Foundation Pre-Trained Model for Underwater Soundscape Ecology

Abstract

Underwater soundscape ecology delves into understanding the unique, dynamic, and rich information within marine habitats, including signals from the environment, species, and structural complexities. Previous approaches have explored automated analysis to reduce reliance on labor-intensive labeling, leveraging techniques such as ecoacoustic indices, statistical-based and machine learning methods, with mixed results due to overfitting or out-of-distribution issues. Recently, passive acoustic monitoring (PAM) has become more popular with the capability of extensive underwater data collection, enabling long-term and continuous investigation of large geographic areas. To unveil opportunities for advancing underwater research with big data, this work introduces an innovative pretraining approach, employing state-of-the-art artificial intelligence techniques on large PAM datasets. Our model, named Underwater-Song, is designed to provide transferable feature embeddings for multiple underwater acoustic downstream tasks. This approach demonstrates usefulness and effectiveness for diverse goals, including fish and mammal monitoring, coral reef location classification, and assessing coral reef environmental state, with up to 11.06% performance improvement over previous methods, overcoming challenges associated with labor-intensive labeling and enhancing scalability and reliability in marine ecosystem research.

Keywords: artificial intelligence, audio pre-training, foundation pretrained model, signal processing, underwater soundscape ecology