Screening for Coffee Adulteration Using Digital Images and SPA-LDA

摘要

In this paper, we propose a new methodology to identify adulterations in ground roasted coffees (due to the presence of husks and sticks) using digital images and the successive projections algorithm for variable selection in association with linear discriminant analysis (SPA-LDA). A simple document scanner was used for capturing the images, and a Petri dish support with eight circular holes (one for each sample) to be scanned was employed. Color histograms in the hue-luminosity-saturation (HLS) channels extracted from the digital images were used as input data and statistically evaluated using supervised pattern recognition techniques. For comparison with SPA-LDA, soft independent modeling by class analogy (SIMCA) and partial least squares discriminant analysis (PLS-DA) were also used. In general, SPA-LDA provided significantly better performance than the other classification models, reaching a mean accuracy of 92.5 % for both the training and test sets, while SIMCA and PLS-DA attained only 71.5 and 85.5 %, respectively. More specifically, all of the models presented high rates (above 90 %) for sensitivity and specificity (in the test set samples classification), except SIMCA, which presented a specificity rate of 76 %. Moreover, the SPA-LDA model generally showed the lowest classification error rates. As such, it is a more adequate chemometric tool for discriminating pure coffee samples and adulterated by husks and sticks. The proposed strategy avoided laborious sample preparation, and additional operational costs, assessing coffee adulteration by husks and sticks.