Data from: Automatic Detection of Fish Schools in Acoustic Data.

Le résumé fourni par la source

Morphometric, energetic, and depth data on fish schools are important features for fisheries management and surveys of the biomass of prey available to marine predators. To address this, we developed two methods to automatically detect fish schools from echograms. They were collected from three uncrewed surface vehicles (USV) equipped with a 200kHz modulated frequency EK80 echosounder that were deployed across three different ecosystems: in Australia, Canada, and Sweden. The first method was based on double thresholding to reconstruct schools through dilation, both thresholds being automatically optimised for each image. The second method is based on deep learning, using Meta's Segment Anything Model (SAM). Their detection performance was compared against analyses conducted by five experts (100 images per site). Both methods showed satisfactory performance, with accuracies (81% for double thresholding, 76% for SAM) similar to the average expert accuracy of 82% ± 3%. In Sweden, there were more schools, in Australia the mean Sv (volume backscattering strength) was higher (indicating denser schools) while in Canada the schools were the biggest and the deepest. Even if schools detected had different characteristics depending on the ecosystems the methods accuracy varied only slightly between sites showing their robustness across a range of ecosystems.Both methods are available https://github.com/liliaguillet/echoedgeSchool features extracted from the echograms of the 3 sites and their analysis is available in the folder. More details in the readme file.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.