Using Machine Learning for Remote Behaviour Classification-Verifying Acceleration Data to Infer Feeding Events in...

Area of Science:

• Wildlife behavior analysis
• Conservation technology
• Machine learning applications

Background:

• Studying elusive wildlife behavior is crucial, especially for threatened species involved in human-wildlife conflicts.
• Accelerometers (ACCs) and machine learning algorithms (MLAs) offer remote behavioral monitoring solutions.
• Cheetahs face conservation challenges due to human-wildlife conflict.

Purpose of the Study:

• To evaluate the effectiveness of ACC data and MLAs in identifying cheetah behaviors.
• To develop and validate a predictive model for cheetah behavior using ground-truthed data.
• To apply the validated model to free-ranging cheetahs for ecological and conservation insights.

Main Methods:

• Utilized ACC data from five captive cheetahs, ground-truthed with direct observations.
• Applied six MLAs, including two ensemble methods and a probability threshold, to classify six distinct behaviors.
• Deployed the trained model on ACC data from four free-ranging cheetah males.

Main Results:

• The MLA models achieved high precision and recall for resting, walking, and trotting/running behaviors (80.1–100.0% precision, 87.3–99.2% recall).
• Feeding behavior identification showed 74.4–81.6% precision and 54.8–82.4% recall.
• The model successfully identified all nine known kill sites and 17 of 18 feeding events in free-ranging cheetahs.

Conclusions:

• The developed behavioral model reliably detects feeding events in free-ranging cheetahs using ACC data.
• This technology can accurately determine cheetah kill sites, crucial for conservation efforts.
• The findings support the use of ACCs and MLAs to mitigate human-cheetah conflicts and inform wildlife management strategies.