Epsilon Archive for Student Projects

Abstract

Youngstock plays a crucial role within dairy operations as the next generation of milking cows. Despite this, the cost of rearing youngstock and the impact of rearing management on growth trajectories and associated reproductive and production performance of replacement heifers are often overlooked. This is partly due to the challenges in efficiently and effectively monitoring youngstock performance throughout rearing. This thesis aimed to validate the concept that the body weight of developing dairy heifers (5 to 20 months) can be accurately estimated through the application of 3D computer vision technology. The hypothesis is that an overhead rearview image can provide sufficient information to extract adequate body measurements for development of an accurate BW estimation model. Thereby enabling timely monitoring and management of youngstock, leading to more cost-effective management and improved health, production, and reproductive performance in future dairy cows.
The study was conducted at the Swedish Livestock Research Centre in Uppsala, Sweden, over a period of four months starting in January 2024. Data was collected on four occasions, on 101 diary heifers of two breeds: Swedish Red (n = 59) and Swedish Holstein (n = 42). The collected data included body weights, 3D images, three manual metric body measurements (hip ischial width, hip height, and external hip joint width), and manual body condition scores for heifers aged ≥ 13 months.
The body measurements were manually acquired as reference values and extracted through photogrammetry applied to 3D time-of-flight imaging. These images were acquired directly after the heifer exited the weighing scale, with an overhead caudal rear view. Pearson correlations showed strong correlations between body weights from the scale and reference body measurements: hip ischial width (r = 0.94), hip height (r = 0.88), and hipometer (r = 0.88). Age in months also showed strong correlations (r = 0.93), while body condition score was weakly correlated (r = 0.27). Model training input features were age and 3D image extracted hip ischial width, hip height, and hip width measures, whereof hip width have been established to have a strong correlation to BW in other research (r = 0.75). Model training was performed with Gradient Boosting and Group K-Fold cross-validation, achieving a regression score of 0.93 (± 0.02) between estimated and reference body weights. The relative error across all 101 heifers was 5.35 % (± 0.51 %) with a relative mean square error of 26.0 kg (± 2.5 kg).
This thesis validated the possibility of accurately estimating body weight in developing dairy heifers using 3D computer vision, indicating the potential for further model enhancement and utilization to advance precision livestock farming (PLF) technology.