A comprehensive explainable machine learning framework for predicting sheep body weight

This study introduced an explainable machine learning (ML) framework that integrates data preprocessing, optimized ensemble tree-based algorithms, and SHapley Additive exPlanations (SHAP) to estimate the body weight (BW) of sheep from various morphometric measurements. The data include records of BW and nine body measurements, such as body length, withers height, face length, chest girth, ear length, length between ears, paunch girth, fat-tail width, and fat-tail length, that served as predictors, from 651 Harnai sheep, an indigenous fat-tailed breed from Pakistan. After checking multicollinearity and identifying anomalies using Isolation Forest, several machine learning models, namely Multilayer Perceptron, Support Vector Regression, and six ensemble tree-based models (Random Forest, Extra Trees, XGBoost, LightGBM, CatBoost, and AdaBoost) were developed. Hyperparameters of models were optimized with Optuna framework inside a nested cross-validation (NCV) setup to avoid overly optimistic performance estimates. Among all models, the Extra Trees performed the best (test R² = 0.94; MAE = 1.63 kg; RMSE = 2.18 kg; MAPE = 4.45%). SHAP results suggested that fat-tail width, face length, and ear length were the strongest contributors to BW predictions, an outcome that makes biological sense for fat-tailed breeds, where tail morphology can reflect overall body condition. The proposed framework offers a combination of good predictive accuracy, interpretability, and computational efficiency. These characteristics make it potentially useful for precise livestock management, morphometric-based selection programs, and low-cost automated weight estimation in resource-constrained settings. A simple graphical user interface was also developed to make the model easier to use in practical field situations.