NIR spectroscopy and pattern recognition methods for classification and characterization of cocoa varieties.

Cocoa is currently a valued product for chocolate processing, produced in several countries in Africa and America. One of the most destructive diseases of cocoa is the fungus Monilliophthora perniciosa, which instigates the ‘witch’s broom disease’, causing ultimate plant damage [1]. As an alternative to prevent this fungus occurrence, the use of resistant varieties of high productivity has been developed by genetic breeding programs [2]. However, these different varieties present a wide range of diverse chemical composition, making it difficult for the processing industry to standardize parameters during processing.
NIRS has been reported as fast and non-destructive method already used in various fields [3], including cocoa characterization. It has been used to determine fat, nitrogen and moisture content of cocoa powder [4], protein, fat, starch and proanthocyanidins in cocoa [5], caffeine, theobromine and epicatechin in cocoa [6] and sucrose in chocolate [7].
The current work proposes application of NIR spectroscopy as an analytical method to classify different varieties of cocoa beans and predict chemical and physical attributes of cocoa for both intact and ground samples. To our knowledge, no previous study was carried out for classification of cocoa samples into different varieties using spectral information.
Five different varieties of cocoa beans were ground and sieved for standardization of particle size for further analyses. Spectral data were collected in reflectance mode for intact and ground samples over the wavelength range of 400 to 2,498nm at 2nm intervals.
PCA was carried out with the purpose to scrutinize the major influence of the different factors of cocoa varieties in the spectral information. A few wavelengths were selected to be used as predictors for the LDA model.
A few broad local absorption maxima were noticeable around 1190, 1460 and 1950 nm; absorption at these wavelengths correspond to O-H, C-H, N-H stretch first and second overtones and combination bands that could be attributed to water absorption and protein changes. PCA was carried out with a full cross-validation to identify the qualitative discrimination in the spectra among samples. Results showed that in general there was satisfactory discrimination between the five varieties of cocoa investigated. The satisfactory group separation results obtained from PCA indicate that the data should provide enough information to develop classification models for cocoa varieties. The first three principal components were responsible for 68%, 27% and 3%, respectively, of the total variance among the whole cocoa samples. It was indicated that the five cocoa varieties had different spectral patterns and can be distinguished into separate classes, meaning that the profiles of objects in the same cluster are very similar and the profiles of objects in different clusters are distinct.
After identifying the optimal wavelengths, spectral data was reduced to the most relevant wavelengths observed. LDA was employed in order to classify cocoa bean varieties based on optimal bands selected from spectral information. For whole beans, LDA could discriminate all samples of five different varieties with 100% accuracy. Regarding ground samples, the accuracy rate was 92% of samples analyzed, with a few misclassified samples appearing.