Potential passive acoustic tool for reef fish monitoring: A learning-based method

Abstract

Visual non destructive methods are widely used for assessing marine biodiversity, but a logistically simpler approach is combining passive acoustic techniques with artificial intelligence (AI). Here, we classified sounds of fish on subtropical rocky reefs in Arraial do Cabo (22°57'S, 41°01'W), RJ, Brazil, using supervised machine learning algorithms. Sound detection and feature extraction were performed manually by inspecting spectrograms using Raven Pro 1.6 software. Five supervised algorithms were implemented for classification: naive Bayes, support vector machine, random forest, decision tree, and multilayer perceptron. A representative subset of samples for each sound class was used to train the supervised algorithms. The accuracy of the algorithms was between 67% and 97%. Four classes of sounds were recognized, consisting of sequences of pulses. In general, temporal features were more important, however high frequency was the most important feature of all. Understanding the contribution of each feature is crucial for sound classification, however it is at an early stage for fishes. The classifier showed promising results, highlighting the effectiveness of applying AI to passive acoustics as a tool for monitoring of fish assemblages.