Assessment of chemometric and deep learning models for blueberry authentication using visible and near-infrared spectral imaging

Reliable identification of blueberry varieties is critical for quality control, supply chain traceability, and fraud prevention. This study evaluated hyperspectral imaging (HSI) in the visible–near infrared (VNIR, 400-1000 nm) and near infrared (NIR, 900-1700 nm) ranges, combined with chemometric and deep learning models, for blueberry varietal discrimination. Mean spectra were extracted and analyzed using Partial Least Squares Discriminant Analysis (PLS-DA), Linear Discriminant Analysis (LDA), Support Vector Machines (SVM), and 1D Convolutional Neural Networks (1D-CNN), and a 3D-CNN was trained on spectral–spatial patches from the full hypercubes. Full-spectrum modeling showed that 1D-CNN achieved the highest accuracy with VNIR data (97.4%), while LDA performed best with NIR (94.5%), reflecting complementary strengths of nonlinear and linear models. To reduce redundancy, the Successive Projections Algorithm (SPA) was applied under three feature level data fusion strategies: fusion-before-selection, separate-selection, and fuse-after-selection. Among these, fuse-after-selection maintained the best performance, with the 3D-CNN achieving the highest accuracy (97.5%), followed by LDA (91.0%) and SVM (90.0%). These findings demonstrate that HSI, coupled with tailored modeling strategies, provides an effective and scalable approach for blueberry varietal authentication.