Wild salmon enumeration and monitoring using deep learning empowered detection and tracking

• Published in: Frontiers in Marine Science
• Main Authors: William I. Atlas, Sami Ma, Yi Ching Chou, Katrina Connors, Daniel Scurfield, Brandon Nam, Xiaoqiang Ma, Mark Cleveland, Janvier Doire, Jonathan W. Moore, Ryan Shea, Jiangchuan Liu
• Format: Article in Journal/Newspaper
• Language: English
• Published: Frontiers Media S.A. 2023
• Subjects: computer vision; deep learning; fisheries management; in-season fishery management; indigenous science; wild salmon; Science; Q; General. Including nature conservation; geographical distribution; QH1-199.5; British Columbia; Canada; Kitwanga; Pacific; Sockeye; First Nations
• Online Access: https://doi.org/10.3389/fmars.2023.1200408; https://doaj.org/article/7b9c5ab60640430a8c870d8f81462ef7

Pacific salmon have experienced declining abundance and unpredictable returns, yet remain vital to livelihoods, food security, and cultures of coastal communities around the Pacific Rim, creating a need for reliable and timely monitoring to inform sustainable fishery management. Currently, spawning salmon abundance is often monitored with in-river video or sonar cameras. However, reviewing video for estimates of salmon abundance from these programs requires thousands of hours of staff time, and data are typically not available until after the fishing season is completed. Computer vision deep learning can enable rapid and reliable processing of data, with potentially transformative applications in salmon population assessment and fishery management. Working with two First Nations fishery programs in British Columbia, Canada, we developed, trained, and tested deep learning models to perform object detection and multi-object tracking for automated video enumeration of salmon passing two First Nation-run weirs. We gathered and annotated more than 500,000 frames of video data encompassing 12 species, including seven species of anadromous salmonids, and trained models for multi-object tracking and species detection. Our top performing model achieved a mean average precision (mAP) of 67.6%, and species-specific mAP scores > 90% for coho and > 80% for sockeye salmon when trained with a combined dataset of Kitwanga and Bear Rivers’ salmon annotations. We also tested and deployed a prototype for a real-time monitoring system that can perform computer vision deep learning analyses on site. Computer vision models and off-grid monitoring systems show promise for automated counting and species identification. A key future priority will be working with stewardship practitioners and fishery managers to apply salmon computer vision, testing and applying edge-capable computing solutions for in-situ analysis at remote sites, and developing tools for independent user-led computer vision analysis by non-computer scientists. ...