Phenomics

Genetic Parameters of Uncertainty in Morphometric Traits Collected by an Autonomous Vision System in Holstein Cows

---

Abstract

Automated computer vision systems provide a scalable approach for generating digital phenotypes in dairy cattle. Autonomously collected image data may vary due to animal-specific factors (e.g., posture, behavior, and coat pattern) and technical factors (e.g., lighting conditions and camera position), which can influence the stability of deep learning (DL) inference. For morphometric traits (MT), true biological variation in mature animals is expected to be minimal at the daily scale. Consequently, daily within-cow variability observed across repeated DL-based estimates of MT can be interpreted as an empirical measure of inference uncertainty and provides an opportunity to investigate the extent to which this uncertainty is affected by animal-specific factors. The objective of this study was to estimate the heritability and repeatability of uncertainty in MT inferred from DL algorithms. RealSense D435 cameras were installed at the sort gates of a voluntary milking system in a commercial Holstein herd in California. Image acquisition was fully automated using a C++ program running on edge-computing devices and triggered by an object-to-camera distance threshold. A total of 276,642 top-down images were collected from 918 cows between September and November 2024. A YOLOv8 keypoint detection model was trained using manually labeled images from September and applied to subsequent images to predict body landmark coordinates at the hips, rump, shoulders, and midsagittal plane. Pixel-based distances between predicted landmarks were used to derive rump width (RW), body length (BL), hip width (HW), shoulder width (SW), right rump length (RLR), and left rump length (RLL). For each cow-day, daily means and standard deviations were computed per trait from multiple DL-based predictions, and trait-specific uncertainty was defined as the natural logarithm of the coefficient of variation (lnCV), yielding 15,469 daily observations per trait. The natural logarithm of the number of images collected per cow per day (lnNI) was also considered. Variance components were estimated using GBLUP repeatability models. The number of images collected per cow-day ranged from 3 to 286, while mean lnCV values ranged from 0.66 ± 0.47 (HW) to 1.67 ± 0.63 (BL). Heritability estimates for uncertainty traits were low but non-zero for most traits, ranging from 0.00 (lnCV_HW) to 0.05-0.06 for lnCV_RW, lnCV_BL, and lnNI, while repeatability estimates ranged from 0.08 to 0.20. In contrast, the original image-derived morphometric traits showed moderate to high heritability (0.25 ± 0.05 to 0.47 ± 0.07) and repeatability (0.47 ± 0.02 to 0.89 ± 0.01), confirming that the automated imaging system generated biologically meaningful phenotypes. The results of this study indicate that only a small proportion of variation in image-derived inference uncertainty is attributable to cow-specific factors. Nevertheless, treating DL inference uncertainty as a quantitative trait may inform the refinement of autonomous imaging pipelines and modeling strategies in livestock imagenomics.

Keywords: 2026

How to Cite:

Alves, A., (2026) “Genetic Parameters of Uncertainty in Morphometric Traits Collected by an Autonomous Vision System in Holstein Cows”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2287383. doi: https://doi.org/10.31274/wcgalp.24286

Rights: 1