Automatic monitoring herbage prehensions in grazing cows using audio signals and deep learning techniques

Background Accurate monitoring of feeding behavior in grazing ruminants, particularly the detection of prehension events, is a central challenge for Precision Livestock Farming (PLF). Traditional methods, such as accelerometers, show limitations in the reliable identification of individual events. Acoustic analysis based on deep learning is emerging as a non-invasive and promising alternative. Methods This study presents two main contributions: (i) a web-based software platform (built on React.js and TensorFlow.js) for the annotation, visualization, and in-browser inference of audio signals; (ii) a comparative analysis of several 2D-CNN architectures (DenseNet-121, ResNet-101, EfficientNet-B7, and YOLO11s-cls) for the classification of prehension events. Models were trained and tested on a dataset of logarithmic spectrograms (500 ms) derived from audio recordings acquired via collars on cattle. Results Analysis revealed high performance across all architectures. Although DenseNet-121 achieved the highest weighted metrics (Accuracy 83.7%, AUC 0.90), the YOLO11s-cls model demonstrated remarkable competitiveness, achieving nearly identical accuracy (83.1%) but with significantly superior computational efficiency (4.5 ms inference time). Crucially for field applications, YOLO exhibited excellent rejection of non-relevant sounds, with a 91% Specificity on the “no-prehension” class. Conclusions The study validates the efficacy of spectrogram-based 2D-CNNs for ingestion monitoring and identifies YOLO as a promising candidate for efficiency-oriented deployment scenarios, offering a favorable trade-off between predictive reliability and low-latency requirements. The developed platform further supports this transition from research to in-field application.