Studying marine soundscapes by detecting known sound events and quantifying their spatio-temporal patterns can provide ecologically relevant information. However, the exploration of underwater sound data to find and identify possible sound events of interest can be highly time-intensive for human analysts. To speed up this process, we propose a novel methodology that first detects all the potentially relevant acoustic events and then clusters them in an unsupervised way prior to manual revision. We demonstrate its applicability on a short deployment. To detect acoustic events, a deep learning object detection algorithm from computer vision (YOLOv8) is re-trained to detect any (short) acoustic event. This is done by converting the audio to spectrograms using sliding windows longer than the expected sound events of interest. The model detects any event present on that window and provides their time and frequency limits. With this approach, multiple events happening simultaneously can be detected. To further explore the possibilities to limit the human input needed to create the annotations to train the model, we propose an active learning approach to select the most informative audio files in an iterative manner for subsequent manual annotation. The obtained detection models are trained and tested on a dataset from the Belgian Part of the North Sea, and then further evaluated for robustness on a freshwater dataset from major European rivers. The proposed active learning approach outperforms the random selection of files, both in the marine and the freshwater datasets. Once the events are detected, they are converted to an embedded feature space using the BioLingual model, which is trained to classify different (biological) sounds. The obtained representations are then clustered in an unsupervised way, obtaining different sound classes. These classes are then manually revised. This method can be applied to unseen data as a tool to help bioacousticians identify recurrent sounds and save time when studying their spatio-temporal patterns. This reduces the time researchers need to go through long acoustic recordings and allows to conduct a more targeted analysis. It also provides a framework to monitor soundscapes regardless of whether the sound sources are known or not. |
