Bat Detective

Bat Detective detects and classifies ultrasonic full-spectrum search-phase echolocation calls of echolocating bats to enable passive acoustic monitoring for biodiversity assessment.

Key Features:

• Convolutional Neural Network (CNN) Pipeline: Bat Detective employs a convolutional neural network-based pipeline to identify and localize ultrasonic full-spectrum search-phase echolocation calls in noisy audio recordings.
• Comprehensive Training Data: The deep learning models were trained on labeled full-spectrum ultrasonic audio collected across Europe with labels contributed via www.batdetective.org.
• Robust Detection Performance: Comparative analyses and testing show superior performance relative to existing algorithms and commercial systems for detecting search-phase echolocation calls, validated on five years of monitoring data from Jersey (UK).
• Minimal Species-Specific Assumptions: The approach uses minimal species-specific design decisions, allowing adaptation to other echolocating or vocalizing species given appropriate training data.
• Indicator-Species Monitoring Capability: Automatic detection and classification of bat calls support the use of bats as indicator species for assessing ecological health and anthropogenic impacts.

Scientific Applications:

• Population Trend Analysis: Quantifying temporal trends in bat activity and occurrence from passive acoustic recordings to inform conservation status assessments.
• Biodiversity Impact Studies: Measuring responses of echolocating bat communities to environmental change and anthropogenic pressures at large scales.
• Citizen Science Integration: Leveraging labels contributed by citizen scientists (via www.batdetective.org) to expand training datasets and monitoring coverage.

Methodology:

A convolutional neural network-based pipeline trained on labeled full-spectrum ultrasonic audio collected across Europe (labels from www.batdetective.org) and designed to localize search-phase echolocation calls in noisy recordings, with performance validated on five years of monitoring data from Jersey (UK).