An electronic nose (E-nose), comprising eight metal oxide semiconductor (MOS) gas sensors and a
moisture-temperature sensor, was used for classifying three quality grades of superior java cocoa beans, namely fine
cocoa dark bean < 20%, fine cocoa dark bean > 60%, and bulk cocoa bean that is a harder task compared to the
discrimination of high versus low-quality cocoa beans. The E-nose signals were pre-processed using the maximum
value method. The capability for discriminating the quality grade of the cocoa beans was checked by applying
multivariate statistical tools, namely, linear discriminant analysis (LDA), support vector machine (SVM) and
artificial neural networks (ANN). For this, the experimental dataset was split into two subsets, one for training (i.e.,
establishing the classification models) and the other for external-validation purposes. Furthermore, hyperparameter
optimization and K-fold cross-validation variant were implemented during the model training procedure to select the
best classification models and to avoid over-fitting issues. The best predictive classification performance was
obtained with the E-nose-MLP-ANN procedure, which allowed 99% of correct classifications (overall accuracy) for
the training dataset and 95% of correct classifications (overall accuracy) for the external-validation dataset. The
satisfactory results clearly demonstrated that the E-nose could be applied as a quality control tool in the cocoa
industry, requiring minimum and simple sample preparation.
The authors thank the Directorate of Research and Community Service, Ministry of Research, Technology and Higher Education, the Republic of Indonesia for providing research grants of PTUPT 2019 (Contract No. 2688/UN1.DITLIT/DIT-LIT/LT/2019). The authors also like to acknowledge the financial support given by Associate Laboratory LSRE-LCM-UID/EQU/50020/2019, strategic funding UID/BIO/04469/2019-CEB, BioTecNorte operation (NORTE-01-0145-FEDER-000004) and strategic project PEst-OE/AGR/UI0690/2014 – CIMO, all funded by national funds through FCT/MCTES (PIDDAC).
