Bat detective-Deep learning tools for bat acoustic signal detection

Area of Science:

• Bioacoustics
• Artificial Intelligence
• Conservation Biology

Background:

• Passive acoustic sensing is crucial for assessing anthropogenic impacts on biodiversity, particularly for echolocating bats.
• Accurate, reliable, and open-source tools are needed for bat call detection and classification in large audio datasets.
• Existing tools often neglect call localization, especially in noisy environments, and many are commercial.

Purpose of the Study:

• To develop an open-source pipeline using deep learning for detecting bat echolocation calls.
• To improve the accuracy and efficiency of bat population monitoring.
• To provide a tool adaptable for detecting other species in audio recordings.

Main Methods:

• Developed a convolutional neural network (CNN) based pipeline for detecting ultrasonic, full-spectrum, search-phase bat calls.
• Trained deep learning algorithms on European road-transect audio data labeled by citizen scientists.
• Compared the pipeline's detection performance against existing algorithms and commercial systems.

Main Results:

• The developed pipeline demonstrated significantly higher detection performance for search-phase echolocation calls compared to existing methods.
• Applied the pipeline to five years of bat monitoring data from Jersey (UK), showing robust performance.
• The system achieved accurate and efficient monitoring, outperforming a widely-used commercial system.

Conclusions:

• The open-source pipeline offers a powerful, accurate, and efficient tool for automated bat population monitoring.
• The approach facilitates the use of bats as large-scale indicator species for environmental changes.
• The methodology is adaptable for detecting other species in audio, highlighting the potential of deep learning in bioacoustics.