The Unrealised Potential of AI Solutions for Pasture-based Dairy Systems
Access status:
Open Access
Type
ThesisThesis type
Doctor of PhilosophyAuthor/s
Azubuike, Blessing NnennaAbstract
Efficient management of pasture-based dairy systems can benefit substantially from integrating
individual cow supplementation, real-time pasture monitoring, and grazing event detection, but
traditional herd-level approaches cannot address the biological variation and dynamic ...
See moreEfficient management of pasture-based dairy systems can benefit substantially from integrating individual cow supplementation, real-time pasture monitoring, and grazing event detection, but traditional herd-level approaches cannot address the biological variation and dynamic pasture growth inherent in commercial operations. This thesis developed and validated AI and machine learning methods across three interconnected domains, establishing empirical foundations for integrated precision management. Feeding optimisation was addressed through two studies. A Random Forest model combined with Differential Evolution reallocated concentrate to 81 cows across 91 days, achieving an 8% theoretical milk yield increase without additional feed cost (Chapter 3). Four evolutionary algorithms applied to 1,053 training cows and 165 optimised cows over 30 days achieved 6.63 to 8.64% theoretical yield improvements, with NSGA-II outperforming all others across 10 runs (Chapter 4). These gains are predictive estimates; controlled field validation remains necessary. Pasture biomass estimation was addressed through satellite and smartphone-based approaches, both validated against rising plate meter ground truth. An XGBoost model trained on Sentinel-2 imagery from 16 farms achieved R² = 0.70 and MAE = 216 kg DM/ha, outperforming the commercial Pasture.io platform (Chapter 5). These estimates supported automated grazing event detection across 12 farms, where Random Forest achieved within-year F1 = 0.878 and One-Class Support Vector Machine achieved cross-year F1 = 0.692, outperforming supervised models by 7.6% on Year 2 data despite average 24.2% supervised degradation (Chapter 6). Smartphone imagery achieved R² = 0.561 as a low-cost complement sensitive across high-biomass ranges where satellite indices saturate (Chapter 7). Each domain functions independently on commercial farms, though the integrated system linking all three in real time remains the frontier for future research.
See less
See moreEfficient management of pasture-based dairy systems can benefit substantially from integrating individual cow supplementation, real-time pasture monitoring, and grazing event detection, but traditional herd-level approaches cannot address the biological variation and dynamic pasture growth inherent in commercial operations. This thesis developed and validated AI and machine learning methods across three interconnected domains, establishing empirical foundations for integrated precision management. Feeding optimisation was addressed through two studies. A Random Forest model combined with Differential Evolution reallocated concentrate to 81 cows across 91 days, achieving an 8% theoretical milk yield increase without additional feed cost (Chapter 3). Four evolutionary algorithms applied to 1,053 training cows and 165 optimised cows over 30 days achieved 6.63 to 8.64% theoretical yield improvements, with NSGA-II outperforming all others across 10 runs (Chapter 4). These gains are predictive estimates; controlled field validation remains necessary. Pasture biomass estimation was addressed through satellite and smartphone-based approaches, both validated against rising plate meter ground truth. An XGBoost model trained on Sentinel-2 imagery from 16 farms achieved R² = 0.70 and MAE = 216 kg DM/ha, outperforming the commercial Pasture.io platform (Chapter 5). These estimates supported automated grazing event detection across 12 farms, where Random Forest achieved within-year F1 = 0.878 and One-Class Support Vector Machine achieved cross-year F1 = 0.692, outperforming supervised models by 7.6% on Year 2 data despite average 24.2% supervised degradation (Chapter 6). Smartphone imagery achieved R² = 0.561 as a low-cost complement sensitive across high-biomass ranges where satellite indices saturate (Chapter 7). Each domain functions independently on commercial farms, though the integrated system linking all three in real time remains the frontier for future research.
See less
Date
2026Rights statement
The author retains copyright of this thesis. It may only be used for the purposes of research and study. It must not be used for any other purposes and may not be transmitted or shared with others without prior permission.Faculty/School
Faculty of Science, School of Life and Environmental SciencesAwarding institution
The University of SydneyShare