Enhancing minority class recognition in cattle monitoring: A robustness analysis of lightweight decision-level fusion

Abstract

Accelerometer-based monitoring has become an important approach in precision livestock farming, but achieving both high sensitivity for rare, health-relevant behaviors and computational efficiency remains challenging. Deep learning methods can address class imbalance but are often unsuitable for edge deployment due to their computational cost. This study evaluates the robustness of a lightweight decision-level fusion framework for imbalanced cattle behavior classification using tri-axial accelerometer data. To ensure rigorous evaluation, the Synthetic Minority Over-sampling Technique (SMOTE) was applied only in the training feature space to prevent data leakage. Because the dataset is strongly imbalanced, with Salt Licking (SLT) as the minority yet health-relevant class, model performance was assessed using Macro-F1 for global robustness and RecallSLT_{SLT}SLT for rare-event sensitivity. Individual tree-based models achieved strong results, with the best single model, XGBoost, obtaining 93.70% accuracy and 0.9421 Macro-F1. The proposed soft-voting fusion of Extra Trees, XGBoost, and CatBoost further improved performance to 94.21% accuracy and 0.9447 Macro-F1, with a statistically significant gain over the best single model (Wilcoxon signed-rank test, (p=0.0326). The framework also maintained strong minority-class recognition, with SLT achieving precision = 0.9571, recall = 0.9853, and PR-AUC = 0.9990. These results show that lightweight decision-level fusion can improve robustness and rare-event sensitivity without temporal deep learning, making it suitable for resource-constrained edge monitoring in livestock systems.