All Times EDT
Virtual
Classification of Animal Sounds in a Hyperdiverse Rainforest Using Convolutional Neural Networks (309921)
Zuzana Burivalova, University of Wisconsin-MadisonTatiana Midori Maeda, University of Wisconsin-Madison
Daniel Pimentel-Alarcon, University of Wisconsin-Madison
Claudia Solis-Lemus, University of Wisconsin-Madison
*Yuren Sun, University of Wisconsin-Madison
Keywords: Convolutional Neural Network, Animal Classification, Rainforest, Soundscapes, Data Augmentation, Conservation
Biodiversity is an integral and irreplaceable part of rainforest ecosystems. In order to protect tropical forest biodiversity, we need to be able to detect it reliably, cheaply, and at large scales. Automated species detection from passively recorded soundscapes is a promising technique to do so, but it is constrained by the necessity of large training data sets. We sampled soundscapes ranging from 1 day to 1 year from about 100 sites in the rainforests of Borneo and labeled around 3700 instances of about 440 ‘sonotypes’ (unique animal vocalizations). We investigated the minimum viable training data set size, and the extent to which data augmentation can overcome the issues of small training data sets when using a Convolutional Neural Network (CNN) with the spectrograms. We found that even relatively high sample sizes (>100 per call type) lead to mediocre accuracy, which however improves significantly with data augmentation, including at extremely small sample sizes. Our results suggest that data augmentation can make the use of CNNs as means to classify species' vocalization feasible for individual eco-acoustic projects with many rare species.