Classification of olives using FT-NIR spectroscopy, neural networks and statistical classifiers

Abstract

Green olives (Olea europaea L. cv. Ayvalik') were classified based on their surface features such as existence of bruise and fly-defect using two NIR spectrometer readings of reflectance and transmittance, and classifiers such as artificial neural networks (ANN) and statistical (Ident and Cluster). Spectral readings were performed in the ranges of 780-2500 and 800-1725nm for reflectance and transmittance modes, respectively. Original spectral readings were used as input features to the classifiers. Diameter correction was applied on reflectance spectra used in ANN classifier expecting improved classification results. ANN classifier performed better in general compared to statistical classifiers. Classification performance in detecting bruised olives using diameter corrected reflectance features and ANN classifier was 99% while it was 98% for Ident and Cluster classification approaches using regular reflectance features. Classification between solid and fly-defected olives was performed with success rates of 93% using reflectance features and 58% using transmittance features with ANN classifier while statistical classifiers of Ident and Cluster performed between 52 and 78% success rates using the same spectral readings. ANN classifier resulted 92% classification success for the classification application considering three output classes of solid, bruised and fly-defected olives using reflectance features while it performed 57.3% success rate using transmittance features.