Predictive modeling of rice milling degree for three typical Chinese rice varieties using interpretative machine...

Area of Science:

• Agricultural Engineering
• Food Science
• Computer Vision

Background:

• Over-milling brown rice leads to significant economic and nutritional losses.
• Accurate detection and prediction of rice degree of milling (DOM) are challenging, especially during moderate processing.
• Maintaining rice quality and nutrition requires precise control over the milling process.

Purpose of the Study:

• To develop an automated, non-destructive, and cost-effective method for predicting rice milling quality.
• To establish a robust model for predicting the degree of bran layer remaining (DOR) and its relationship with DOM.
• To identify key image features influencing milling quality prediction.

Main Methods:

• A custom grain image acquisition platform was utilized for capturing rice images.
• Image processing techniques were employed to extract color, texture, and shape features from rice grains.
• Machine learning models, including Catboost, were developed and optimized using cross-validation and grid search for DOR prediction.

Main Results:

• The optimized Catboost model achieved a prediction accuracy of 91.24%, with precision, recall, and F1-score exceeding 90%.
• Shapley additive explanations revealed that color features (e.g., YCbCr-Cb_ske) have the highest importance, followed by texture (e.g., GLCM-Contrast) and shape.
• The model demonstrated significant improvements in accuracy from an initial 84.28% to the optimized 91.24%.

Conclusions:

• Image processing combined with machine learning offers a viable solution for automated and non-destructive rice milling degree prediction.
• Feature importance analysis provides valuable insights for refining future milling quality prediction models.
• This approach can guide improvements in rice milling practices to enhance nutritional retention and reduce broken rice yield.