부산대학교 도서관

Avian songs, a fundamental element of bird ecology, serves a crucial function in communication, mate selection, territorial dynamics, and many more. Precise identification of avian vocalizations is imperative for comprehending ecosystem health and facilitating efficient conservation endeavors. Conventional approaches encounter difficulties owing to the intricate nature of avian vocalizations, marked by a variety of song types, regional distinctions, and individual subtleties. Studies show that Machine Learning (ML) applications have the potential to address these issues accurately and conveniently. This study proposes a method to identify avians with their songs in real life environment using two different sizes of Convolutional Neural Network (CNN) models. The proposed methodology involves converting audio files into spectrograms, followed by training and validating on 40 different bird species. The conclusive prediction is derived from a dataset comprising 10-minute soundscapes, processed into 120 spectrograms with a 5-second interval each. According to the experimental findings, the integrated model demonstrates an accuracy of approximately 86% in forecasting the overall count of bird species within each soundscape.