Behavior recognition of grazing cattle with hybrid segmentation of multi-sensor features

• Dynamic behaviors showed irregular and rapid variations of the signal profiles. • Static behaviors remain constant signal amplitudes and stable signal profiles. • The hybrid segmentation method, HSF, is developed to extract multi-behavior features. • A behavior recognition framework based on HSF was developed for grazing cattle. • HSF had similar accuracy of 98.85% while being faster and lighter than baseline models. The use of multiple sensors to monitor grazing cattle behavior can improve feed intake estimation, optimize pasture utilization, and support cattle health management in livestock production. Current behavioral recognition studies predominantly rely on static time windows to extract sensor features globally across all behavior types. While static behaviors (e.g., standing, lying) can be recognized with high accuracy, the recognition of dynamic behaviors (e.g., walking), especially mixed activities such as walking-ruminating, remains suboptimal. Static windows often lead to behavior truncation or the contamination of a single segment with multiple actions. Therefore, we propose a hybrid segmentation framework (HSF) that first employs a pre-classifier to distinguish between dynamic and static behaviors. Static behaviors are then processed using fixed length windows, whereas dynamic behaviors are segmented with variable length windows based on the PELT change point detection algorithm. Each category subsequently undergoes tailored feature extraction and classifier training. The framework is applied to classify eight behaviors of grazing cattle, standing, lying, foraging-biting, foraging-chewing, standing-ruminating, lying-ruminating, walking, and walking-ruminating, using data from integrated pressure and inertial measurement unit (IMU) sensors. Experimental results clearly show that HSF achieved a mean F1 Score of 98.85%, delivering absolute improvements of 2.52 and 3.07 percentage points (averaging 2.80%) over the standard static window (SWS, 96.33%) and the sliding static window (SWS-Slide, 95.78%) baselines, respectively. Compared to other advanced approaches, the proposed framework delivers better performance while requiring less computational time and memory. This study provides a solid algorithmic foundation for recognizing multiple behaviors in grazing cattle, with potential applications in feed intake estimation, parturition detection, and other precision livestock farming tasks.