Rice Science

Rice Science

• Research Paper • Previous Articles     Next Articles

Physical and Physicochemical Classification of Parboiled Rice Using VNIR-SWIR Spectroscopy and Machine Learning

  1. Laboratory of Postharvest, Campus Cachoeira do Sul, Federal University of Santa Maria, Cachoeira do Sul, Rio Grande do Sul 96506-322, Brazil; Campus de Chapadão do Sul, Federal University of Mato Grosso do Sul, Chapadão do Sul, Mato Grosso do Sul 79560-000, Brazil
  • Contact: Paulo Carteri CORADI
  • Supported by:

    The authors thank This study was supported by the Coordination for the Improvement of Higher Education Personnel, Brazil (Grant No. 001), National Council for Scientific Technological Development, Brazil (Grant No. 304966/2023-1), and Research Support Foundation of the State of Rio Grande do Sul, Brazil (Grant No. 24/2551-0001150-1). The authors would like to thank Research Group at Postharvest Innovation: Technology, Quality and Sustainability, Laboratory of Postharvest (LAPOS), Federal University of Santa Maria (UFSM), the University of Passo Fundo (UPF)-Cereal Laboratory, and the Federal University of Mato Grosso do Sul (UFMS) for their contributions in the research project.

PDF

Supplement

1

Abstract: For the classification of parboiled rice into types, the use of machine learning (ML) algorithms can optimize data processing, providing greater speed and accuracy. The objectives of this study were: (i) to investigate the spectral behavior of different types of parboiled rice (Types 1–5 and Off-Type); (ii) to identify the most effective ML algorithm for classifying parboiled rice types and (iii) to determine the optimal kernel configuration and preprocessing methods for spectral data. Samples were selected based on the maximum defect limits tolerated for each type, according to the Technical Rice Regulation. Spectral data were acquired using a spectroradiometer in the range of 350–2500 nm and were subsequently processed with different methods, including baseline correction, standard normal variation, multiplicative scattering correction, combination of these techniques with Savitzky-Golay smoothing and application of the first derivative of Savitzky-Golay smoothing. The data were analyzed using six different machine learning algorithms: artificial neural networks, decision trees, logistic regression, REPTree-decision tree variant, random forest, and support vector machine. Rice types were used as output variables and spectral features as input variables. Logistic regression and support vector machine algorithms showed the best classification performance, with accuracy rates above 97%, F-scores around 0.98 and Kappa values exceeding 0.97. Spectral preprocessing did not provide substantial improvements and incurred a high computational cost. Therefore, using raw data is a viable, efficient alternative. For practical implementation in the rice storage industry, we recommend acquiring a VNIR-SWIR hyperspectral sensor (350–2500 nm), and developing a classification model based on the Support Vector Machine algorithm with a linear kernel trained on representative local samples. In addition, the implementation of an automated real-time classification, representative sample collection, and detailed reporting for inventory and logistics optimization.

Key words: Oryza sativa L., artificial intelligence, supervised classification, support vector machine, logistic regression