A general procedure for estimating the composition of fish school clusters using standard acoustic survey data

Abstract

An algorithm to identify classes of fish in acoustic backscatter images would improve the accuracy of acoustic biomass estimates over manually scrutinized images. A generalized Bayesian procedure for such identification called BASCET is presented, and two implementation strategies for the procedure are compared using simulated acoustic survey data. The procedure has several unusual characteristics: it evaluates schools not individually but in clusters; it makes use of human experience at cluster identification; it presents measures of uncertainty in all estimation results; and it constructs the training set required for supervised learning automatically using spatial and temporal assumptions. The simulation study comparison suggests that making use of temporal and spatial structure in the acoustic data leads to improved estimation performance. On the simulated data, the BASCET algorithm correctly identified the dominant fish class in 15 of 16 cases. However, the simulation model generates acoustic survey data based on the same assumptions used in BASCET, assumptions that may differ from a real acoustic survey. The study also assumed that the human experience incorporated in the Bayesian prior distributions was not misleading. Performance of BASCET on real acoustic data is presented in a companion paper.