Improving Livestock Monitoring: Multi-Classifier Approach for Cattle Behavior Recognition

Advancements in wearable sensors and artificial intelligence have enabled real-time monitoring of livestock behavior, transforming traditional farm management practices. This paper presents an enhanced machine learning framework for dairy cattle behavior classification, building upon the open-source MmCows dataset. The dataset comprises multimodal sensor data collected from 16 Holstein cows during a two-week period in the summer of 2023 at the Arlington Agricultural Research Station, University of Wisconsin-Madison. We focus on data captured from an inertial and magnetic measurement unit placed on the neck halters of the cattle to measure motion patterns. While other works utilized a single deep neural network classifier, our expanded approach explores multiple machine learning models, Support Vector Machine, Random Forest, and Extreme Gradient Boosting. The data preprocessing pipeline includes 10-second window segmentation with 50% overlap, discrete wavelet transforms for feature extraction, and z-score normalization. Our findings demonstrate that incorporating diverse classifiers improves behavior classification performance, mitigating overfitting, while maintaining high accuracy.