Machine learning based framework for the detection of mushroom browning using a portable hyperspectral imaging system

Abstract

White button mushrooms (Agaricus bisporus) are soft-cellular and susceptible to color changes accounting significant postharvest losses due to brown spots on their cap surface. In this study, a portable hyperspectral imaging camera in the visible-near infrared wavelength range (400–1000 nm) was explored to determine browning effects in time series on white button mushrooms stored at 4 °C while relative humidity kept constant at 60 % and 80 % relative humidity (RH), respectively. This study proposed the combination of unsupervised training algorithms using principal component analysis (PCA) combined with fuzzy C-means clustering (FCM) for mushroom image segmentation and calibration data selection for further supervised training approaches. Thus, the supervised classification models of k-nearest neighbor (k-NN) and partial least square-discriminant analysis (PLS-DA) were developed for the determination of browning patterns on mushrooms and achieved the correct classification rate (CCR) values of 97.6 %-99.8 % and 94.7 %-97.7 %, respectively. Overall, this time-series study during storage demonstrated the potential of a portable hyperspectral imaging camera combined with machine learning models for post-harvest mushroom quality control purposes.