Seg2Reg-Net: An Explainable AI Analysis of Predictive Limitations in Cattle Weight Estimation

Abstract

Accurate measurement of cattle body weight is critical for improving nutrition, health management, breeding strategies, and market value. Automated methods using computer vision and deep learning provide a promising non-contact alternative for weight estimation. However, these approaches face challenges because 2D images often fail to capture full body volume, key anatomical details, and variations caused by lighting, occlusion, or background clutter, which lead to inconsistent and inaccurate predictions. To address these challenges, we propose Seg2Reg-Net, an end-To-end framework that integrates an Attention U-Net-based segmentation module with a regression network. The proposed model achieves enhanced predictive accuracy, reducing RMSE from 53.36 with standard regression to 41.56. Pixel-level explainability techniques, including LIME and SHAP, are applied to systematically interpret model predictions. LIME highlights top positive contributors such as the mid-body and hindquarters, while SHAP reveals positive, negative, and near-zero effect regions, exposing spurious correlations and background influences. These findings contribute to the refinement of data preprocessing and model design, facilitating the practical application of accurate weight predictions in livestock management.