A machine learning approach to simulate cattle growth at pasture using remotely collected walk-over weights

The growing interest in the use and implementation of remote and automated technologies, such as walk-over (WO) weighing, has made the availability of a large volume of data. However, animal behaviour, sensitivity and repeatability of WO weighing scales and animal's physiological state could impact the accuracy of recorded weights. Therefore, WO weight data needs through processing before it can be utilized effectively for developing livestock management tools. This study is designed to implement a machine learning (ML) approach to develop a simulation model capable of predicting cattle growth using remotely collected WO weight data from pasture-based beef production. WO weight data collected from two beef production farms (Belmont and Tremere) of central Queensland, Australia has been used to train an XGBoost ML model which is then implemented in an algorithm that takes the number of male and females of Belmont Red, Brahman, Composite and unknown breed of animals, number of days to simulate, birth date range of animals and weather conditions at the farm to simulate as input parameters and predicted daily weights as the output. The simulation model obtained Lin's concordance coefficient (CCC), root mean square error (RMSE), mean absolute percentage error (MAPE) and coefficient of determination (R 2 ) of 0.84, 32.60 kg, 15.23 % and 0.77 respectively for test set 1 and CCC, RMSE, MAPE and R 2 of 0.82, 38.30 kg, 15.20 %, and 0.73 respectively for test set 2, when compared with the actual WO weights. The mean difference between the simulated and observed weights is found to be −1.2 kg (SD 27.3 kg), where the 95 % agreement limit (at 1.96 SD) is −54.8 kg to 52.3 kg. The results suggest that autonomously collected WO weights can be an important data source for developing strategies for livestock management, and the simulation model developed in this study can be used by cattle producers to predict the growth patterns of their herd and by researchers to generate cattle growth data to implement in the development of livestock managements tools. • Autonomously collected walk-over weights were implemented to train XGBoost machine learning model. • A simulation algorithm capable of outputting daily weight distribution of cattle herd was developed. • The simulated results were compared statistically and graphically with the observed walk-over weights. • The model can be used to predict daily weight distribution of cattle pastured in tropical or subtropical conditions.