Weakly supervised classification of acoustic echo-traces in a multispecific pelagic environment

In trawl-acoustic methods, machine learning can objectively assign species composition to echo-traces, providing a reproducible approach for improving biomass assessments and the study of schooling behaviour. However, the automatic classification of schools in multispecies environments is challenging due to the difficulty of obtaining ground truth information for training. We propose a weakly supervised approach to classify schools into seven classes using catch proportions as probabilities. A balancing strategy was used to address high dominance of some species while preserving species mixtures. As the composition of schools from multispecific catches was unknown, model performance was evaluated at the school and haul level. Accuracy was 63.5% for schools from single-species catches or those identified by experts, and a 20.1% error was observed when comparing predicted and actual species proportions at the haul level. Positional and energetic descriptors were highly relevant, while morphological characteristics showed low discriminative power. The highest accuracies were obtained for juvenile anchovy and Muller’s pearslide, while sardine was the most challenging to classify. Our multioutput approach allowed the introduction of a metric to assess the confidence of the model in classifying each school. As a result, we introduced a method to classify echo-traces considering prediction reliability.