Predicting body weight and efficiency traits in beef cattle based on depth image data

Abstract

Advancements in sensor technologies have enabled the development of automated, non-invasive, and cost-effective phenotyping tools capable of capturing novel traits relevant to livestock management and genetic selection. The objective of this study was to compare the performance of different analytical approaches for predicting body weight, dry matter intake, and feed efficiency-related traits in beef cattle using image-derived biometric traits, assess the relative contribution of individual biometric traits to the predictive performance, and evaluate how the number of images per animal and image selection strategy affect model performance. A total of 196 commercial Angus steers were video recorded using a RealSenseTM D455 depth camera, from which the biometric body traits projected volume, surface area, length, width, and height were extracted. These traits were and used as predictors in several analytical approaches, including linear regression, partial least squares, elastic net, random forest, support vector machine, gradient boosting machine, and neural networks. Model performance was evaluated using leave-one-out cross-validation. Across all traits, linear regression achieved performance comparable to or exceeding more complex models (e.g., random forest, gradient boosting machine) while offering greater interpretability and computational efficiency. Body weight was predicted with high predictive performance (R2 = 0.96; MAE = 10.12 kg) using projected volume, flat surface area, and body length, whereas dry matter intake was moderately predicted (R2 = 0.59, MAE = 0.602 kg) using projected volume and flat surface area. In contrast, image-derived biometric traits showed limited ability to predict residual feed intake, residual average daily gain, or residual intake and body weight gain (R2 < 0.01). Prediction accuracy was highest when biometric traits were summarized across all available images per animal, although comparable performance was achieved using one or more randomly or centrally selected images. Overall, these results demonstrate that image-derived biometric traits can accurately predict body weight and provide informative predictors of dry matter intake, with limited predictive value for the feed efficiency traits investigated in this study. These findings demonstrate that image-derived biometric traits provide a scalable, non-invasive phenotyping approach that supports both management and breeding selection decisions in beef cattle.