Early prediction of final weight in nomadic sheep flocks using machine learning algorithms.

The current study explores the implementation of machine learning algorithms in nomadic sheep flocks. For this purpose, a total of 1,379 Güney Karaman sheep, a breed highly favoured by nomads, were used to develop an early prediction model for final weight. As target variables; animal records such as birth date, age of dam at the birth, age of sire at the time of insemination, sex, birth type and weights taken at birth and weaning were used along with the meteorological values such as precipitation, dew point, wind speed and temperature humidity index (THI). Single decision tree (DT) and ensemble learners built by multiple decision trees namely random forest (RF), extra tree regressor (ERT) and extreme gradient boosting (XGB) were selected as machine learning models. These algorithms were assessed in terms of coefficient of determination (R2), adjusted coefficient of determination (R2adj), relative absolute error (RAE), coefficient of variation (CV), mean squared error (MSE), root relative mean squared error (RRMSE), mean absolute error (MAE), mean absolute percentage error (MAPE). Ensemble learners were found to have similar prediction performances, with a minimum R2 score of 0.86, a minimum R2adj of 0.86, a maximum MSE of 2.17, a maximum RRMSE of 5.17, a maximum MAE of 1.02, a maximum MAPE of 3.66, a maximum RAE of 0.33, and a maximum CV of 5.17) in the evaluated metrics overperforming single decision tree (R²=0.77, R2adj = 0.77, MSE = 3.56, RRMSE = 6.62, MAE = 1.33, MAPE = 4.73, RAE = 0.43, CV = 6.62). The findings of this study suggest that ensemble machine learning algorithms, trained on animal records combined with meteorological data, can be effectively utilized for early weight prediction in sheep raised under nomadic management systems. This approach provides valuable insights that can aid nomads in planning and decision-making.