Scalable, probabilistic echo inversion for multifrequency and broadband acoustic surveys

Abstract

Identifying biological scatterers is a perennial challenge in fisheries acoustics. Most practitioners classify backscatter based on direct sampling and frequency-difference thresholds, then integrate at a single frequency. However, this approach struggles with species mixtures, and discards multi-frequency information when integrating. Inversion methods do not have these limitations, but are seldom used, because their species identifications are often ambiguous and their algorithms complicated to implement. We address these shortcomings with a probabilistic, Bayesian inversion method. Like other inversion methods, it handles species mixtures, uses all available frequencies, and extends naturally to broadband signals. Unlike prior approaches, it leverages Bayesian priors to rigorously incorporate information from direct sampling and biological knowledge, constraining the inversion and reducing ambiguity in species identification. Because it is probabilistic, it can be trusted to run automatically: it should not produce solutions that are both wrong and confident. Unlike some data-driven machine learning models, it is based on acoustical scattering processes, so its inferences are physically interpretable. Finally, the approach is straightforward to implement using existing Bayesian libraries, and is easily parallelized for large datasets. We present examples using simulations and field data from the Gulf of Alaska, and discuss possible extensions and applications of the method.