U.S. flag An official website of the United States government.
Official websites use .gov

A .gov website belongs to an official government organization in the United States.

Secure .gov websites use HTTPS

A lock ( ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites.

i

Automated fish detection in videos to support commercial fishing sustainability and innovation in the Alaska walleye pollock (Gadus chalcogrammus) trawl fishery



Details

  • Journal Title:
    ICES Journal of Marine Science
  • Personal Author:
  • NOAA Program & Office:
  • Description:
    Bycatch reduction devices (BRDs) are used in the Alaska walleye pollock (Gadus chalcogrammus) fishery to reduce Pacific salmon (Oncorhynchus spp.) bycatch. Evaluation of BRD effectiveness often requires people to process collected or live-feed video. Deep learning can be used to detect and classify fish in video to support BRD and other fisheries bycatch work. We fine-tuned and evaluated the detection model EfficientDet and YOLO11 to find salmon and pollock in videos collected inside a trawl using 11 572 salmon and 73 394 pollock annotations from 16 989 video frames. We evaluated model performance across all data and during high abundances of krill, varying fish density, camera occlusions, low lighting, and combinations of these using five-fold cross validation. The best performing model was further evaluated by applying it to videos from three fishing tows not used for model training, using it in a salmon presence algorithm that was developed to assess whether an efficient semi-automated video review process was feasible, and comparing it with the performance of a salmon-only detection model. We found that the YOLO models performed better than EfficientDet and on average detected 90% of salmon and pollock with 72% accuracy using a 50% detection overlap threshold. The YOLO models performed comparably to annotators for fish detection: the detection performance was higher for pollock and lower for salmon than the variability measured between annotators. Model performance across trawl and video conditions was more variable for salmon and generally lowest during high fish densities. The YOLO salmon and pollock model performed better than the salmon-only model when using an optimal confidence score threshold. When applied to full fishing tows, the YOLO salmon and pollock model incorrectly detected Pacific herring (Clupea pallasi) as salmon, and correctly predicted 99.3% of salmon presences while reducing the number of video frames needing to be reviewed by 85%. Overall, the models detected salmon and pollock well inside a pollock trawl, but camera placement, lighting, and occlusions presented challenges. We provide our annotated dataset, salmon presence algorithm, and recommendations for optimising video quality in trawls.
  • Source:
    ICES Journal of Marine Science, 82(9)
  • DOI:
  • ISSN:
    1054-3139 ; 1095-9289
  • Format:
    pdf
  • Publisher:
    Oxford University Press (OUP)
  • Document Type:
    Journal Article
  • License:
  • Rights Information:
    CC0 Public Domain
  • Compliance:
    Submitted
  • Main Document Checksum:
    urn:sha-512:5aec50173954b60becd258b66d58c2354239dc115186ac0dd772e580a2a1fbc74188d3da2a10a3b1f241f7798f578d2c82a1ed203ea4258ea1224d5eb8eb3df4
  • Download URL:
  • File Type:
    Filetype[PDF - 2.18 MB ]
PREVIOUS
ON THIS PAGE
Details
The NOAA IR serves as an archival repository of NOAA-published products including scientific findings, journal articles, guidelines, recommendations, or other information authored or co-authored by NOAA or funded partners. As a repository, the NOAA IR retains documents in their original published format to ensure public access to scientific information.
BACK TO TOP