Semi-supervised target classification in multi-frequency echosounder data

Abstract Acoustic target classification in multi-frequency echosounder data is a major interest for the marine ecosystem and fishery management since it can potentially estimate abundance or biomass of species. A key problem current methods heavy dependence on manual categorization samples. As solution, we propose novel semi-supervised deep learning method leveraging few annotated samples together with vast amounts unannotated samples, all single model. Specifically, two inter-connected objectives, namely, clustering objective objective, optimize one shared convolutional neural network an alternating manner. The exploits underlying structure data, both unannotated; enforces certain consistency to given classes using We evaluate our from sandeel case study North Sea. In setting only tenth training annotated, achieves 67.6% accuracy, outperforming conventional by 7.0 percentage points. When applying proposed fully supervised setup, achieve 74.7% surpassing standard 4.7