Application of XGBoost and Random Forest algorithms for body weight prediction in Blackbelly sheep using biometric measurements

BACKGROUND: Reliable and cost-effective live weight determination is critical for improving both management and selection efficiency in small-scale sheep farming. In recent years, machine learning approaches based on phenotypic biometric measurements have emerged as a practical alternative for producers with limited access to weighing infrastructure. This study aimed to compare the performance of two machine learning algorithms (XGBoost and Random Forest) for estimating live weight based on biometric measurements in Blackbelly sheep (60 females, 60 males) raised in humid tropical climates. RESULTS: Using the obtained biometric measurements, both algorithms were evaluated separately for training and test sets. The XGBoost algorithm demonstrated high accuracy on the training data (R2 = 0.981; MSE = 0.416; RMSE = 0.645; MAE = 0.486; AIC = 150.513 and BIC = 157.841) but had limited generalization to the test data (R2 = 0.813; MSE = 3.296; RMSE = 1.816; MAE = 1.440; AIC = 146.125 and BIC = 150.791). In contrast, the Random Forest algorithm produced more balanced and stable predictions on both training (R2 = 0.969; MSE = 0.630; RMSE = 0.794; MAE = 0.617; AIC = -31.230 and BIC = -21.459) and test data (R2= 0.873; MSE = 2.188; RMSE = 1.479; MAE = 1.105; AIC = 35.401 and BIC = 41.623). Optimal hyperparameters were determined for both models, and model fit criteria were compared. The results revealed that Random Forest offers a more reliable and stable option for estimating live weight. CONCLUSIONS: The current study demonstrates that machine learning models based on phenotypic biometric measurements can be used in decision-support processes in small-scale sheep farming operations. The Random Forest algorithm stands out as a suitable tool for improving production and selection efficiency in live weight estimation. Furthermore, validation studies encompassing different sheep breeds and larger data sets will strengthen the reliability and generalizability of the developed models.