Portable system for cocoa bean quality assessment using multi-output learning and augmentation

Abstract

Cocoa is an essential raw commodity in worldwide trade and requires extreme quality measurement. Precise measurement of cocoa bean ingredients, notably moisture content (MC) and fat content (FC), is essential for quality control. This paper describes a combined system that reliably uses spectroscopy, spectrum preprocessing, data augmentation, dimension reduction, wavelength selection, and advanced machine learning (ML) models to forecast these critical characteristics. The multi-output ML technique was used for MC and FC prediction. Furthermore, spectral augmentation and wavelength selection strategies were used to improve the effectiveness. The proposed method obtained a coefficient of determination (R²) = 0.992, root mean square error (RMSE) = 0.072, and a ratio of performance to deviation (RPD) = 10.620 for MC prediction, while R² = 0.984, RMSE = 0.093, and RPD = 7.919 for FC prediction. Classification analysis was also performed, and the proposed method obtained an accuracy of 96.0% for MC prediction and 90.0% for FC prediction. Moreover, statistical analysis found a t-statistic of 44.445 and a p-value of 0.001. These findings demonstrate the usefulness of this non-destructive technique, which provides a dependable, efficient, and practical option for detecting the quality of cocoa beans and has tremendous potential for use in quality control operations within the cocoa trade.