Assessment of Macadamia Nutrients Using Hyperspectral Data and Machine Learning

With the increasing demand for high-quality macadamia products, there is a pressing need for efficient and accurate nutrient analysis methods to maintain product quality and nutritional integrity. Compared to existing nutrient analysis methods which are destructive, expensive, and slow, hyper-spectral imaging (HSI) offers a promising solution providing non-destructive and rapid analysis capabilities. This research focuses on three primary research questions. First, we assess the predictive capabilities of visible to near-infrared (VisINIR) for determining nutrient concentrations in macadamia kernels. Second, we examine how data proportioning, particularly the test/train split, affects model prediction accuracy, demonstrating improved accuracy with optimised data proportioning. Third, we investigate the effectiveness of smoothing functions in eliminating problematic spectra and enhancing model accuracy, finding no success in removing light-polluted samples. We utilise various machine learning algorithms, including artificial neural network (ANN), partial least squares regression (PLSR), k nearest neigh-bors (KNN) regressor, and random forest regressor. Additionally, we explored three feature selection algorithms: Relief, mean decrease impurity (MDI), and Boruta, to refine our predictive models and enhance their accuracy. The research showed that MDI achieved acceptable results for certain nutrients (e.g., Mn, Na) with fewer features compared to Relief, indicating a potential advantage in feature efficiency.