 0.45 kg lead fishing weight. Additionally, a preliminary analysis of bottom-mounted Gemini data from the FORCE site collected by HTEL Energy in February 2022 is included as a case study on algorithm performance. In all cases, the effective range of the Gemini sonar was found to be 30 m to 40 m depending on the size of the target and environmental conditions (bubbles, sediment, zooplankton, and other acoustic scatterers). Beyond this effective range targets are not easily visible amongst the background noise, although this is site, target and orientation specific, to some extent.
The analysis methodology was developed to evaluate the performance of the two methods based on two key metrics: “precision”, here defined as the portion of all predicted targets which were true targets, and “recall”, also known as the true positive rate, evaluates what portion of targets in a database were found by the algorithm.
Both the optical-based deep learning method and the spatial and temporal filtering method are compared to data based on manual annotations made by a trained technician using Tritech’s proprietary SeaTec data acquisition and processing software. SeaTec’s built-in processing tools for automated object detection and tracking are also evaluated and used in the comparison. The Gemini SeaTec software has proven to be reliable for instrument setup and data collection with a user- friendly interface. The software writes proprietary Gemini .ecd files, though the raw data can be accessed programmatically for conversion and potential compression into alternate formats for storage and analysis. Although lossless conversion was used in this project, note that any conversion or compression may cause data losses which can significantly affect data analysis.