Goats on the Move: Evaluating Machine Learning Models for Goat Activity Analysis Using Accelerometer Data

Area of Science:

• Animal behavior analysis
• Machine learning applications
• Wearable sensor technology

Background:

• Automated animal activity recognition using body-worn sensors offers insights into animal welfare.
• Previous algorithms struggled with complex accelerometer data, but deep learning shows promise.
• A need exists to compare deep learning models and input types for robust activity recognition.

Purpose of the Study:

• To evaluate the generalizing capabilities of different deep learning models for animal activity recognition.
• To compare orientation-independent data transformation techniques for accelerometer data.
• To identify optimal model-input combinations for accurate animal behavior classification.

Main Methods:

• Experimented with two orientation-independent data transformations: vector magnitude (L2-norm) and Discrete Fourier Transform.
• Trained three deep learning models: Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and a hybrid CNN (ensemble of MLP and CNN).
• Assessed model generalization using mixed cross-validation and goat-wise leave-one-out cross-validation.

Main Results:

• Orientation-independent data transformations yielded promising results for animal activity recognition.
• The hybrid CNN model, using L2-norm input, achieved high classification accuracy and low standard deviation.
• Misclassifications were concentrated in behaviors with similar accelerometer patterns or in minority classes.

Conclusions:

• Hybrid CNNs combined with orientation-independent accelerometer data processing offer superior generalization for animal activity recognition.
• Future improvements can be made by using larger, more balanced datasets to address misclassifications of similar or minority behaviors.