Caracterização, classificação e predição da qualidade de grãos de arroz polido branco para segregação de lotes em unidades de beneficiamento, utilizando técnicas UV/VIS/SWIR e modelos preditivos

• Ano de defesa: 2025
• Autor(a) principal: Carneiro, Letícia de Oliveira
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Santa Maria; Brasil; UFSM; Programa de Pós-Graduação em Engenharia Agrícola; Centro de Ciências Rurais
• Palavras-chave em Português: Controle de qualidade; Processamento de grãos; Reflectância espectral; Modelagem supervisionada; Quality control; Grain processing; Spectral reflectance; Supervised modeling; CNPQ::CIENCIAS AGRARIAS::ENGENHARIA AGRICOLA
• Link de acesso: http://repositorio.ufsm.br/handle/1/36876

The modernization of rice processing requires methods capable of integrating analytical efficiency and precision in grain quality assessment. However, conventional classification processes are still predominantly based on physical attributes, disregarding centesimal and nutritional aspects that influence the commercial value of the product. In this context, the present study aimed to characterize, classify, and predict the physical and centesimal quality of commercial batches of white polished rice using spectral techniques and machine learning (ML) models. The experiment was conducted in four stages: (i) collection, physical classification, and preparation of samples; (ii) determination of centesimal composition by NIR spectroscopy; (iii) acquisition and analysis of spectral signatures in the UV/VIS/NIR/SWIR ranges; and (iv) classification and prediction with ML algorithms. Individual samples (healthy grains and physical defects) and commercial batch samples (Types 1 to 5 and Out of Type) were analyzed. The centesimal analyses covered moisture, starch, protein, lipids, fiber, and ash. The supervised algorithms used were Random Forest (RF), Gradient Boosting (GB), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Multilayer Perceptron (MLP), XGBoost (XGB), LightGBM (LGBM), CatBoost (CAT), and Partial Least Squares (PLS), in addition to the Logistic Regressor (LR) model as a reference. The characterization of physical-chemical quality revealed that the segregation of commercial batches based on these attributes tends to reduce the currently established classes, since certain ones show equivalence between categories regardless of physical condition. It was also observed that batches of better physical quality do not necessarily exhibit the highest levels of all compounds associated with nutritional quality, such as Type 1, which had higher levels of starch and fiber only. Spectral analysis revealed greater variation and reflectance in the UV-VIS regions, associated with the absence of pigments, and lower reflectance in the NIR-SWIR regions, due to absorption related to the molecular bonds of nutritional compounds. In the modeling stage, MLP performed best for batch classification (accuracy > 97%), followed by decision tree-based models (RF, XGB, LGBM, and CAT), with accuracy greater than 95%. In predicting physicochemical constituents, the PLS model showed the best overall performance, especially for starch and fiber (R > 0.7). The UV, VIS, and SWIR regions, particularly between 350 and 400 nm, showed greater discriminatory power between groups for both classification and prediction. It is concluded that the use of machine learning models associated with UV/VIS/NIR/SWIR reflectance spectroscopy is a promising, non-destructive, and efficient approach for the characterization and segregation of commercial batches of white polished rice, integrating physical and physicochemical parameters and contributing to the improvement of quality control in processing units.