Journal Paper

The quality and safety of sheep farming increasingly depend on automated monitoring systems, with parturition detection representing a critical challenge due to its direct impact on lamb survival. This study presents the development and evaluation of two machine learning approaches for predicting sheep parturition: a lightweight model designed for deployment on collar-mounted devices with limited computational resources, and a full-featured model intended for scenarios without such constraints. The models were developed using data from 53 parturition events collected via collar-mounted accelerometers and thermometers, within a system that supports real-time data acquisition, processing, and visualization of parturition predictions. The lightweight model, operating under strict computational and memory limitations, achieved an accuracy of 0.74 and a Matthews Correlation Coefficient (MCC) of 0.71. In contrast, the full-featured model delivered superior performance, reaching an accuracy of 0.81 and an MCC of 0.79 when predicting time to birth up to 11 h in advance. Additionally, temporal filtering optimization contributed to stable performance in extended validation scenarios for both models. The work included the development of a detection tool based on the use of an MQTT broker, that includes two subscribers that perform detection, and alarm triggering, and a producer that streams monitoring data gathered by the the collars. Overall, this work advances precision livestock farming by offering two practical solutions for automated parturition monitoring, enabling farmers to select either a resource-efficient or a high-performance approach according to their specific operational requirements and available infrastructure.