Identification of fish species in estuaries and rivers using recorded soundscape with supervised machine learning

Abstract

Identification of various fish species in estuarine and riverine environments is an important ecological problem at the intersection of various disciplines, including underwater acoustics, signal processing, and fisheries. Advancements in this field can be made by applying machine learning and deep learning technologies. From the vast amount of existing work on machine learning techniques, of special relevance to this problem are those aiming at the detection and labeling of speakers on overlapping speech, where the existing literature is much more reduced. In this paper, we present a dataset collected with a passive recorder placed on a tripod in Delaware Bay estuary during summer 2013 to continuously measure environmental soundscapes for several weeks. By using a library of mid-Atlantic region fish sounds as prior knowledge, we adapt the aforementioned techniques to the detection and identification of various recorded sounds. This adaptation is challenging, as the statistical properties of fish sounds are not the same as those of speech, and requires careful research on how to perform data pre-processing, feature extraction, and model development. The application of the proposed machine learning techniques to future observations will enhance the potential for long-term autonomous ecological studies in bays and rivers.