Comparative Performance Analysis of Multiple Linear Regression and Machine Learning Models for the Prediction of Methane Emissions in Korean Holstein Cows

Abstract This study aimed to develop and compare multiple linear regression (MLR) and machine learning (ML) models, including scikit-learn linear regression and artificial neural networks (ANN), to predict methane emissions from Holstein dairy cows in Korea. Methane emissions and associated variables, including body weight, dry matter intake, energy-corrected milk, and CH₄/CO₂ ratio, were measured using the GreenFeed system. The MLR models demonstrated moderate predictive performances, with adjusted R² values ranging from 0.811 to 0.820. Machine-learning models, particularly ANN models, outperformed MLR, achieving higher adjusted R² values (up to 0.839), lower root mean square errors (RMSE), and higher concordance correlation coefficients (CCC). Among the ANN models developed Models 4 and 5 showed the best predictive performance, with CCC values of 0.925 and 0.929, respectively. These results highlight the potential of artificial intelligence approaches to predict enteric methane emissions in dairy cattle accurately. Future studies should focus on expanding the datasets and validating the models in diverse production environments to enhance their practical applicability in sustainable livestock management and greenhouse gas mitigation strategies