A Deep Transfer Learning Model for the Identification of Bird Songs: A Case Study for Mauritius

Abstract

Birds communicate with their colonies through sound and inform them of potential problems like forest fires. The identification of bird sounds is therefore very important and has the potential to solve some global problems. Convolutional neural networks (CNNs) are sophisticated deep learning algorithms that have proven to be effective in image processing and in sound classification. This paper describes the work done to develop a tool using a deep learning model for classifying Mauritius bird sounds from audio recordings. A dataset obtained from the Xeno-canto bird song sharing site, which hosts a vast collection of labeled and classified recordings, is used to fine-tune three pre-trained CNN models, namely InceptionV3, MobileNetV2 and RestNet50 and a custom model. The neural network's input is represented by spectrograms created from downloaded mp3 files. Time shifting and pitch stretching have been used for data augmentation. The best performing model has been integrated into a website to identify birds sounds recordings. In this work, transfer learning has been used successfully to produce a model with a weighted accuracy of 84%. Although a custom CNN was trained, better accuracy was achieved through the use of transfer learning.