Deep Transfer Learning-Based Automated Identification of Bird Song

Abstract

Bird species identification is becoming increasingly crucial for avian biodiversity conservation and assisting ornithologists in quantifying the presence of birds in a given area. Convolutional Neural Networks (CNNs) are advanced deep learning algorithms that have proven to perform well in speech classification. However, developing an accurate deep learning classifier requires a large amount of data. Such a large amount of data on endemic or endangered creatures is frequently difficult to gathered. Also, in some other fields, such as bioinformatics and robotics, the high cost of data collection and expensive annotation limit their progress, so large, well-annotated data creating a set is also difficult. A transfer learning method can alleviate overfitting concerns in a deep learning model. This feature serves as the inspiration for transfer learning, which was created to deal with situations where the data are distributed across a variety of functional domains. In this study, the ability of deep transfer models such as VGG16, VGG19 and InceptionV3 to effectively extract and discriminate speech signals from different species of birds with high prediction accuracy is explored. The obtained accuracies using VGG16, VGG19 and InceptionV3 were equal to 78, 61.9 and 85%, respectively, which are very promising.