Learning relationships between bat habitats and biosonar behavior

Abstract

Echolocating bats are known to adapt their biosonar behaviors according to their surroundings. Vespertilionid bats, for example, adapt their pulse duration and bandwidth as a function of their proximity to foliage. However, these relationships have been based on mostly qualitative observations so far. More powerful methods that are capable of handling large quantitative datasets could produce further and more detailed insights. In the current work, aligned multimodal datasets have been gathered to address this challenge: habitat geometry (laser scans), acoustic scene (biomimetic active sonar system), bat flight trajectories (infrared camera array), and biosonar behavior (microphone array). Machine learning approaches can then be used to analyze relationships between the components of these multimodal datasets. To do this, representations of the bat’s surroundings that are suitable for biosonar problems have been developed, for example a probability density function of the angles and distances from the laser scanned points of a habitat model relative to the bat’s position and orientation. These representations will serve as inputs to machine learning algorithms that can predict the quantitative features of the bat’s biosonar behavior. This research could lead to a better understanding of bat pulse adaptation and also aid in developing biomimetic sonar navigation systems.Echolocating bats are known to adapt their biosonar behaviors according to their surroundings. Vespertilionid bats, for example, adapt their pulse duration and bandwidth as a function of their proximity to foliage. However, these relationships have been based on mostly qualitative observations so far. More powerful methods that are capable of handling large quantitative datasets could produce further and more detailed insights. In the current work, aligned multimodal datasets have been gathered to address this challenge: habitat geometry (laser scans), acoustic scene (biomimetic active sonar system), bat flight trajectories (infrared camera array), and biosonar behavior (microphone array). Machine learning approaches can then be used to analyze relationships between the components of these multimodal datasets. To do this, representations of the bat’s surroundings that are suitable for biosonar problems have been developed, for example a probability density function of the angles and distances from the las...