Detection of defective cocoa beans using machine learning techniques and NIR spectral data fusion.

Cocoa beans are graded according to the degree of fermentation and the presence of defects, which are typically assessed using conventional cut tests. However, this method is invasive and leads to the destruction of randomly selected beans. To reduce or eliminate the destructive tendency of the conventional cut-test method, correlative, non-destructive, rapid, and objective alternative techniques are required. In this research, two portable NIR spectrometers with short wavelength ranges (740-1070 nm and 900-1700 nm) coupled with feature extraction techniques (Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA)) and multivariate algorithms (Support Vector Machine (SVM), Artificial Neural Network (ANN), and Random Forest (RF)) were used to build an optimum discrimination model for the rapid identification and grading of cocoa beans. The data from the individual spectrometers were analysed separately and fused to improve the identification results. The models were evaluated in terms of accuracy, sensitivity, and precision in both the calibration and prediction sets. The results showed that LDA-SVM outperformed other models (PCA-SVM, PCA-RF, PCA-ANN), achieving over 96.39% accuracy for individual spectrometer data and 100.00% accuracy for fused data. This study demonstrates the potential of integrating two portable NIR spectrometers with the LDA feature extraction technique and SVM algorithm as reliable, non-destructive alternatives to traditional cut tests, enabling rapid and onsite cocoa bean grading and defect detection.