Evaluation of a deep neural network for acoustic classification using simulated echo sounder data

Abstract

An important part of fisheries acoustics is the classification of fish species. Sound waves are transmitted through water to detect fish species, and the echoes returning from the fish are categorized to be used for fish abundance estimates. These estimates are import for fishery management. Recently, it has been shown that a deep learning model performs well on the task of classifying acoustic data. However, these models are often criticized for being “black boxes” and hard to interpret. We have created a pipeline to test a neural network model, in order to shed light on what features of the data impact the predictions of the model. In this pipeline, simulated data is utilized, created by a model that emulates the performance of a multi-frequency echo sounder. The simulated data enables the possibility of adjusting one feature of the data at a time. We have concentrated on two features: the relative frequency response, an energetic characteristic of the data, and the shape of the fish schools. A neural network is trained to recognize two types of fish schools, dissimilar only in shape and relative frequency response. The network is then tested on data where either shape or relative frequency is changed, to evaluate the importance of each feature. From these tests we conclude that the relative frequency response affects the model's performance more than shape.