Aplicação de espectroscopia VNIR – SWIR e aprendizado de máquinas para predição da qualidade de grãos de arroz em unidades armazenadoras e indústrias beneficiadoras

Abstract

Rice (Oryza sativa L.) is an essential food in the global diet, and its physicochemical quality is a determining factor in the production chain. The objective of this study is to evaluate the predictive and characterization capacity of the physicochemical quality of different rice types using hyperspectral sensors and machine learning algorithms. Samples of White, Black, Red, and Parboiled rice were analyzed through hyperspectral spectroscopy (350–2500 nm) and subjected to Linear Regression; traditional models such as Support Vector Machine (SVM), Random Forest (RF), and Gradient Boosting (GB); and deep learning models, including Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN). Spectral and physicochemical data were explored using multivariate analysis techniques (PCA) and cross- validation with the following metrics: Correlation Coefficient (r); Coefficient of Determination (r²); Mean Absolute Error (MAE); and Root Mean Square Error (RMSE). The results indicated that the SVM, RF, and GB models showed superior performance, with r and r² ranging from 0.95 to 1.0 and MAE and RMSE below 0.2, while deep learning models showed lower performance, especially with datasets of moderate size. Black rice stood out for its high levels of protein (~9), lipids (~2), and ash (~1.45); parboiled rice showed a higher fiber content (~2.8), while white rice was characterized by its starch content (~73). Hyperspectral spectroscopy proved to be effective in differentiating rice types, allowing the selection of relevant bands for optimized sensor development. It is concluded that the use of non-destructive technologies integrated with machine learning is promising for quality control in the rice industry, providing speed, accuracy, and sustainability in the post-harvest process.