● NIRS enables rapid, non-destructive classification of 18 walnut cultivars. ● Six spectral matrix strategies simulate varying acquisition scenarios. ● Single-surface spectra also yielded >97.8% accuracy for practical applications. ● The method supports scalable, real-time walnut variety authentication. Different varieties of walnut contain varying levels of bioactive compounds, including unsaturated fatty acids and polyphenols, and it is generally believed that the higher the bioactive compound, the better the walnut quality. Accurate variety identification is therefore essential for producers, processors, and consumers. This study employed near-infrared spectroscopy (NIRS) to collect spectral data from 18 walnut varieties within the 950–1650 nm wavelength range. Spectral data were organized into matrices based on the number of sampled surfaces, from single-surface spectra to averaged signals from six surfaces. After applying five different pretreatments, seven classification models were evaluated for their performance. The result shows that the support vector machine (SVM) model combined with second order derivative (2nd) preprocessing on five-surface averaged spectra, yielding an accuracy of 99.42%. Even the single-surface strategy achieved over 97.8% accuracy, supporting the development of streamlined, low-cost, and scalable methods for real-time walnut variety classification in industrial settings.
Zhou et al. (Wed,) studied this question.