Zualkernan, ImranArshad, Muhammad Arbab2021-10-042021-10-042021-0835.232-2021.40http://hdl.handle.net/11073/21556A Master of Science thesis in Ccomputer Engineering by Muhammad Arbab Arshad entitled, “Unsupervised Deep Learning for Classification Of Bats Calls Using Acoustic Data”, submitted in August 2021. Thesis advisor is Dr. Imran Zualkernan. Soft copy is available (Thesis, Completion Certificate, Approval Signatures, and AUS Archives Consent Form).Analysis and understanding of bat behaviors have taken on an increased importance post-Covid 19. Manual analysis of echolocation calls in bats to deduce behavior is cumbersome, time-consuming and costly. Previous attempts to automate this process have relied on labeled data which is expensive and difficult to collect. This thesis explored the use of state-of-the-art unsupervised learning algorithms like IMSAT, IIC, SCAN, JULE and DeepCluster to determine if interesting bat behaviors can be automatically determined based on unlabeled bat echolocation data which is readily available. The algorithms originally developed for image classification were adapted to work with audio data. One small labeled echolocation data set from the UAE Al-Hajar mountains and a large unlabeled dataset from an urban space in Dubai from the Emirates Nature - World Wildlife Foundation (WWF) were utilized. A coding scheme for interpreting bats' behavior was also developed. The results are that different algorithms capture different behavior. For example, IIC and IMSAT identified the presence of multiple bats, DeepCluster was better able to identify prey capture attempts, SCAN could distinguish bat calls in a close habitat and JULE could capture different species types. Based on Mutual Information (MI) the most similar pairs of algorithms were IIC and IMSAT (0.429), IIC and DeepCluster (0.374), and IMSAT and DeepCluster (0.266). On the small labeled data set, IIC performed the best with an accuracy of 48.28% followed by IMSAT (43.59%), JULE (43.13%), DeepCluster (39.84%) and SCAN (29.38%). A baseline K-Medoid algorithm only had an accuracy of 23.75%. For future work, better audio augmentation techniques can be explored and other unsupervised learning algorithms like DAC, DEC and K-Autoencoders can be investigated as well.en-USUnsupervised Deep LearningAudio ClassificationWildlife MonitoringThesis