Answers are grounded in this paper and cite their sources.
Commercial fishing vessels equipped with calibrated echosounders continuously collect large volumes of acoustic data that could improve the estimates of fish distribution and abundance. However, the data requires manual scrutiny using echograms to identify fish schools before acoustic backscatter can be integrated into fish estimates and biomass indices, making large-scale routine analysis impractical. Manual annotations are time-consuming and prone to error, especially given the large amount of data involved. To solve this problem, we developed a deep-learning model based on U-Net convolutional neural network to automate the preprocessing and labelling of fish schools. The model was trained using manually annotated echograms from ten years of calibrated 38 kHz echosounder data collected during 30 blue whiting fishing trips by seven pelagic freezer trawlers from the pelagic freezer trawler fishing industry. Manual annotations of fish schools and seabed generated in Large Scale Survey System LSSS served as the ground truth for training and evaluation. The network trained for 50 epochs achieved a Dice coefficient of 0.69, demonstrating consistent segmentation performance between predicted and manually annotated segmentation masks across years and vessels. To assess the suitability of automated segmentation for fisheries applications, we compared manually derived Nautical Area Scattering Coefficient (NASC) values with NASC calculated from the acoustic backscatter contained within the automatically detected fish-school regions. Image-based segmentation performance was further evaluated using the area under the precision-recall curve (AUC), which measures the model’s ability to distinguish fish-school pixels from other classes across decision thresholds. While AUC ranged from 0.81 to 0.98 depending on the vessel, agreement between manually and automatically estimated NASC (R2 0.31 to 0.90) showed that image-based segmentation metrics alone do not adequately reflect final accuracy. Instead, using NASC directly within segmented fish schools, provides a performance measure directly relevant to stock assessment. Our results demonstrate that deep learning can reduce manual scrutiny of large industry echosounder datasets while producing biomass estimates that are consistent with expert analyses. More importantly, they show that acoustic backscatter-based metrics such as NASC provide a more appropriate evaluation of segmentation models than image-overlap alone when the objective is fish biomass estimation.
No takes yet. Share an insight, caveat, or question.
Possenti et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: