Quality analysis of cocoa using portable Near-Infrared Spectroscopy (NIR): Challenges towards geographical differentiation

Abstract

Introduction. The combination of calibration curves generated by Near-Infrared (NIR) spectroscopy with deep learning offers an opportunity to develop methods for discriminating the quality and origin of cocoa, supporting strategies for territorial valorization and traceability of differentiated cocoas. Objective. To validate the applicability of a method for discriminating cocoa according to its geographical origin using portable NIR spectroscopy and deep learning techniques. Materials and methods. A total of 193 sample of fermented and dried cocoa beans from different regions of Costa Rica were collected, using 72 samples for calibration and 121 for prediction. The samples were analyzed to determine their proximal composition, titratable acidity, and phenolic compounds. A NIR spectrophotometer was used to collect spectral data (400-1700 nm). Spectral data preprocessing allowed for the development of regression models to predict chemical characteristics. For the geographical classification model, noise was removed from the spectra, and cluster analysis was performed using the Gower distance and Ward’s clustering method on components obtained through Principal Component Analysis (PCA). Results. Fat was the main component present in the sample set (>39.67%). Spectral analysis demonstrated that NIR can differentiate cocoa based on the degree of fermentation and phenolic compound content. The extended partial least squares (XLS) regression model showed the best predictive capacity for chemical properties. Clustering by geographical origin identified four groups mainly influenced by chemical properties related to post harvest practices. Conclusion. The linear regression model used proved superior in predicting proximal chemical characteristics. It was observed that limited genetic diversity and standardized post-harvest practices could reduce the quality variability associated with geographical origin, limiting the utility of NIR in origin identification and traceability. It is suggested to explore broader spectra and additional equipment for advanced multivariate analyses.