Improving the accuracy of green bean palm civet coffee purity classification using wrapper feature selection

Abstract

Palm civet coffee, a highly prized specialty coffee, faces counterfeiting due to its limited production. The lack of reliable detection methods necessitates the development of non-destructive sensing techniques. This study investigates the use of machine vision and feature selection to classify the purity of palm civet coffee. A dataset of 101 image features (11 color and 90 textural) was extracted from coffee bean images. A wrapper-based feature selection approach, employing K-Nearest Neighbors (KNN), Random Forest (RF), and Support Vector Machine (SVM) classifiers with four optimization algorithms (Bat Algorithm, Cuckoo Search, Genetic Algorithm, and Grey Wolf Optimizer), was used to identify the most informative features. The results demonstrate that a Random Forest classifier, optimized using Grey Wolf Optimizer with 500 trees, achieved the highest accuracy (0.981) using a subset of five features: Blue_Mean, Hue_Entropy, Gray_Inverse, S_HSL_Correlation, and Green_Cluster. These findings suggest that machine vision, combined with feature selection, holds promise for developing a robust and non-destructive method for detecting palm civet coffee counterfeiting.