Scholarlink

Body Condition Scoring (BCS) is essential for evaluating livestock health, nutrition, and reproduction. For sheep, especially those with long wool, manual scoring requires a level of expertise that many farmers do not possess. Despite its importance, to date, no automated approach has been proposed in the existing literature. In this context, we propose an innovative approach that automatically predicts the BCS of sheep from images, followed by an explainable AI (XAI). The pipeline follows a three-step process: (1) data collection, which was conducted under real-world farming conditions; (2) sheep detection using the Florence-2 object detection model and BCS classification via a Vision Transformer (ViT) trained on 5,848 images (11,000 images after the augmentation), and (3) visual explanation using multiple XAI methods. Our method achieved 72% accuracy using exact BCS class matches and 95% when allowing adjacent classes. This tolerance reflects the variability typically observed among expert scores.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Predicting Sheep Body Condition Scores via Explainable Deep Learning Model

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nourelhouda Hammouda,
  • Mariem Mahfoudh,
  • Rima Grati,
  • Khouloud Boukadi

摘要

Body Condition Scoring (BCS) is essential for evaluating livestock health, nutrition, and reproduction. For sheep, especially those with long wool, manual scoring requires a level of expertise that many farmers do not possess. Despite its importance, to date, no automated approach has been proposed in the existing literature. In this context, we propose an innovative approach that automatically predicts the BCS of sheep from images, followed by an explainable AI (XAI). The pipeline follows a three-step process: (1) data collection, which was conducted under real-world farming conditions; (2) sheep detection using the Florence-2 object detection model and BCS classification via a Vision Transformer (ViT) trained on 5,848 images (11,000 images after the augmentation), and (3) visual explanation using multiple XAI methods. Our method achieved 72% accuracy using exact BCS class matches and 95% when allowing adjacent classes. This tolerance reflects the variability typically observed among expert scores.