FLOUR CLASSIFICATION USING HYPERSPECTRAL IMAGING AND MACHINE LEARNING

Hyperspectral imaging (HSI) has been widely used for classification or prediction of parameters in food matrices. The aim of this study was to evaluate the potential of near infrared hyperspectral imaging (900-2500 nm) and machine learning to detect different types of pure flour and flour added of fiber in different percentages (3 types of flour, 10 types of fiber, 3 percentages). A total of 1260 images were acquired and Principal Component Analysis (PCA) and Neural network (NN) were used to visualize and classify the samples, respectively. The data set was split into training set (80% of samples) and validation set (20% of samples) and the metrics used to evaluate the results were precision, recall and accuracy. For the first case, the Machine learning (ML) algorithm was able to correctly classify 100% of flour samples, both in training and validation sets, independently if it was pure or mixed samples. The second case seemed more complex to solve, since there were 10 types of fiber mixed together with different flours in different percentages. The PCA scores showed no clear separation among samples. Still, NN was able to classify 75% of samples in the training set and 71% in the validation set. In the third case, the data set was classified according to the percentage of fiber in each sample and the ML algorithm correctly classified in 91% both training and validation sets. The results showed that HSI associated to machine learning can be very efficient to evaluate the same data set, which include different flours and fibers, based in different outputs (type of flour, type of fiber or percentage of flour).