272 Acoustic burp detection for scalable enteric methane monitoring in precision livestock farming.

Abstract This study pursued two objectives: (1) to verify the detectability of cattle burp acoustic patterns and their utility for robust classification, and (2) to develop an automated detection pipeline from long-term recordings for large-scale burp monitoring in real-world conditions. Enteric methane emissions account for an estimated 30–32% of global methane output and approximately 72% of total livestock-related greenhouse gas emissions, making them a critical environmental concern. Among these enteric emissions, ruminant burps generate 87–93%, underscoring the feasibility and impact of automating burp detection to mitigate greenhouse gases in livestock systems. A total of 6,810 hours of audio were recorded from 30 beef cattle under an open rotational grazing system, of which 100 hours were examined in detail to identify 301 expert-confirmed 5-second burp clips and 602 non-burp clips. Based on Mel-spectrogram analyses, we proposed two key preprocessing strategies—low-pass filtering (1,024 Hz) and upsampling (16 kHz)—and systematically tested them under four configurations. The Audio Spectrogram Transformer (AST) was trained using a 10-fold Monte Carlo Cross-Validation, yielding accuracy, mean average precision (mAP), burp precision, burp recall, and burp F1-score; Two-way ANOVA was conducted to assess the effectiveness of each preprocessing approach. Subsequently, an ensemble model was built from the best AST checkpoints across folds, employing a sliding-window approach (5-second window, 0.5-second step) with probability averaging and an optimized detection threshold from validation and test sets. Two-way ANOVA revealed that both low-pass filtering and upsampling significantly improved accuracy and burp F1-score (p < 0.05), with only upsampling being significant for burp recall, and no significant differences observed for mAP and burp precision. Under the best-performing setup—combining filtering and upsampling—accuracy reached 0.919 ± 0.0218, mAP 0.952 ± 0.0186, burp precision 0.935±0.0406, burp recall 0.818 ± 0.0605 and F1-score: 0.871 ± 0.0401. When tested on 10 hours of new audio (five cows, day/night), the ensemble-based pipeline delivered a burp recall of 0.821, precision 0.867, and an F1-score of 0.844, confirming strong generalization in different ambient conditions. In conclusion, this low-cost audio-based approach provides a scalable, non-invasive tool for continuous monitoring of enteric methane emissions. By accurately identifying burp events—responsible for the majority of methane release in ruminants, the proposed method holds promise for large-scale adoption in Precision Livestock Farming, ultimately contributing to more effective greenhouse gas mitigation strategies in the livestock industry.