Hyperspectral imaging predicts macadamia nut-in-shell and kernel moisture using machine vision and learning tools

Abstract

Tree nuts are a convenient and nutritious food source and recently considerable attention has been placed on quality assessment to provide high quality nuts and improve consumer satisfaction. Moisture is a critical parameter for tree nut quality and is routinely monitored throughout post-harvest processing. However, current direct methods to assess nut moisture are based on using limited numbers of representative sub-sets and are destructive. This study aimed to use hyperspectral imaging and machine learning (ML) to predict moisture of individual macadamia nuts during post-harvest processing. Specifically, we aimed to compare data extraction methods (automatic vs. manual) and nut orientation (base-up, base-down and combined orientations) during imaging in predicting moisture for nut-in-shell and kernels. We also explored minimum wavelength numbers to predict moisture. Spectra were obtained from images of nuts in two orientations and extracted using manual and automatic methods prior to development of partial least squares (PLSR), artificial neural network (ANN), support vector machine (SVM) and Gaussian process regression (GPR) models. Kernel moisture prediction was more accurate using automatically extracted spectra, whereas nut-in-shell moisture prediction accuracy was similar for either method. For kernels, combining the spectra from two images of nuts in base-up and base-down orientations provided similar prediction accuracy (RMSET = 0.308 %), compared with spectra from one image (RMSET ≥ 0.341 %), and for nut-in-shell, using spectra from one image also provided similar accuracy (RMSET ≈ 1.2 %) as using both images combined. PLSR models predicted moisture with very high accuracy for both nut-in-shell (R2T = 0.96, RMSET = 1.20 %, RPD = 5.15) and kernels (R2T = 0.99, RMSET = 0.308 %, RPD = 11.05) following selection of ten important wavelength bands between 760 and 967 nm. ANN and GPR also achieved equivalent (R2T = 0.99) highest accuracy predictions for kernels, however, all wavelengths were required, which would increase computational processing time for high volume applications. The important wavelength bands required to develop accurate models for macadamia moisture prediction are consistent with other food and nut products and prediction accuracies are possible for process control applications using only 10 wavelength bands. Several ML models including PLSR, ANN and GPR are suitable for use with Vis/NIR hyperspectral images to predict macadamia moisture, however, for industrial applications where high volume through-put is required, using PLSR with limited selected wavelength bands is recommended. Overall, hyperspectral imaging combined with computer vision software and ML models showed significant potential to predict moisture concentration of macadamia during post-harvest processing.