Machine Learning for Conservation: Evaluating Deep Learning and Feature Extraction in Bird Species Classification in New Zealand

Abstract

Automated classification of bird sounds plays an im-portant role in monitoring and protecting biodiversity. Recently, similar efforts have been carried out for birds from all over the world, but New Zealand is one that has been overlooked. Hence in this study, we will be comparing feature extraction methods and machine learning models using a dataset that primarily contains bird species from New Zealand. Machine learning models such as Gated Recurrent Unit (GRU), Long Short-term Memory (LSTM) Recurrent Neural Network, Artificial Neural Network (ANN), and Convolution Neural Network (CNN) were used for audio classification. The accuracies achieved from the training of these models resulted in GRU-MFCC with 0.78, LSTM-MFCC with 0.91, ANN-MFCC with 0.091, and CNN-MFCC with 0.997 accuracy respectively. In order to design a user interface that can anticipate bird sounds and identify them appropriately, we employed our highest-performing model, CNN, with MFCC acting as the extractor.