摘要
Label-free sensing of complex botanical matrices has long posed significant challenges in analytical sensing and automated quality assurance due to the limitations of exhaustive chemical extraction and the data scarcity bottleneck inherent to modern deep learning. To overcome this challenge, this work presents a label-free, data-efficient botanical characterization framework, using terahertz time-domain spectroscopy combined with ensemble learning. Specifically, this framework bypasses the necessity of chemical extraction by profiling the macroscopic optical properties of whole botanical matrices, which are in the form of pellets prepared by a streamlined, label-free protocol. To accurately resolve the highly similar spectral features while mitigating the risk of overfitting, a soft-voting ensemble of deep neural networks is implemented, architecturally reinforced by a stratified six-fold cross-validation strategy. The framework is further rigorously evaluated in two aspects: an input feature ablation analysis is conducted to ensure the physical interpretability of the network, and its classification performance is systematically benchmarked against conventional machine learning algorithms. Finally, this complete framework is tested on a highly demanding case study, the discrimination between two closely related medicinal herbs, Panax ginseng and Panax quinquefolius. The proposed framework yields a discrimination accuracy of 97.67%, highlighting its potential for real-world phytochemical profiling of high-value botanical materials.