Application of short-wave infrared hyperspectral imaging combined with machine learning on chilling injury detection in fresh coriander

Abstract

Chilling injury in coriander (Coriandrum sativum L.) occurs easily during transportation at low temperatures and cold storage to maintain freshness, leading to quality deterioration. Such chilling injury is difficult to notice at the early stage, thus rapid and accurate detecting is necessary. In this study, the application of short-wave infrared hyperspectral imaging (SWIR-HSI) combined with machine learning (ML) techniques was demonstrated in quality control of coriander. Selection of the relevant wavelength in SWIR region was conducted prior to investigation of ML algorithms for effective classification of chilling injury. The early stage of chilling injury was explored with SWIR-HSI in comparison with conventional measurement of moisture, total chlorophyll and color. Notably, early chilling injury was detected after 7 days of cold storage with SWIR-HSI, whereas it required 14 days for observation by conventional measurements. In addition, different profiles of volatile organic compounds from different cultivation batches were recognized by the current application of SWIR-HSI, which was in agreement with results from gas chromatography-mass spectrometry. These findings suggested the potential of SWIR-HSI in combination with ML, as a rapid and accurate technique for determining quality of coriander. Among the ML algorithms tested, i.e., K-Nearest Neighbors, Support Vector Machine, Random Forest Classifier, Multilayer Perceptron, and Convolutional Neural Network (CNN), CNN coupled with CNN-based wavelength selection yielded the best performance in classifying the early chilling injury with 94% accuracy.