Abstract
Fritillaria spp. has an extremely high edible and medicinal value. Different parts of it exhibit significant variations in medicinal efficacy. To rapidly and accurately identify the origin and adulteration of Fritillaria spp., a handheld near-infrared spectrometer was combined with a convolutional neural network (CNN) to establish an efficient and convenient quality assessment method. First, for the origin of Fritillaria spp., the CNN could achieve high accuracy, with 100 ± 0%. The features contributing to the origin of Fritillaria spp. were visualized using gradient-weighted class activation mapping (Grad-CAM). For the adulteration of Fritillaria spp., compared with partial least squares regression (PLSR), the CNN yielded the best performance, with the R(2) of the test set being 0.9897. Additionally, to improve the interpretability of the adulteration model, a CNN model was established using data whose dimensions had been reduced by PCA (PCA-CNN), which also achieved an R(2) of 0.9876. The features extracted by PCA focused on 1400-1500 nm, which was consistent with Grad-CAM. The visualization of Grad-CAM and the adulteration detection model achieved mutual validation, showing the effectiveness of both methods in analyzing the samples. The experimental results demonstrated that the integration of a handheld near-infrared spectrometer with a CNN enabled both reliable authentication of Fritillaria spp. geographical origins and quantitative determination of adulteration levels, establishing a novel analytical framework for rapid quality evaluation of Fritillaria spp. materials.