One size fits all? Adaptation of trained CNNs to new marine acoustic environments

Convolutional neural networks (CNNs) have the potential to enable a revolution in bioacoustics, allowing robust detection and classification of marine sound sources. As global Passive Acoustic Monitoring (PAM) datasets continue to expand it is critical we improve our confidence in the performance of models across different marine environments, if we are to exploit the full ecological value of information within the data. This work demonstrates the transferability of developed CNN models to new acoustic environments by using a pre-trained model developed for one location (West of Scotland, UK) and deploying it in a distinctly different soundscape (Gulf of Mexico, USA). In this work transfer learning is used to fine-tune an existing open-source 'small-scale' CNN, which detects odontocete tonal and broadband call types and vessel noise (operating between 0 and 48 kHz). The CNN is fine-tuned on training sets of differing sizes, from the unseen site, to understand the adaptability of a network to new marine acoustic environments. Fine-tuning with a small sample of site-specific data significantly improves the performance of the CNN in the new environment, across all classes. We demonstrate an improved performance in area-under-curve (AUC) score of 0.30, across four classes by fine-training with only 50 spectrograms per class, with a 5% improvement in accuracy between 50 frames and 500 frames. This work shows that only a small amount of site-specific data is needed to retrain a CNN, enabling researchers to harness the power of existing pre-trained models for their own datasets. The marine bioacoustic domain will benefit from a larger pool of global data for training large deep learning models, but we illustrate in this work that domain adaptation can be improved with limited site-specific exemplars.