Abstract
Body weight (BW) is an important phenotype in livestock farming as it allows the monitoring of animal health and production, breeding selection, as well as feed management and efficiency. Particularly, in the context of beef cattle, final BW is an important aspect of decision making for optimal harvesting and efficient use of resources. Since performing BW measurements on a large scale is costly and labor-intensive, various studies have investigated the application of Computer Vision (CV) and Machine Learning (ML) techniques for BW predictions. However, existing works primarily focus on predictions of current BW, but do not explore forecasting BW months into the future. To address this gap, we propose an approach that leverages CV and ML to predict BW several months ahead based on animals’ currently available data. In this study, we evaluated the use of Gradient Boosting Trees (GBT) to predict BW at 12 months of age of 95 beef-on-dairy animals followed from a starting age of 159 ± 33 d to a final age of 399 ± 15 d, with final BW 630.2 ± 51.8 kg (mean ± SD). Monthly data consisted of manual measurements (height, length, girth, hip width, weight, age, birthweight, bodyweight, and gender) and biometrics extracted from 3D images captured from the dorsal area (volume, area, circularity, extent, eccentricity, perimeter, major axis length, and pixel area). A pre-processing step was implemented to fill in the data between consecutive monthly manual measurements with a linear interpolation, allowing a consistent age selection in the following experiments. When predicting BW at 360 days of age using all available features, GBT achieved RMSE = 25.8 kg, RMSEp = 4.62 kg, R2 = 0.55 using input features at 160 days of life, RMSE = 23.48 kg, RMSEp = 4.3%, R2 = 0.65 at 230, and RMSE = 15.5 kg, RMSPE = 2.8%, R2 = 0.85 at 300 d. We removed three-dimensional features such as height, and volume, and other complex features such as girth and birthweight. With these features, GBT achieved RMSE = 25.35 kg, RMSEp = 4.5%, R2 = 0.57 using input features at 160 days of life, RMSE = 22.68 kg, RMSPE = 4.17%, R2 = 0.67 at 230 days of life, and RMSE = 15.59 kg, RMSPE = 2.81%, R2 = 0.84 at 300 days of life. Interestingly, GBT seems to better learn growth patterns when fewer features are provided. In summary, this work shows the potential of CV and ML to forecast body growth in beef cattle, which can be a helpful tool to improve growth development based on optimized nutritional strategies. However, further investigations are needed to replace observed bodyweight with its predictions as input feature, and to evaluate performance on more heterogeneous datasets.
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.