Accurate prediction of body weight (BW) is essential for the management and genetic improvement of poultry in smallholder systems. This study evaluated and compared the performance of three tree-based machine learning (ML) algorithms—decision trees (DT), random forest (RF), and gradient boosted trees (GBT)—for predicting the BW of local Ethiopian ducks. Models were developed using two predictor sets: morphometric traits alone and a combination of morphometric and categorical predictors (sex and region). Results indicated that ensemble methods (RF and GBT) significantly outperformed the DT model. The GBT model achieved the highest predictive accuracy using morphometric traits alone (test-dataset R² = 0.75, RMSE = 0.40 kg), with RF performing comparably (R² = 0.73). The inclusion of categorical variables did not enhance predictive performance, confirming that morphometric traits alone are sufficient for optimal accuracy and that they capture the biological variance linked to sex and region. However, the ensemble methods maintained this high performance with mixed data types, demonstrating their robustness. Across all models, chest circumference (CC) was consistently identified as the most important predictor. This study demonstrates that tree-based ensemble algorithms, particularly GBT, provide a robust, interpretable, and field-applicable framework for accurate BW prediction in ducks, offering a valuable, low-cost tool for selection and management in resource-limited settings where weighing scales are unavailable.
Megersa et al. (Wed,) studied this question.