Journal Paper

This article delves into the development of a two-stage machine learning system designed to detect and estimate lambing in sheep using wearable collar devices sensor data. The system is built by combining a binary classifier for early lambing detection and a multiclass classifier for accurate estimation of birth time. In three experiments, the performance of lightweight and non-lightweight machine learning classifiers has been assessed at the sample level and using k-fold cross-validation. The results emphasize the trade-offs between model complexity and deployment feasibility. Even though more complex models achieve higher accuracy, these require significantly more computational resources, while lightweight models offer faster, resource-efficient predictions, and easier deployment with potential precision. While addressing the resource constraints normally associated with wearable collars used in livestock monitoring, this study also underscores how sampling frequency and historic window size affect model accuracy to estimate with precision ewe lambing. Based on the experimental results, it is reasonable to deduce that future data collection can take place at 1 or 2 Hz while preserving dependable model performance. Finding the optimal trade-off combination between sample rate and time window size can greatly enhance prediction results and reduce computational resources for automated lambing monitoring. For the optimal trade-off between sampling frequency and historic window size, the combined lightweight two-stage system achieved an MCC of up to 0.93, effectively balancing accuracy and resource efficiency. Thus, this work can be a step forward and a foundation for future research into the design of practical machine learning solutions for animal welfare, livestock farming, and agricultural automation towards the design and deployment of scalable and efficient systems in real-world scenarios, by coupling model complexity with deployment ability.