ANN-Integrated Electronic Nose and zNosetm System for Apple Quality Evaluation

Abstract

The fresh produce industry generates more than one billion dollars each year in the U.S. market. However, fresh produce departments in grocery stores experience as much as 10% loss because the apples contain undetected defects and deteriorate in quality before they can be sold. Apple defects can create sites for pathogen development, which can cause foodborne illness. It is important to develop a non-destructive system for rapid detection and classification of defective fresh produce. In this study, an artificial neural network (ANN) based electronic nose and zNoseTM system was developed to detect physically damaged apples. Principal component analysis was used for clustering plot and feature extraction. The first five principal components were selected for the electronic nose data input, and the first ten principal components were selected for the zNoseTM spectrum data. Different ANN models, back-propagation networks (BP), probabilistic neural networks (PNN), and learning vector quantification networks (LVQ), were built and compared based on their classification accuracy, sensitivity and specificity, generalization, and incremental learning performance. For the Enose data, the BP and PNN classification rate of 85.3% and 85.1%, respectively, was better than the LVQ classification rate of 73.7%; for the zNoseTM data, the three ANN models had similar performances, which were less favorable than the Enose, with classification rates of 77%, 76.8% and 74.3%. The three ANN models' performances were also measured by their sensitivity, specificity, generalization, and incremental learning.