Probabilistic School Classification of Multiple Species in Acoustic Echograms Based on Machine Learning  ^†

Abstract

Multifrequency trawl-acoustic surveys are used worldwide for continuous monitoring of pelagic ecosystems. Acoustic backscattering energy partitioning in different species is typically done by visual scrutiny of the echograms with the aid of trawl species composition, which may be subjective and time-consuming. Alternatively, machine learning techniques may provide well-stablished, objective, and reproducible methods for automatic school classification in acoustic echograms. The pelagic ecosystem is a diverse one, where many species co-occur in space and time, being mixed catches very common during scientific surveys. However, most of the school classification models are built using single species composition trawls due to difficulties to assign a class to each school in multispecific trawls. The present study has the aim of developing and comparing different probabilistic multivariate models to identify pelagic species in mixed scenarios based on trawl catch proportions. In addition to the standard predictors, a novel variable, collective mean TS per nautical mile measured on the periphery of the schools, has shown to play an important role in species discrimination. The methods were applied on data from seven consecutive years of an acoustic survey in the Bay of Biscay. Preliminary results yielded classification performances near 90% in classifying 5 different pelagic species.