• End-to-end estrus detection from ear-tag accelerometry for on-device use • Compact deep-learning behavior model fits Cortex-M4F MCU for on-tag inference • Causal 14-day baselines normalize behavior deviations per animal to reduce drift • Regularized logistic regression accurately detects estrus from behavior deviations • Suites digital-ag decision support: automated alerts and workflow integration Timely and accurate estrus detection is essential for improving cattle reproductive efficiency and herd management. We present a lightweight deployable pipeline that infers estrus from daily behavior patterns estimated from ear-tag triaxial accelerometer data. As the key enabler, we propose a compact deep-learning model that classifies core cattle behaviors from raw accelerometer readings and supports construction of per-animal daily behavior profiles. To accommodate individual differences and temporal variability, we compute individualized baseline profiles using causal dynamic trimmed means over preceding days and define features as deviations of daily behavior durations from these baselines. We use a regularized logistic regression model to map the behavior-deviation features to the daily probability of estrus. Evaluated on data from 33 Angus cows fitted with smart ear-tags during a four-month grazing trial in temperate Australia, using five-fold stratified non-shuffled cross-validation, the proposed approach achieves mean precision 0.967, recall 0.917, F1 score 0.941, specificity 0.998, and MCC 0.938. The results show that subtle yet systematic behavioral shifts measured by low-cost accelerometers are reliable predictors of estrus, enabling scalable, infrastructure-light, and autonomous reproductive monitoring in extensive grazing systems. The pipeline is compatible with on-tag or edge deployment and integrates readily with digital-agriculture workflows for alerting and decision support.
Arablouei et al. (Tue,) studied this question.