Bayesian estimation of fish school cluster composition applied to a BeringSea acoustic survey

This paper applies BASCET, a Bayesian Spatial Composition Estimation Toot f or clusters of acoustically identified schools, to Bering Sea acoustic surv ey data collected during 1994. As the method employs prior information from an acoustic expert, procedures for eliciting such information arc suggeste d and pitfalls of the process are indicated. Techniques for model checking using the posterior predictive distribution are employed, as is a mufti-cha in method for evaluating the convergence of the Markov-Chain Monte Carlo al gorithm used in BASCET. Unlike methods based on neural networks, BASCET is able to provide confidence regions for its estimates of school cluster comp osition. In addition, it can indicate which school cluster attributes were most influential in determining a given estimate, a useful tool for model c hecking that is here demonstrated on a randomly selected cluster. Estimated abundance ratios of juvenile to adult pollock (Theragra chalcogramma) were compared, in two regions, to the values used by expert technicians. Ratios differed from expert values by less than 0.03 in both regions, The encoura ging results reported here suggest that the BASCET method, originally teste d on simulated data, may be usefully applied to real surveys.