Advanced Multimodal Prediction of Components of Livestock Feed Materials Using Knowledge Distillation

Accurate analysis of livestock feed quality is critical for enhancing productivity and supporting sustainable farming practices. This study explores the integration of red, green, and blue (RGB) and near-infrared (NIR) imaging modalities, leveraging their complementary strengths, RGB for physical properties and NIR for chemical compositions, to predict key nutritional metrics. A novel knowledge distillation model was developed to transfer insights from a complex teacher model to a simpler student model. The approach involved training three types of models: single-channel (RGB or NIR), double-channel (RGB and NIR), and knowledge distillation models. Key evaluation metrics, including mean squared error (MSE), mean absolute error, and root MSE, validated the model's predictive accuracy. Experimental results demonstrated that the knowledge distillation model significantly outperformed both single- and double-channel models, achieving a 91.86% reduction in the MSE compared to RGB single-channel models, an 89.68% reduction compared to NIR single-channel models, and a 63.43% improvement over double-channel models. This study provides a robust, efficient, and cost-effective solution for feed quality assessment, highlighting the transformative potential of multimodal imaging and machine learning in precision agriculture.