Authentication of Edible Bird's Nest Beverage-Derived Powders Using Block-Based Near-Infrared Hyperspectral Imaging and Multi-Model Agreement Mapping.

Abstract

Authentication of edible bird's nest (EBN) in heterogeneous matrices is challenging because whole-region spectral averaging may obscure localized variation. This study developed a block-based near-infrared hyperspectral imaging (NIR-HSI) framework integrating complementary classifiers and spatial multi-model agreement mapping. The development dataset comprised 90 independently prepared physical samples, each with four independently sampled and repacked scans; external validation used 21 additional physical samples (84 scans). Spectra (900-1700 nm) were represented by mean-spectrum or block-based approaches and classified using partial least squares discriminant analysis (PLS-DA), random forest (RF), and Multiple Instance Adaptive Cosine Estimator (MI-ACE). Model development used leave-one-physical-sample-out cross-validation. In external validation, mean-spectrum PLS-DA achieved accuracies of 0.7143-0.8571, whereas mean-spectrum RF correctly classified all 21 samples. For 9 × 9 blocks, PLS-DA achieved accuracies of 1.0000, 0.9048, and 1.0000 across the three preprocessing conditions, while RF and MI-ACE correctly classified all 21 samples under all conditions. All-ROI analysis identified a significant second-derivative difference at approximately 1505.65 nm between EBN-positive and EBN-free classes (U = 0, p < 0.001). This region is potentially informative but should not be considered EBN-specific. Overall, the framework provided proof-of-concept evidence of strong physical-sample-level classification while preserving spatial information.