Identifying animal behaviours from accelerometers: Improving predictive accuracy of machine learning by refining the...

Area of Science:

• Animal behavior analysis
• Machine learning in ecology
• Biologging technology

Background:

• Observing wild animals presents challenges, but animal-borne accelerometers offer insights into unobservable behaviors.
• Automated machine learning (ML) models streamline behavior identification from large datasets, but accuracy relies on training data quality.

Purpose of the Study:

• To investigate how data processing influences the predictive accuracy of random forest (RF) models for animal behavior identification.
• To leverage domestic cats (Felis catus) as a model organism for terrestrial mammalian behaviors.

Main Methods:

• Nine indoor domestic cats were fitted with accelerometers, and their behaviors were video-recorded for calibration.
• Eight datasets were created with variations in descriptive variables, acceleration data frequencies (40 Hz vs. 1 Hz), and behavior durations.
• RF models were trained on these datasets, validated on indoor cats, and then applied to free-ranging cats.

Main Results:

• RF models accurately predicted indoor cat behaviors (F-measure up to 0.96), with improvements from post-collection data processing.
• Additional variables, standardized durations, and higher recording frequencies enhanced model accuracy.
• Prediction accuracy varied by behavior; high frequencies were better for locomotion, while lower frequencies better identified slower behaviors like grooming in free-ranging cats.

Conclusions:

• RF modeling provides a robust method for identifying behaviors from accelerometer data.
• Field validation is crucial for confirming model accuracy in free-ranging animals.
• Optimized data processing methods can significantly enhance behavior identification accuracy, benefiting wildlife ecology, welfare, and management.