Botanical Origins Classification and Adulteration Detection in NZ Honey using Machine Learning and Hyperspectral Imaging

Abstract

Honey fraud is a significant problem globally, with honey being the third most adulterated food product globally. Certain kinds of honey can be very high-value products, such as New Zealand (NZ) Manuka honey; quality assurance is vital to keep consumer trust in these products. This thesis aims to develop hyperspectral imaging (HSI) and machine learning techniques for honey quality assurance in a reliable and non-invasive way. Existing methods for quality assurance are typically invasive chemical tests that measure different properties of the honey. Preliminary work has used hyperspectral imaging and traditional machine learning techniques to identify honey quality, specifically in botanical origins classification. One of the most significant barriers in this vital application is the lack of available data. A large dataset of honey from different botanical origins has been previously captured and made available for this thesis. This dataset has had a significant impact on the field of honey quality assurance using HSI. Using spectral imaging, existing work on adulteration detection has previously used small datasets, sometimes only containing a single honey type. One of the significant contributions of this thesis is capturing the largest HSI dataset of adulterated honey. The dataset includes adulterated honey at four concentrations across 11 different honey types. This thesis develops two new autoencoder architectures for feature reduction, the class embodiment autoencoder (CEAE) and variational class embodiment autoencoder (VCEAE). These new structures aim to encompass both reconstruction of the input data, generalisation in the case of the VCEAE, and classification performance of the resulting features. This thesis also develops a new multi-class support vector machine (SVM) ensemble technique to solve a significant drawback of the often overlooked hierarchy based SVMs. This thesis highlights that these new techniques improve the classification accuracy over traditional methods and end-toend neural networks for the honey botanical origins classification and adulteration detection problems. This thesis evaluates the new techniques’ performance and several relevant benchmarks on five sub-problems incorporating adulteration and botanical origins classification. A novel classification system for honey quality assurance is presented, using the best methods for each sub-problem. This system encompasses adulteration detection and botanical origins classification to quickly and non-invasively identify honey fraud.