Phenomics

Automated detection of lambing phases in ewes using sensor-based phenotypes and behavioral modeling

---

Abstract

Lambing success has a direct impact on flock profitability, animal welfare, and long-term sustainability in sheep production. Thus, there is interest in developing lambing detection and phenotyping tools. Manual lambing phenotyping, conducted through direct human observations, is labor-intensive and prone to judgment errors. Moreover, due to the variation in gestation length, it is challenging for producers to accurately predict when parturition will occur. Thus, automated, sensor-based phenotyping offers a continuous and objective alternative for monitoring ewe behavior, eliminating the need for constant supervision. The objective of this study was to evaluate whether continuous, sensor-based ewe monitoring could detect lambing phases. Real-time detection of lambing phases can aid in identifying abnormal labor, enabling timely interventions to improve lamb survival, while providing objective phenotypes for genetic improvement programs. To accomplish this, a custom collar-mounted multimodal sensor was developed. The sensor board was equipped with a triaxial accelerometer and a gyroscope to continuously monitor locomotion and activity. Nine purebred Hampshire ewes were monitored for 14 days during the lambing season. Five overhead cameras recorded ewes over three pens to establish ground-truth activity before and during lambing. Video data was continually watched by one observer using a predetermined ethogram that defined lambing phases: baseline (no lambing activity), preparation (associated with nesting behavior), labor (onset of contractions), delivery, and post-partum lamb tending. Across ewes, an average of 6.8 preparation bouts and one continuous labor and delivery event per lambing were annotated. Phase durations and frequencies were summarized with principal component analysis (PCA) to identify dominant sources of variation. Sensor data was orientation-corrected and analyzed with within-ewe ANOVA to determine phase-specific differences across eight sensor-derived features. Random-forest models and a large-language-model (LLM) "sentiment-style" framework with leave-one-animal-out (LOAO) validation were used to assess the classification of sensor data into lambing phases. Class imbalance was addressed by oversampling less represented classes and adding Gaussian noise and by including class weighting in the loss functions. PCA analyses revealed distinct behavioral profiles between ewes, with labor duration contributing most to overall variance. Within-ewe analyses revealed that gyroscope magnitude and vertical acceleration features showed the strongest significance differences between phases. Predictive models showed inter-ewe variability, with baseline predicted most accurately (F1≈0.87-0.90), followed by delivery (~0.72-0.77), preparation (~0.19-0.31), and labor (~0.17-0.28). Wearable sensors effectively distinguished baseline and delivery behaviors, though labor remains difficult to distinguish from preparation. This shows potential for on-farm monitoring to support timely intervention. These findings suggest ewe-specific movement signatures may require individualized or deterministic modeling approaches. Additionally, sensor-derived measures, such as labor duration and pre-labor activity, are repeatable phenotypes that can be accurately recorded. Incorporating behavioral phenotypes into evaluations may help improve maternal ability and lambing ease.

Keywords: 2026

How to Cite:

Sliger, C., Park, Y., Pandey, S., Johnson, A., Burgett, R. & Steibel, J., (2026) “Automated detection of lambing phases in ewes using sensor-based phenotypes and behavioral modeling”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2284997. doi: https://doi.org/10.31274/wcgalp.23623

Rights: 1