Pre-processing spectrogram parameters improve the accuracy of bioacoustic classification using convolutional neural networks. Issue 3 (3rd May 2020)

• Record Type: Journal Article
• Authors: Knight, Elly C.; Poo Hernandez, Sergio; Bayne, Erin M.; Bulitko, Vadim; Tucker, Benjamin V.
• Is Part Of: Bioacoustics. Volume 29:Issue 3(2020)
• Page Start: 337
• Page End: 355
• Publication Date: 2020-05-03
• Subjects: Autonomous recording unit -- birdsong -- classification -- signal processing -- machine learning -- spectrogram; Bioacoustics -- Periodicals; Sound production by animals -- Periodicals; Animal sounds -- Periodicals; Sound recordings -- Periodicals; 591.59405
• DOI: 10.1080/09524622.2019.1606734
• Languages: English
• ISSNs: 0952-4622

Abstract

ABSTRACT: A variety of automated classification approaches have been developed to extract species detection information from large bioacoustic datasets. Convolutional neural networks (CNNs) are an image classification technique that can be operated on the spectrogram of an audio recording. Using CNNs for bioacoustic classification negates the need for sophisticated feature extraction techniques; however, CNNs may be sensitive to the parameters used to create spectrograms. We used AlexNet to classify spectrograms of audio clips from 19 species of birdsong. We trained and tested AlexNet with the spectrograms and observed that mean classification accuracy ranged from 88.9% to 96.9% depending on the parameters used to create the spectrogram. Classification accuracy was highest when we used a composite of four spectrograms with different combinations of scales for frequency and amplitude. Classification accuracy also varied depending on the FFT window size of the spectrogram. Overall, our results suggest that optimal spectrogram parameters for CNN classification may differ from those used for human visualization or other classification approaches. We suggest that if spectrogram parameters are appropriately selected, classification accuracy similar to current state-of-the-art methods can be achieved using off-the-shelf software and without the need to extract domain-specific features.