Machine learning goes wild: Using data from captive individuals to infer wildlife behaviours

Area of Science:

• Behavioral Ecology
• Machine Learning
• Animal Tracking

Background:

• Behavior classification models using acceleration data are successful in captive animals.
• Transferring these models to wild animals is crucial for ecological studies but challenging without direct observation.
• Current methods often require direct observation of target animals in the wild.

Purpose of the Study:

• To infer wild animal behavior from acceleration data using models trained on captive individuals.
• To avoid the necessity of observing wild conspecifics during model development.
• To develop methods for validating the accuracy of extrapolated behavior classifications.

Main Methods:

• Trained Random Forest (RF) and Support Vector Machine (SVM) algorithms on captive red fox (Vulpes vulpes) data.
• Applied RF and SVM models to acceleration data from wild red foxes.
• Tested an Artificial Neural Network (ANN) with a moving window for behavior classification.
• Investigated four strategies to validate the classification output.

Main Results:

• RF and SVM models performed well in training but failed to classify behaviors when transferred to wild foxes.
• The ANN with a moving window successfully inferred distinct behaviors in wild foxes with consistent results.
• Kappa values under training conditions were high: RF (0.82), SVM (0.78), ANN (0.85).

Conclusions:

• The ANN with a moving window approach significantly improves upon existing methods for classifying wild animal behavior.
• This framework enables the inference of behaviors in wild animals without prior direct observation.
• The method is applicable to various species, advancing behavioral ecology research.