Convolutional Neural Network-Based Models for Near-Infrared Prediction of Nutritional Quality in Multi-Product Animal...

Area of Science:

• Agricultural Science
• Analytical Chemistry
• Machine Learning

Background:

• Near-infrared spectroscopy (NIRS) is crucial for rapid, non-destructive feed quality analysis.
• Developing robust NIRS calibrations for heterogeneous multi-product feed datasets remains a significant challenge.

Purpose of the Study:

• To evaluate convolutional neural network (CNN)-based models for predicting crude protein (CP) and acid detergent fiber (ADF) in diverse feed types.
• To compare the performance of CNN models against conventional Partial Least Squares Regression (PLSR) models.

Main Methods:

• Development of a 1D CNN, CNN+PLS, and CNN+XGBoost models using a NIR database of forage and grain-based feeds.
• Comparison of model performance using an internal hold-out test set.
• Analysis of prediction errors and model generalization across different feed categories.

Main Results:

• CNN-based models demonstrated strong performance for CP prediction, outperforming global PLSR models with high R² and low RMSEP values.
• CNN and CNN+PLS showed modest improvements for ADF prediction, while CNN+XGBoost exhibited weaker generalization.
• Feed matrix and product category had a greater impact on ADF prediction than CP prediction.

Conclusions:

• CNN-based models, particularly CNN+PLS, show potential for enhancing NIRS-based CP prediction in complex feed mixtures.
• The advantage of CNN models for ADF prediction was limited, suggesting challenges related to feed matrix variability.
• Further validation with external datasets and varied instrument conditions is necessary for routine application of these advanced models.