Batch-Level Average Market Weight Estimation and Interpretability Analysis for Pigs with Regional K-Means Clustering and AMFormer

In modern pig farming, batch-level weight estimation is important for optimizing feeding management and improving economic efficiency. Current vision-based weight estimation methods are often affected by acquisition conditions and mainly focus on individual pigs, making them less suitable for batch-level decision-making. To address this problem, this study focuses on concurrent estimation of batch-level final average market weight and explores the use of tabular production data for non-invasive body weight estimation in pig groups. Based on more than 35,000 production records from a modern pig farming enterprise, we propose RC-AMFormer. The method uses region-wise K-means clustering to characterize underlying farming heterogeneity, introduces cyclic month encoding and organizational hierarchy features, and adopts the arithmetic attention mechanism in AMFormer to model feature interactions. Among the ten comparison models, RC-AMFormer achieves an MAE of 2.64, an MSE of 11.77, and an R2 of 0.67 on the test set. Compared with XGBoost, it reduces MSE and MAE by 4.03% and 1.43%, respectively. SHAP analysis identifies important features associated with market weight. These results suggest that tabular production data provide a feasible basis for batch-level non-invasive body weight estimation and may also support similar production scenarios.