HYPERSPECTRAL IMAGING SYSTEM FOR MATURITY STAGE CLASSIFICATION OF DURIAN PULP USING BAYESIAN OPTIMIZED MACHINE LEARNING ALGORITHMS

Non-destructive classification of fruits based on the maturity stage is beneficial to the consumer and fruit industry. Improper ripening can lead to low eating quality and economic loss for the producers. In this research, a hyperspectral image (HSI) of durian pulp was obtained using a reflectance-based system. The mean raw spectra of the durian pulp were extracted and pre-treated using standard normal variate (SNV). An assessment of maturity stage classification (unripe, ripe, and overripe) on the full wavelength (900-1600 nm) was performed. The comparison among the machine learning (ML) algorithms (random forest (RF), support vector machine (SVM), and k Nearest Neighbours (kNN)) was carried out, where the hyperparameters were tuned using Bayesian optimization and the 3-fold cross-validation method. The samples were split into training (70%) and test (30%) set using stratified random sampling. In terms of overall classification accuracy and kappa coefficient, SVM (88.5%, 0.83) performed better than RF (84.6%, 0.77) and kNN (73.1%, 0.59). The results show that the classifiers (SVM and RF) can fairly differentiate the ripening stage of durian pulp using HSI.