Honey Botanical Origin Classification using Hyperspectral Imaging and Machine Learning

Abstract

Honey is a world commodity which makes a significant contribution to the economy, especially, for honey producing countries like New Zealand. This research focuses on the development of non-invasive honey botanical origin classification based on hyperspectral imaging and machine learning. Based on current knowledge, there has been no prior comprehensive study about that topic. Development of non-invasive methods for honey classification is targeted as alternatives to time-consuming, laborious and destructive chemical-oriented methods. Hyperspectral imaging is an advanced imaging technology widely used for food quality and safety assessment. This technology is able to acquire spectral information over a particular spatial region (spectral-spatial information) carrying characteristics of assessed objects. The present research develops a hyperspectral image acquisition methodology that consists of: samples preparation, acquisition stage equipped with homogeneous lighting for reflectance and transmittance configurations, hyperspectral imager working on 399.40–1063.79 nm, applied to 58 honey products available in New Zealand. A specific module written in Python was designed to greatly ease raw data processing and machine learning method development. A systematic method development was applied to investigate the best implementation of spectra processing, feature selection and classification of honey botanical origin. The spectra processing included methods for noise-invariant spectra extraction on segmented regions, adaptive calibration, robust noisy band elimination, and normalisation. The feature selection obtained bands that are relevant to botanical origins through type-guided methods combinations. The classification methods adopted a hierarchical design for filtering parasitic spectral information and predicting botanical origin using the potential features. Well-defined experimental scenarios were implemented to examine the proposed methods. The experimental results showed that the proposed spectral processing methods successfully cleaned and normalised spectral information from which the relevant bands could be extracted based on the proposed feature selection strategy. The developed classification models demonstrated promising accuracy for predicting honey botanical origins using the potential bands. The proposed approach of hyperspectral imaging and machine learning for honey botanical origin classification is a successful, non-invasive, fast and automatic approach.