Tracing the geographical origin of forest-sourced medicinal materials by hyperspectral imaging: a critical review for smart forestry
TL;DR
A structured narrative synthesis of the 2015–2026 literature indicates that deep learning architectures—particularly two-dimensional correlation spectroscopy (2DCOS) combined with residual networks—have been reported to achieve 98–100% classification accuracy for several forest-sourced species under controlled laboratory conditions.
Abstract
Forests supply a large and economically significant fraction of the medicinal materials used in traditional Chinese medicine (TCM), comprising woody medicinal taxa, understory-cultivated herbs and forest-dwelling macrofungi. For all three resource classes the therapeutic value, and hence the market price, is tied to geographical origin and to cultivation mode, yet conventional authentication remains slow, destructive and poorly suited to supply-chain deployment. Hyperspectral imaging (HSI) offers a rapid, non-destructive alternative that is well matched to the monitoring requirements of smart forestry. This review evaluates the HSI-based authentication pipeline across three methodological dimensions: spectral preprocessing, feature extraction, and classification modelling, and then maps reported performance onto the three forest resource classes. A structured narrative synthesis of the 2015–2026 literature indicates that deep learning architectures—particularly two-dimensional correlation spectroscopy (2DCOS) combined with residual networks—have been reported to achieve 98–100% classification accuracy for several forest-sourced species under controlled laboratory conditions. Within-study comparisons, in which a single dataset is modelled by several algorithms, show dataset-dependent outcomes rather than a universal deep-learning advantage over partial least squares discriminant analysis (PLS-DA) or support vector machines; values compared across different studies are confounded by differences in species, sample size, instrument and validation design and should not be read as benchmarks. In one fully verifiable within-study example, wavelength selection retained 11.2–16.1% of the original bands with zero to 1.12 percentage points of accuracy loss, while several data-fusion studies reported improvements over their own single-source baselines. These individual results do not establish universal advantages. Critically, no study has directly compared spectral signatures of identical species grown under forest canopy versus open-field conditions—a gap with significant implications for Daodi quality verification and for the certification of understory cultivation as a forest-management practice. External validation, cross-instrument transfer and the confounding effect of a changing climate on metabolite profiles remain the principal barriers to field deployment. Future priorities include standardized spectral databases, portable devices, transfer learning, and explainable AI.
