Abstract
Monitoring heifer growth development is an important aspect of decision making in dairy farms, as a tool for health assessments, feed management, and breeding selection. Monitoring growth requires frequent body weight (BW) measurements; however, weighing calves is costly and labor-intensive. Computer vision and machine learning (ML) have been proposed as a powerful tool to predict BW in livestock animals. However, most ML techniques involve tuning a large number of parameters that increase the risk of overfitting in small data sets. Besides, these models lack interpretability and do not reveal biological associations between predictors and response variables. Symbolic regression (SR) is a ML approach that leverages genetic algorithms to find analytical equations and their coefficients to describe the output based on the input variables. In this study, we evaluated the use of SR versus gradient boosting trees (GBT) to predict BW in 67 pre-weaning Holstein dairy calves from 2 to 8 weeks of age, and BW of 57.0 ± 14.7 kg. A total of 400 3D images were captured from the dorsal area (top-down view) during weighing on a digital scale, and we extracted 27 biometric features from depth images (area, volume, length, 11 heights and widths along the dorsal area, eccentricity, and extent). Both SR and GBT were evaluated against observed BW using a nested cross-validation (5-fold for hyperparameter tuning and leave-one-out for testing). GBT achieved root mean squared error of prediction (RMSEP) = 7.7 kg, mean absolute error (MAE) = 5.8 kg, R2 = 0.59, and concordance correlation coefficient (CCC) = 0.78. Among the frequently high-ranked equations we found (1) BW = a + b × Volume, (2) BW = √Area + Width6, and (3) BW = (Width5 + c)^d . The median values for a, b, c, and d were 31.53, 0.22, 12.14, and 1.16. SR presented better predictive performance than GBT, with (1) presenting RMSEP = 6.0, MAE = 4.9, R2 = 0.67, and CCC = 0.86. While further investigations are needed on more heterogeneous data sets, SR shows the potential to predict BW using simple linear and nonlinear equations that may generalize well, with low computational cost, and with the benefit of interpretability.
Enrico Casella
Assistant Professor of Data Science for Animal Systems
Multi-disciplinary computer scientist with a focus on Artificial Intelligence and Computer Vision applications for animal systems.