A Machine Learning Approach for Automatic Identification of Tawny Owl Individuals From Acoustic Features

In this thesis, supervised machine learning was used to discover if supervised machine learning techniques can be used to recognize individual Tawny Owls ( Strix aluco ) for use in ecological research and conservation. Therefore, engineered acoustic features of a range of acoustic features were extracted from a dataset of annotated owl vocalizations: MFCCs, pitch statistics, formants, spectral descriptors, harmonic to noise ratio, etc., and standardized. Furthermore, using spectrograms along with DTW and standardized Euclidean distance, misclassifications were investigated from an Extra Trees Classifier trained with high accuracy on individuals. The acoustic plausibility of errors were revealed through these analyses with the model limitations. Moreover, the system demonstrated resistant performance through invariance tests for various signal modifications including pitch variations and amplitude changes and noise disturbances. The process of distinguishing individuals worked well in most cases but specific vocal similarities caused classification difficulties. The presented work paves the path to pursue upcoming work by combining self-supervised learning with sequence-aware models to enhance classification capabilities and attribute recognition in feature analysis. This work presents an understandable framework to detectindividual animals through their vocal characteristics.