APPLICABILITY OF VITAL DATA COLLECTED FROM A NON-CONTACT SENSOR FOR ESTIMATING AN INDIVIDUAL COW’S METHANE EMISSION WITH A LASER METHANE DETECTOR

抄録

 This paper explores the applicability of estimating methane emission from cows using non-contact sensors and deep learning techniques. The study was conducted on a Holstein cow housed in a tie-stall barn at Rakuno Gakuen University in Ebetsu-shi, Hokkaido, Japan. Methane concentration in the cow's breath and vital data, including heart rate, respiratory rate, and body movements, were measured using a laser methane detector (LMD) and a non-contact sensor, respectively. The LMD data was preprocessed to “extracted Mini-peaks”, which represent exhalation events, to be used as the target variable dataset. In addition, the vital data was used as the explanatory variable dataset. The Long Short-Term Memory (LSTM) model was implemented to estimate methane concentration in a cow’s breath, and the performance of the model was evaluated based on the root mean square error (RMSE) value. The results showed that the LSTM model trained with the “extracted Mini-peaks” data outperformed the model trained with the “raw LMD” data, indicating that the Mini-peaks were more closely related to vital data. Furthermore, the LSTM model trained with the “extracted Mini-peaks” data exhibited a relative error ranging from 1.9% to 11.6% in estimating daily methane emissions, compared to that calculated from observed methane concentration. The study demonstrated the applicability of estimating methane emissions from cows using non-contact sensors and the LSTM model with achieving estimation accuracy that is comparable to the LMD method. This approach could provide a cost-effective and efficient method for monitoring methane emissions from cows, contributing to the development of sustainable livestock farming practices.