417 A data-driven approach for near real-time calving detection in extensive cattle systems.

Abstract Timely detection of calving events in extensive cattle systems is critical to improving reproductive efficiency, animal welfare and overall farm management. Traditional monitoring methods are labor intensive and ineffective in large-scale extensive systems. This study presents a near real-time, data-driven approach to automated calving detection using GNSS tracking data and anomaly detection techniques. A structured calving detection workflow was developed consisting of seven steps: (i) Data acquisition: GNSS positions were transmitted from collars to servers via SigFox, GSM, or LoRa networks; (ii) data preprocessing: including removing positions outside farm boundaries (>200 m), eliminating days with insufficient data, and resampling GNSS coordinates at 30-minute intervals; (iii) indicator calculation: individual cow behavior indicators (e.g., trajectory length, sinuosity, and number of trajectory segments >25 m) and social behavior indicators (e.g., distance to herd centroid, distance to nearest cow, and trajectory length relative to herd mates) were computed from GNSS coordinates in a 24-hour moving window; (iv) indicator transformation: z-score normalization to standardize indicator values; (v) anomaly detection: identification of anomalies with a ± 2.5 threshold applied to z-score values; (vi) rolling accumulation of anomalies: adding anomalies over a 3-hour moving window; and (vii) model evaluation: calving detection performance based on accuracy, precision, recall, F1-score and specificity. The effect of varying the number of anomalies required to trigger a calving alarm, ranging from 1 to 20 anomalies accumulated over the last 3 hours, was evaluated. The described workflow was tested using GNSS coordinates of 78 grazing cows during peripartum, collected in two different farms in Spain. The results (Figure 1) showed that increasing the number of anomalies required to trigger a calving alarm improved accuracy (from 0.73 to 0.95) and specificity (from 0.74 to 0.99), reducing false positives. However, this came at the expense of recall, which dropped from 0.60 to nearly 0.00, meaning fewer calving events were detected. F1 score, which balances precision and recall, peaked at 0.30 when using 11 anomalies, but declined rapidly beyond this threshold. These results highlight a trade-off between precision and recall: using few anomalies improves recall but increases false alarms, while including more anomalies improves specificity at the expense of recall. A key advantage of this approach is its near real-time operation, which enables rapid intervention and reduces calving-related risks. This GNSS-based methodology provides a scalable and automated solution for precision livestock management, minimizing unnecessary farmer intervention. In addition, the flexibility of the model allows for dynamic threshold adjustments to accommodate different management strategies. Future research will focus on refining the feature selection process, optimizing anomaly detection, and improving the adaptability of the system to different environmental and management conditions.