Evaluation of a global repository of published dairy cattle methane emission prediction models on a northern irish dataset: Individual performance vs a stacked ensemble approach

Dairy Cattle (DC) methane (CH4) emission prediction models offer an accessible alternative to traditional recording technologies but struggle to maintain performances when applied outside the scope of their original training data. This study evaluated 23 DC CH4 emission prediction models upon a local Northern Irish data set to highlight influential factors affecting model generalization and proposed a novel machine learning based stacked ensemble of the 23 models combined to address the limitations they faced. Individual model concordance correlation coefficients (CCC) ranged from 0.50 to 0.86, and root mean square prediction errors (RMSPE), from 27.99% to 10.82%. The traditional linear approach the majority of models employed subjected them to high mean and slope biases when conditions differed from training. Nonlinear models proved more adaptable through their reflection of genuine biological reactions, such as zero emissions at zero intake, offsetting mean bias, and the decrease in CH4 energy output as feeding levels increase above maintenance, offsetting slope bias. Yet the novel machine learning based stacked ensemble proposed here was the best performing model overall, achieving a RMSPE of 9.61% and CCC of 0.89, its meta learner algorithm able to combine the strengths of multiple individual models together simultaneously. Thus, DC CH4 emission prediction research should continue to explore a diverse range of modeling approaches in various developmental ranges, as once combined within a machine learning based stacked ensemble during external prediction, their collective strengths can offset individual weaknesses, with the more diverse its foundation, the more data sets it can adapt to.