Hyperspectral imaging-based dual-task learnable spectral mask convolutional neural network for hierarchical adulteration detection in Fritillaria (Beimu) powder.

Reliable, non-destructive, and interpretable methods for detecting hierarchical adulteration in powdered herbal medicines are essential for stringent quality control. This study introduces a dual-attention Learnable Spectral Mask Convolutional Neural Network (DLSM-CNN) to simultaneously identify adulterant type and quantify their concentration in Fritillaria powder. Near-infrared hyperspectral images (945-1620 nm) were acquired, yielding a massive dataset of 39,600 independent superpixel spectral samples derived from 33 physical powder samples. These encompassed both inter-species foreign matter adulteration (starch/talc) and highly deceptive intra-genus congeneric adulteration involving four homologous species (Fritillaria pallidiflora Schrenk, Fritillaria thunbergii Miq., and two distinct commercial varieties of Fritillaria ussuriensis Maxim.). The embedded LSM module acts as a dynamic physical gateway, adaptively highlighting chemically informative wavelengths while suppressing environmental noise. Compared to traditional chemometric baselines and feature selection methods, DLSM-CNN achieved superior superpixel-level predictions. For foreign matter detection, it yielded a classification accuracy of 97%, alongside quantitative regression metrics of R2 = 0.9692 and RMSE = 5.65%. For the exceptionally complex intra-genus discrimination, the framework maintained an overall accuracy exceeding 99%. Crucially, visualizing the learned saliency maps demystified the model's "black box" nature, revealing its reliance on rigorous chemical logic-autonomously targeting specific functional groups including inorganic/organic hydroxyls (O-H), carbohydrate skeletons (C-H), and nitrogenous profiles (N-H). The proposed framework offers a robust, chemically interpretable, and high-resolution spatial mapping solution for industrial botanical drug inspection.