A comparison of convolutional neural networks and few-shot learning in classifying long-tailed distributed tropical bird songs

Abstract

Biodiversity monitoring depends on reliable species identification, but it can often be difficult due to detectability or survey constraints, especially for rare and endangered species. Advances in bioacoustic monitoring and AI-assisted classification are improving our ability to carry out long-term studies, of a large proportion of the fauna, even in challenging environments, such as remote tropical rainforests. AI classifiers need training data, and this can be a challenge when working with tropical animal communities, which are characterized by high species richness but only a few common species and a long tail of rare species. Here we compare species identification results using two approaches: convolutional neural networks (CNN) and Siamese Neural Networks (SNN), a few-shot learning approach. The goal is to develop methodology that accurately identifies both common and rare species. To do this we collected more than 600 hours of audio recordings from Barro Colorado Island (BCI), Panama and we manually annotated calls from 101 bird species to create the training data set. More than 40% of the species had less than 100 annotated calls and some species had less than 10. The results showed that Siamese Networks outperformed the more widely used convolutional neural networks (CNN), especially when the number of annotated calls is low.