Behavioural compass: animal behaviour recognition using magnetometers.

Area of Science:

• Animal behavior analysis
• Bio-logging technology
• Machine learning in ecology

Background:

• Animal-borne data loggers collect high-frequency multi-sensor data for fine-scale behavioral insights.
• Magnetometers offer unique behavioral detection capabilities beyond accelerometers, crucial for limited training data scenarios.
• A unified framework is needed to integrate and compare magnetometer and accelerometer data for robust behavior recognition.

Purpose of the Study:

• To develop an end-to-end approach for recognizing common animal behaviors using magnetometer data.
• To create a classification framework accommodating and comparing data from both magnetometers and accelerometers.
• To assess the robustness and accuracy of magnetometer-based behavior recognition compared to accelerometer-based methods.

Main Methods:

• Developed biomechanical descriptors from magnetometer data, utilizing static (tilt/posture) and dynamic (movement intensity/periodicity) components.
• Employed a hybrid scheme combining biomechanics and machine learning for behavior recognition.
• Validated the method on triaxial magnetometer data from ten wild meerkats, with video recordings as groundtruth, and compared results with accelerometer data.

Main Results:

• Magnetometer data achieved over 94% recognition accuracy, comparable to accelerometer data.
• Magnetometers demonstrated higher robustness to inter-individual variability in dynamic behaviors.
• Accelerometers showed superiority in estimating animal posture.

Conclusions:

• Magnetometers accurately identify common animal behaviors, particularly excelling in dynamic behavior recognition.
• The biomechanically informed hybrid scheme accommodates data from magnetometers, accelerometers, or both, leveraging individual sensor strengths.
• This study provides a method to evaluate the added benefits of magnetometers in behavior recognition frameworks for future research.