AI-Enabled Sonar Systems for Fish Counting and Biomass Estimation: A Systematic Review

Abstract

Accurate fish counting and biomass estimation are central to precision aquaculture because they support feeding control, harvest planning, welfare monitoring, and stock management. Optical monitoring is widely used in digital aquaculture, but its reliability can decrease in turbid, high-density, or poorly illuminated environments. Sonar-based sensing provides a non-invasive alternative because acoustic signals can operate under low-visibility conditions. This paper presents a systematic review of artificial intelligence methods applied to sonar, echosounder, forward-looking sonar, multibeam sonar, side-scan sonar, and acoustic-camera data for fish counting, detection, behavior inference, and biomass estimation. Following PRISMA 2020 reporting guidance, studies published from 2016 to 2026 were screened from Google Scholar, Scopus, Lens.org, and IEEE Xplore, yielding 13 eligible studies for qualitative synthesis. The reviewed studies used convolutional neural networks, Mask R-CNN, U-Net and DeepLab-style semantic segmentation, density regression, transformer-based domain adaptation, Gaussian process regression, random forests, support vector machines, and clustering. The synthesis shows that model choice is strongly shaped by sonar modality: image-like acoustic-camera and forward-looking-sonar data favor convolutional detection, segmentation, or density-estimation models, whereas echograms and low-dimensional acoustic features favor semantic segmentation or classical machine-learning approaches. Reported outcomes suggest promising feasibility, but the evidence base remains limited by small datasets, single-site validation, inconsistent reporting, and limited deployment in commercial aquaculture. Future work should prioritize open benchmark datasets, prospective multi-site validation, standardized reporting, uncertainty estimation, and integration with farm decision-support systems.