Breaking Barriers to U-Net Adoption in Hyperspectral Imaging: Using PCA and Spatial Resolution for Enhanced Food Fraud Detection

Hyperspectral imaging (HSI) is highly valued for its minimal sample preparation, nondestructiveness, and rich spatial information. However, its large data volume poses computational challenges. To efficiently process HSI data and detect adulteration in saffron, a high-value product, this study employs a U-shaped deep neural network (U-Net) to retain spatial information and improve classification accuracy. The algorithm aims to determine whether saffron samples contain adulterants and to identify their precise locations. U-Net, as a supervised algorithm, requires ground-truth images for each sample with pixel-level labels. However, creating these labels is both time-consuming and challenging. A key challenge, pixel labeling or semantic segmentation, is addressed by automating the process with principal component analysis, enabling the precise localization of adulteration by classifying individual pixels. To meet the data requirements of deep learning, data augmentation was applied by extracting 100 two-dimensional images from each hypercube within the 500–600 nm wavelength range, where authentic and adulterated saffron exhibit distinct spectral characteristics. The model was evaluated on independent saffron samples, saffron packaged with poly crystals, and saffron with mixed adulterants to test its generalizability. Results were impressive: image classification achieved perfect accuracy (1.00), while pixel-level accuracy ranged from 0.93 to 1.00. The U-Net model outperformed traditional methods such as partial least-squares discriminant analysis, support vector machine, and artificial neural network in both reliability and precision. These findings underscore the power of deep learning for hyperspectral data analysis and highlight U-Net’s potential for accurate adulteration detection.