Automated acoustic identification of echolocating bats by machine learning : application for biodiversity monitoring for the Australia cotton industry

• Main Author: Chan, Jaslyn Jia Hui
• Other Authors: -
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2018
• Subjects: DRNTU::Science::Biological sciences::Ecology
• Online Access: http://hdl.handle.net/10356/76124
• Institution: Nanyang Technological University

Global biodiversity and ecosystem services face great threats from the intensification of agriculture as global population and affluence increases. However, when practiced sustainably, agriculture can support the growth of biodiversity while increasing agricultural yield and economic profit. Cotton is one important agricultural crop which has seen major intensification over the past three decades in Australia. Many farmers are committed to meeting global environmental standards but they lack an effective method to survey and monitor ecosystem health, which could be a proxy for their compliance with environmental standards. // The acoustic monitoring of soniferous animals has gained popularity as an effective approach to monitoring and studying biodiversity in a wide array of habitats, providing insight into environmental and ecological conditions. In Australia, bats may allow us to conduct such acoustic surveys for the cotton fields. Bats possess ideal bioindicator qualities and echolocate to orientate themselves and locate objects. Their echolocation calls contain temporal and spectral information which can be extracted using various analysis techniques and software packages. Echolocation call structures differ amongst most species but may converge in certain situations or have great flexibility within a species. This results in the differentiation of certain species being more complicated than others. // This project investigates the feasibility of acoustic identification of Australian bat species, and hence their viability for use for acoustic monitoring. To do so, the project aimed to created an automated identifier for echolocating bats found in the cotton regions of Australia through the use of artificial neural networks. A library of reference echolocation calls for bats was created and collated containing 23 species, 159 individuals and 674 calls. A range of parameters was extracted using three different sound analysis programmes (AnalookW, Raven and SonoBat) to compare their accuracy in representing the calls. The parameters extracted from the calls were then statistically compared for their species discriminatory power, and machine learning techniques (i.e. back propagation artificial neural networks) were used to create and evaluate the accuracy of an automated acoustic identifier. // The results show that in general, an increase in number of parameters used to differentiate the species increased both the overall correct classification rate and the correct classification rate of the most confused species. There is also no clear difference in the overall correct classification rate of each software package’s range of parameters in comparison to one another. The best network used all 59 parameters from the software package Sonobat and had an overall correct classification rate of 99.7%. It only misclassified a single pulse from Scotorepens balstoni as Scotorepens sp. // The research work presented in this paper represents a significant step forward in creating automated acoustic identification systems for bats in Australian cotton. They reveal the effectiveness of artificial neural networks for Australian bat species, which had not been attempted prior despite its widespread success and use on bat species in other regions. However, the trained networks from this project are in their infancy, lacking a sufficient sample size and complete library of Australian bat species known to occur over cotton-growing regions. There is still a considerable amount of work required to acquire a complete reference library that would include the missing species as well as echolocation calls that could provide for intraspecific and geographical variation. While the results may be particular to the species in this study, it is nevertheless an optimistic prediction of the success for the future work in creating an automated acoustic identifier for echolocating bats to be used for biodiversity monitoring in the Australian cotton industry.