Deep Learning Approaches for Automatic Acoustic Detection of the Bachman's Sparrow and its Application to Assessing its Response to Prescribed Burns in Subtropical Habitats of Central Florida
Abstract
Automatic birdsong detection is a powerful tool that can enable cost-efficient monitoring of birds at large time and spatial scales. For the development of birdsong classifiers, the use of deep learning algorithms such as convolutional neural networks (CNNs) have shown considerable success. However, to successfully train a CNN-based birdsong classifier, a large dataset with labeled recordings of the target species is needed, a condition that represents a major limitation for researchers and conservationists. In this study, the influence of some machine learning techniques used to overcome data scarcity such as transfer learning, data augmentation, and Siamese Neural Networks were evaluated on the performance of Bachman's Sparrow (Peucaea aestivalis)detection. For this purpose, CNN-based models were trained under five different treatments in nine different training set sizes. In addition, to estimate the generalization ability of each model, the birdsong detection performance was evaluated with different signal-to-noise ratio for each of the treatments. My results reveal that the use of pre-trained networks allows obtaining good results with false positive rates close to zero even with low training set sizes. In addition, the indiscriminate use of data augmentation can decrease the detection capability of the model, particularly for recordings with more background noise. My results indicated that increasing the training set does not necessarily improve model performance, so it is recommended to incorporate techniques that measure the potential of a dataset beyond its size.
Degree
M.S.
Advisors
Pijanowski, Purdue University.
Subject Area
Acoustics|Wildlife Conservation|Remote sensing|Artificial intelligence|Conservation biology|Electrical engineering|Engineering|Environmental management
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.