Résumé
The annual pelagic ecosystem monitoring survey PELGAS conducted by IFREMER in the Bay of Biscay aim at assessing the biomass of small pelagic fish populations exploited by commercial fisheries. To achieve this goal, expert operators combine various types of information to annotate the singlebeam echosounder (SBES) water column data. Information used include multi-frequency echograms (Kongsberg EK80), trawl results, and expert knowledge of species' aggregative behaviour in context of their biotic and abiotic environment. In this study, echointegrated SBES data and expert annotations are used to train a convolutional neural network (CNN) for semantic segmentation. The aim is to distinguish echogram pixels belonging to fish schools to those belonging to other biological scatterers (zooplankton and micronekton). The annotation scheme is not adapted for deep learning training purposes, as noise and uncertainties in labels hinder the learning and generalization abilities of the model. In this study, we propose to use a weakly supervised framework instead of a supervised one, to manage uncertainty in labels. We present a methodology for generating pseudo-labels at the resolution of acoustic images provided to the network. Furthermore, a representation of the pseudo-labels confidence is used during the learning process to account for uncertainty introduced during the refining process. The results indicate that the trained model has good precision in terms of pixel-wise metrics. Masks predicted by the CNN are used to compute the nautical areal scattering coefficient (NASC) of fish and sound scattering layer class. The model accurately predicted fish NASC, aligning with experts' estimations.