Audio Classification of Bird Species Using Convolutional Neural Networks

Abstract

As the total number of birds has declined in the billions over the last 50 years, an accurate method for classifying bird species is necessary for conservation efforts and population monitoring. One promising method is using machine learning models to classify birds by their sounds, which has emerged due to benefits such as being less affected by environmental factors (eg. habitat, time of day), and lower disturbances to bird species during the data collection process, contrary to other processes such as image classification. As audio processing may eventually become the main method of classifying birds and may be used as an important conservation tool, it is imperative to understand the challenges that must be overcome before it can be successfully applied. In this work, the programming language Python and the machine learning model Convolutional Neural Networks were used to process and classify audio recordings from over 150 different bird species. This study demonstrates that although audio classification is a promising method of classification, many challenges are still present in the field, such as the amount of variety in the different calls of a single bird, the presence of background noises in many audio recordings, and the difficulty in efficiently representing an audio signal with images, highlighting the importance of overcoming these challenges for conservation efforts.