High accuracy at low frequency: detailed behavioural classification from accelerometer data.

Area of Science:

• Animal behavior and physiology
• Biologging and sensor technology
• Ecological monitoring

Background:

• Direct observation of animal behavior and physiology is often unfeasible.
• Interpreting multivariate acceleration data from sensors presents a significant challenge.
• Accelerometers offer a non-invasive method for collecting behavioral data in free-ranging animals.

Purpose of the Study:

• To develop and validate a method for classifying animal behaviors using low-frequency tri-axial accelerometer data.
• To compare the performance of different machine learning models for behavior classification.
• To assess the relationship between overall dynamic body acceleration (ODBA) and activity levels.

Main Methods:

• Collected tri-axial accelerometer data at a low sampling frequency (1 Hz) from dingoes (Canis dingo).
• Employed out-of-sample validation to compare random forest, k-nearest neighbour, support vector machine, and naïve Bayes models.
• Investigated the impact of predictor variable selection and moving window size on classification accuracy.

Main Results:

• Random forest models achieved the highest out-of-sample classification accuracy (mean 87% for 14 behaviors).
• Overall dynamic body acceleration (ODBA) was significantly higher for high-activity behaviors.
• A positive correlation was observed between ODBA and movement intensity.

Conclusions:

• A relatively simple random forest model can effectively classify animal behaviors from accelerometer data, overcoming key interpretation challenges.
• The developed approach is broadly applicable to terrestrial quadrupeds of similar size.
• The use of low sampling frequencies allows for extended deployment of accelerometers, capturing long-term behavioral and physiological data across life stages.