Improved dairy cattle methane emission prediction via novel machine learning methods for enhanced selective breeding potential

Abstract

Excess methane (CH4) emissions produced by dairy cattle (DC) represent an inefficiency within the animal which increases agricultural greenhouse gas emissions whilst lowering production levels. The selective breeding of low CH4 emitting DC allows this inefficiency to be reduced. Prediction models enable optimal candidates for selective breeding strategies to be identified, with Machine Learning (ML) offering a promising solution. However, the compilation of training datasets from multiple experiments and different sets of cattle introduce random effects (REs) which violate their core assumptions.

To optimise the implementation of ML in the prediction of DC CH4 emissions, this thesis develops several novel solutions. A comprehensive comparative modelling framework is produced to guide researchers in optimal model selection during local development and external application(Chapters 2 and 3). A novel ML based Stacked Ensemble (SE) of DC CH4 emission prediction models published in the literature is presented, offering an alternative approach to external application capable of overcoming individual model limitations (Chapter 3). A novel RE based simulation system is developed capable of replicating compiled DC CH4 emission datasets, lowering the entry barrier to prediction model development and validation (Chapter 4). A novel online web platform is developed capable of automating the evaluation of DC CH4 emission prediction models across various regional datasets, streamlining traditional processes (Chapter 5). A novel Mixed Effect Machine Learning (MEML) framework is developed capable of correcting RE substructures via a custom covariance component, enhancing the local optimisation and external generalisation of ML models in the prediction of DC CH4 emissions (Chapter 6). And finally, novel dual-phenotype classification systems are developed capable of predicting DC emission and production levels simultaneously, promoting balanced breeding decisions (Chapter 7).

Together, these novel contributions optimise the implementation of ML in the prediction of DCCH4 emissions, enhancing the potential accuracy and accessibility of selective breeding.

Date of Award	Apr 2026
Original language	English
Sponsors	Department of Agriculture Northern Ireland & Agri-Food and Biosciences Institute
Supervisor	Haiying Wang (Supervisor) & Huiru (Jane) Zheng (Supervisor)