Tracing the geographical origin of forest-sourced medicinal materials by hyperspectral imaging: a critical review for smart forestry

1 Introduction

1.1 Forest-sourced medicinal materials in a smart-forestry context

A substantial share of the raw materials of traditional Chinese medicine (TCM) originates not from arable farmland but from forests. These forest-sourced medicinal materials fall into three ecologically distinct classes: woody medicinal taxa harvested for bark, wood, fruit, flower or root (for example Eucommia ulmoides, Gardenia jasminoides, Lonicera japonica, Lycium barbarum and several Dalbergia species); herbaceous species cultivated beneath a managed forest canopy, the so-called understory economy; and macrofungi that complete their life cycle on forest substrates. All three depend on stand structure, canopy light transmission and forest soil conditions, and all three are therefore legitimate objects of forest management rather than of agronomy alone. Traditional Chinese medicine has served as a cornerstone of healthcare in East Asia for over two millennia, with the therapeutic efficacy of medicinal materials intimately linked to their geographical origin—a concept formalized as Daodi in Chinese pharmacological tradition (Zhao et al., 2010). China’s 2017 Traditional Chinese Medicine Law codified this principle by defining Daodi materials as those produced in specific regions with documented superior quality and therapeutic outcomes. The economic significance of origin authentication has grown alongside national policy support for the understory economy (National Forestry and Grassland Administration, 2024). Exact area and output-value estimates are not reproduced here because the available yearbook citation does not identify a verifiable edition and page for those figures. Within this landscape, understory-cultivated TCM species—including Panax notoginseng, Gastrodia elata, Dendrobium officinale, Ganoderma lucidum, Fritillaria spp., and Coptis chinensis—represent a substantial share of high-value medicinal production, with individual species commanding price premiums of 3- to 10-fold depending on origin certification (Liu et al., 2023).

Understory cultivation is increasingly promoted as a mechanism for reconciling economic return with forest conservation, because it generates income from standing forest without conversion of land cover (Li, X. et al., 2025). Its expansion creates a management problem that is squarely within the remit of smart forestry: verifying, at scale and non-destructively, that a given batch of medicinal material was produced in the declared region and under the declared cultivation mode. Conventional forest inventory and remote-sensing tools describe the stand; they do not authenticate the harvested product. Conversely, laboratory authentication methods characterize the product but are blind to the stand. Spectral authentication of harvested material is the point at which these two information streams can be joined.

1.2 Authentication challenges and the promise of hyperspectral imaging

The confluence of high market value and complex supply chains has created persistent challenges in TCM quality control. Adulteration through origin mislabeling, substitution of cheaper varieties, and blending of materials from multiple sources undermines both consumer trust and therapeutic reliability (Palhares et al., 2015; Simmler et al., 2018). Conventional authentication methods—including morphological examination, DNA barcoding, and chromatographic analysis (HPLC, GC–MS)—while effective, suffer from limitations that restrict their deployment at scale: they are time-consuming (hours to days per sample), often destructive, require expensive reagents and skilled operators, and are poorly suited to high-throughput screening scenarios such as wholesale market inspection or production line quality control (Liang et al., 2004; Gao et al., 2025; Jiang et al., 2025).

Hyperspectral imaging (HSI) technology has gained substantial traction as a non-destructive, rapid alternative for TCM authentication over the past decade (Pan et al., 2024; You et al., 2025). By acquiring spectral information across hundreds of contiguous narrow wavelength bands—typically spanning the visible (400–700 nm), near-infrared (700–1,000 nm), and short-wave infrared (1000–2,500 nm) regions—HSI captures a comprehensive chemical fingerprint that encodes information about pigments, moisture, proteins, lipids, carbohydrates, and secondary metabolites within each pixel of the imaged sample (Gowen et al., 2007; Lu et al., 2020). When coupled with appropriate chemometric or machine learning algorithms, this rich spectral information enables geographical discrimination with reported accuracies that frequently exceed 95% for major TCM species under laboratory conditions (Pan et al., 2024; Medina-García et al., 2025).

The understory cultivation environment introduces a distinctive dimension to spectral authentication that has received insufficient attention. Forest canopy cover modulates the light environment, temperature regime, humidity, and soil microbiome experienced by understory herbs, directly influencing their biosynthesis of secondary metabolites—the very compounds that drive spectral signatures in the NIR and MIR regions (Zubay et al., 2021). Research has demonstrated that shade-tolerant species such as C. chinensis accumulate more berberine under moderate canopy cover, while P. polyphylla produces optimal steroid saponins at montane understory temperatures of 15–20 °C (Li, X. et al., 2025). A 2024 review confirmed that wild-tending and imitation-wild cultivation modes—the core understory approaches—yield chemical profiles closest to wild specimens and significantly different from open-field cultivated materials (Wang et al., 2024). Yet no study has directly compared the spectral signatures of understory-grown versus open-field materials of the same species, representing a critical research gap, and one that is revisited at the end of every methodological section of this review.

1.3 Climate change as a driver of, and a challenge to spectral authentication

Two features of a warming climate bear directly on the authentication problem addressed here. The first is spatial: the climatic envelopes that define Daodi production regions are not stationary. Suitable cultivation zones for montane and understory medicinal species are shifting upward in elevation and poleward in latitude, so that the set of localities legitimately able to claim a given geographical indication may change within the working life of a calibration model. The second is physiological: drought, heat and altered canopy structure modify the biosynthesis of precisely those secondary metabolites that give a material its spectral identity. Correlations between climate variables and marker-compound content are already documented for forest-sourced species—annual mean temperature, sunshine duration and precipitation covary with ginsenoside content across P. notoginseng production regions in Yunnan (Liu et al., 2023), and dry matter content in G. elata varies with precipitation, altitude, temperature and soil pH (Ma et al., 2023).

The consequence for spectral authentication is a form of concept drift. A model trained on material harvested under one climatic regime encodes a chemotype that the same locality may no longer produce a decade later; the accuracy losses reported for cross-year transfer are therefore not purely instrumental but partly ecological in origin. Climate change also alters the reference itself, because canopy disturbance—drought-induced dieback, insect outbreak, windthrow—changes the light environment under which understory material is produced. This review does not attempt a comprehensive treatment of climate impacts on forest ecophysiology, which lies outside its methodological scope; it does, however, treat climate-induced chemotypic drift as a first-order constraint on model longevity, and returns to it when considering database design, transfer learning and the research agenda (Sections 8.2, 9.2 and 9.3).

1.4 Aim and organization of this review

Several reviews have addressed aspects of this research landscape. Pan et al. (2024) provided a broad overview of HSI combined with machine learning for TCM quality control, while You et al. (2025) focused on texture information extraction from hyperspectral images of Chinese herbal medicines. Medina-García et al. (2025) critically reviewed HSI data-analysis strategies for food quality and safety over the preceding 5 years, and Miao et al. (2025) surveyed the transition from classical chemometrics to machine learning in VIS/NIR and HSI applications; Jiang et al. (2025) compared HSI, near-infrared spectroscopy, terahertz spectroscopy and DNA barcoding for food-medicine homologous herbal materials. Wang et al. (2024) reviewed novel spectral imaging techniques for TCM but remained shallow on HSI-specific geographical authentication. Liang et al. (2004) focused exclusively on NIR spectroscopy, predating the rapid advances in deep learning that have transformed the field since 2020. None of these reviews adopted a forest-resource perspective, distinguished woody from herbaceous and fungal taxa, or systematically evaluated the methodological pipeline—from preprocessing through feature selection to classification—specifically in the context of forest-sourced and understory TCM authentication, nor did they address the unique spectral implications of forest-grown versus open-field cultivation.

Against this background, the present review traces the full analytical pipeline of HSI-based geographical authentication as it applies to forest-sourced TCM species. Section 2 defines the scope, sets out the three-class resource framework used throughout, and describes how the literature was selected. Section 3 sets out the physical fundamentals. Spectral preprocessing follows, where method choice proves model-, species- and morphology-dependent (Section 4), then feature extraction and wavelength selection (Section 5). Classification modelling is evaluated across three generations of approaches, from PLS-DA through SVM and Random Forest to deep learning (Section 6), with an explicit separation of within-study from cross-study evidence (Section 6.5). These threads converge in a resource-class and species-by-species synthesis (Section 7). The review then maps the emerging frontiers—portable devices, transfer learning, explainable AI (Section 8)—and closes with the principal challenges, the case for community spectral infrastructure, and a research agenda for climate-resilient authentication within smart forestry (Sections 9 and 10). Each methodological section ends with a short statement of its unresolved questions, so that the critical appraisal is distributed through the review rather than deferred to its conclusion.

2 Scope, definitions and literature selection

2.1 Defining forest-sourced medicinal materials

The medicinal materials considered in this review are those whose production depends on a forest stand. Three classes are distinguished, and the distinction is maintained throughout Section 7 because the three differ in the tissues sampled, in the environmental drivers of their chemistry, and in the authentication question that matters commercially.

Woody medicinal taxa are trees, shrubs and woody lianas whose bark, wood, root, flower or fruit constitutes the drug. Members of this class are subject to conventional silvicultural management, and their chemistry responds to stand density, rotation age and site quality. Examples appearing in the spectral authentication literature include Dalbergia spp. (heartwood), Gardenia jasminoides (fruit), Lonicera japonica (flower bud) and Lycium barbarum (fruit).

Understory-cultivated herbaceous taxa are perennial herbs grown beneath a retained canopy under wild-tending, imitation-wild or managed understory regimes. The canopy is not incidental to production but is the production system: light transmission, throughfall and forest soil microbiota define the chemotype. This class contains the highest-value species in the Chinese market, including P. notoginseng, G. elata, D. officinale, Paris polyphylla, C. chinensis, Fritillaria spp. and Anoectochilus roxburghii.

Forest-dwelling macrofungi complete their life cycle on woody substrates or on the roots of forest trees. Ganoderma lucidum, cultivated on hardwood logs or sawdust substrate, is the principal example in the spectral literature. Fungi are not plants, and their spectral response is dominated by polysaccharides, chitin and triterpenoids rather than by chlorophyll or cellulose; they are retained here because their production is a forest-based industry and because their authentication problem—species substitution combined with origin mislabeling—is formally identical to that of the plant classes.

Materials that are grown predominantly as field crops on arable land, and whose chemistry is not conditioned by a forest canopy, fall outside this scope and are cited only where a study provides a methodological result of general relevance. Where such a study is cited, the taxon is identified so that the reader can judge transferability. Gastrodia elata occupies an intermediate position and is treated as understory: it is a mycoheterotroph without chlorophyll, dependent on Armillaria and hence on woody substrate, and is cultivated almost exclusively in forest settings.

2.2 Relevance to smart forestry

Smart forestry seeks to instrument the forest and its products with sensing, data infrastructure and decision support. Spectral authentication of forest-sourced medicinal materials contributes to that agenda in three ways. First, it provides a non-destructive, sensor-based measurement of a forest product, closing an information gap between stand-level remote sensing and product-level quality control. Second, if the spectral consequences of canopy cover can be characterized—the central research gap identified in this review—then spectral measurement of harvested material becomes an indirect, retrospective measurement of the growing environment, and therefore a verification tool for silvicultural practice and for payments tied to it. Third, portable and conveyor-mounted instruments of the kind reviewed in Section 8.1 are compatible with deployment at forest-farm gate, cooperative or first-processing plant, which is where value is captured and where mislabeling most often enters the chain.

2.3 Literature search and selection strategy

This article is a structured narrative review rather than a systematic review conducted under PRISMA. No claim is made that the search was exhaustive, and no meta-analytic pooling of effect sizes was attempted; the numbers reported in Tables 1, 2 are the values stated by the original authors and are not directly comparable across studies (Section 6.5). The search strategy and selection rules are set out here so that the basis of the synthesis is transparent and reproducible.

Table 1. Forest-resource characterization of the focal medicinal taxa, with the authentication challenge and best-supported spectral approach for each.

Resource classes are defined in Section 2.1. “Strength of validation” refers to the strongest validation reported in the cited studies for that species: internal test partition denotes a held-out split of a single acquisition campaign, whereas external validation denotes evaluation on material described by the source as a separately designated external set. Two studies in Table 2 report an external test set. Accuracies are given in Table 2 and are not comparable across rows (Section 6.5).

Table 2. Summary of HSI and spectral methods reported for geographical authentication of forest-sourced medicinal species. Numeric citation keys used in the previous version have been replaced by author–year citations.

Accuracies are the values reported by the original authors and were obtained under controlled laboratory conditions on datasets differing in species, number of origin classes, sample size, instrument, spectral range and validation design. They are therefore not directly comparable across rows and do not constitute benchmarks; see Section 6.5. Values of 100.0 indicate that the classification task was saturated under the reported validation scheme, not that the method is exact. Where a study reported both internal and external validation, both values are given.

Web of Science Core Collection, Scopus, PubMed and CNKI were searched for records published between January 2015 and March 2026; during revision, targeted searches for reviewer-requested topics and publications were updated through 6 August 2026, using combinations of the terms “hyperspectral,” “near-infrared,” “spectroscopy,” “spectral imaging,” “traditional Chinese medicine”, “medicinal plant”, “herbal”, “geographical origin”, “provenance”, “authentication” and “traceability”, each combined with the Latin binomials of the focal taxa. Reference lists of the retrieved reviews were screened manually for additional records, and the citation network of the most frequently cited primary studies was followed forward.

Records were retained when they satisfied all of the following: (i) the study reported original spectral measurements of a medicinal material, or was a review of such measurements; (ii) the analytical target was geographical origin, cultivation mode, species authenticity or adulteration; (iii) sufficient methodological detail was reported to identify the spectral range, the preprocessing applied, the modelling algorithm and the validation scheme; and (iv) the article was published in English or Chinese in a peer-reviewed venue. Records were excluded when the spectral measurement served only to quantify a single analytic with no classification objective, when the material was a finished formulation rather than a crude drug, when validation was performed on the calibration set alone with no independent test partition, or when the report was a conference abstract without a full methods section. Methodological studies on non-forest taxa were admitted only where they established a preprocessing, feature-selection or modelling principle of general applicability, and are flagged as such in the text.

Retained studies were then classified along four axes for synthesis: resource class (woody, understory herbaceous, forest-dwelling fungal, or non-forest methodological reference); pipeline stage addressed (acquisition, preprocessing, feature selection, classification, fusion, deployment); modelling paradigm (chemometric, classical machine learning, deep learning); and strength of validation (internal cross-validation only, independent test partition, or external validation on a separately acquired batch, instrument or year). The last axis governs the weight placed on reported accuracies throughout Sections 6 and 7.

3 Fundamentals of HSI technology for authentication of forest-sourced medicinal materials

Hyperspectral imaging integrates conventional imaging with spectroscopy by recording the spectral reflectance or transmittance of a sample across a continuous range of wavelengths, producing a three-dimensional data cube (x, y, λ) where each spatial pixel contains a complete spectrum (Gowen et al., 2007). Acquisition modes in this field are principally line-scanning (push-broom) for laboratory imaging, area-scanning for static samples, and point-scanning for bulk material. The relevant spectral ranges span the visible–near infrared (VNIR, 400–1,000 nm), short-wave infrared (SWIR, 1000–2,500 nm), and mid-infrared (MIR, 2,500–25,000 nm) regions, each encoding distinct molecular information (ElMasry and Sun, 2010).

The physical basis for geographical discrimination lies in the relationship between a material’s chemical composition and its spectral response. VNIR absorption arises from electronic transitions in pigments and from vibrational overtones, with the red edge (680–750 nm) repeatedly identified as discriminative (Liu et al., 2023). SWIR carries overtone and combination bands of O–H, C–H, N–H and C=O, reporting moisture, lipids, proteins and carbohydrates (Rinnan et al., 2009). The MIR fingerprint region resolves fundamental vibrations with high specificity but suffers stronger water absorption and lower signal-to-noise in biological matrices (Li et al., 2018).

Environmental factors associated with geographical origin—including soil composition, altitude, temperature, precipitation, and critically for understory species, canopy cover—modulate the biosynthesis of secondary metabolites in TCM plants. Liu et al. (2023) demonstrated significant correlations between climate factors (annual mean temperature, annual sunshine duration, annual precipitation) and ginsenoside content (Rb1, Rg1, Rd., and notoginsenoside R1) across five production regions of P. notoginseng in Yunnan Province. These chemotypic differences, while sometimes subtle in individual spectral bands, create distinguishable multivariate patterns that can be captured and classified by appropriate analytical pipelines. It should be emphasized that this chain of inference—environment shapes metabolite profile, metabolite profile shapes spectrum, spectrum permits classification—is an indirect one. Classification accuracy is therefore not by itself evidence that a model has learned a chemically meaningful, environmentally driven signal rather than a batch- or instrument-specific artefact, a distinction taken up in Sections 6.5 and 8.3. The general workflow of HSI-based forest-sourced TCM geographical authentication—from sample collection through spectral acquisition, preprocessing, feature extraction, and classification—is illustrated in Figure 1.

4 Spectral preprocessing strategies

4.1 Rationale and common methods

Raw spectral data invariably contain systematic artifacts introduced by scattering effects, baseline drift, instrumental noise, and sample heterogeneity—all of which can obscure the chemical information essential for geographical discrimination (Rinnan et al., 2009; Mishra et al., 2020). Preprocessing serves to attenuate these non-informative spectral variations while preserving or enhancing chemically relevant features. The choice of preprocessing method is not trivial: inappropriate preprocessing can degrade model performance by removing genuine spectral differences between origins or amplifying noise in uninformative regions. This section evaluates the most widely applied preprocessing techniques in TCM spectral authentication and their demonstrated impacts on classification outcomes.

Standard Normal Variate (SNV) transformation centres each spectrum to zero mean and unit variance, correcting the multiplicative scatter and path-length variation that arise from differences in particle size and packing density (Barnes et al., 1989). Multiplicative Scatter Correction (MSC) removes the same components by regressing each spectrum against a mean reference (Geladi et al., 1985). The two typically perform comparably in this literature, with SNV slightly favoured for its independence from a reference spectrum (Yang et al., 2022).

Savitzky–Golay (SG) smoothing fits a polynomial by least squares within a moving window, reducing noise while preserving spectral shape (Savitzky and Golay, 1964); wider windows smooth more strongly but risk attenuating narrow metabolite-specific features. Derivative transformations resolve overlapping bands and remove baseline offsets (Norris and Williams, 1984), and are commonly paired with smoothing to offset their noise amplification. Both first- and second-derivative transformations have improved individual classification models, but their effects depend on the sample, spectral range, model and validation design and should not be generalized across studies (Li et al., 2018; Ma et al., 2023).

4.2 Species-, morphology-, and model-specific preprocessing effects

A pivotal finding emerging from the recent literature is that optimal preprocessing is both model-dependent and species-specific, challenging the common practice of applying a default SNV or first-derivative correction across all datasets. Chen et al. (2024) demonstrated this clearly in a Dalbergia—a woody genus—species identification study: PCA-based classification accuracy actually decreased with SNV and SG preprocessing, PLS-DA benefited from scatter correction that addressed multicollinearity, SVM gained from SG smoothing that clarified class boundaries, while CNN performance was enhanced by MSC which preserved spectral integrity while reducing noise. Under optimal preprocessing conditions, both SVM and CNN reached 100% accuracy on the independent test partition in NIR and full spectral ranges—a result attributable to preprocessing rather than algorithm sophistication alone, and one obtained on a closed set of 10 species under uniform laboratory acquisition.

Sample morphology is a third and frequently overlooked determinant. Dai et al. (2026) compared NIR-HSI provenance classification of Curcumae Radix presented as intact root slices and as milled powder under nested cross-validation, and reported morphology-dependent preprocessing outcomes. The accessible bibliographic record and abstract-level indexing did not provide the complete numerical results needed to reproduce the exact accuracy values independently, so those values are not repeated here. The practical implication remains that a preprocessing protocol validated on powdered material cannot be assumed to transfer unchanged to intact drug.

Hybrid preprocessing combinations can outperform simpler transformations within a given dataset, but the gain must be evaluated within the same model and validation design. For G. elata origin classification, Ma et al. (2023) reported that, without data augmentation, normalization alone gave a 1D-CNN test accuracy of 91.67%, whereas a second-order derivative followed by normalization gave 100% across 11 origins. The same source reported corresponding test-set changes from 91.67 to 95.56% for PLS-DA, from 51.67 to 96.67% for KNN, and from 65.00 to 98.33% for SVM, making this a directly verifiable within-study comparison across four models. Cui et al. (2025) evaluated NIR preprocessing and machine-learning classifiers for Gardeniae Fructus, the fruit of the woody shrub Gardenia jasminoides; exact performance values are not repeated because the full results were not accessible for independent audit. Wavelet transform-based denoising has been evaluated for spectra acquired with portable devices, where signal-to-noise ratios may be lower than with laboratory instruments (Wu et al., 2024). The 2DCOS approach converts one-dimensional spectra into two-dimensional correlation images and therefore combines preprocessing with feature engineering; the limited available evidence suggests that it may reduce reliance on conventional preprocessing, but this has not been established across independent datasets (Liu et al., 2023; Zuo et al., 2025). The comparative effects of these preprocessing strategies on model performance and on representative spectra are illustrated in Figure 2.

4.3 Limitations and open questions

Three issues remain unresolved. First, no consensus protocol exists for selecting a preprocessing pipeline a priori. The prevailing practice—exhaustive search over a grid of preprocessing options with the test-set accuracy as the criterion—risks optimistic bias, because the preprocessing choice is then itself a fitted hyperparameter that has seen the evaluation data; nested cross-validation of the kind applied by Dai et al. (2026) is the exception rather than the rule. Second, the interaction between preprocessing and sample morphology is documented for one genus only, and it is not known whether the intact-versus-powder gap is a general phenomenon or specific to root drugs with pronounced anatomical structure. Third, the claim that 2DCOS and related correlation transforms make conventional preprocessing unnecessary rests on a small number of studies from a small number of groups, and has not been tested on field-acquired spectra where the scatter and illumination variability are considerably larger than in the laboratory. Whether preprocessing choices that are optimal under controlled illumination remain optimal under the variable geometry of a handheld or conveyor-mounted instrument is, at present, unknown.

5 Feature extraction and wavelength selection

5.1 The dimensionality challenge

A typical HSI dataset for TCM authentication contains 200–500 spectral bands per pixel, creating a high-dimensional feature space that presents two interrelated challenges: computational burden and the “curse of dimensionality,” whereby classification performance can paradoxically degrade as the number of features increases relative to the number of training samples (Hughes, 1968). The problem is aggravated in this application domain because the sample sizes attainable for forest-sourced materials are small: collection from multiple remote montane origins is expensive, and a study with several hundred spectra spread over 10 origin classes may have fewer samples per class than it has retained wavelengths. Medina-García et al. (2025) argue, in a critical review of 5 years of HSI data analysis for food quality and safety, that the decisions taken at this stage—region-of-interest definition, signal extraction from the cube, and dimensionality reduction—exert more influence on final model performance than the choice of classifier itself, and are reported with markedly less consistency. Effective wavelength selection or feature extraction—reducing the spectral dimensionality while retaining or enhancing discriminative information—is therefore a critical step in the analytical pipeline. The approaches available can be broadly categorized into unsupervised feature extraction (e.g., PCA), supervised wavelength selection (e.g., CARS, SPA, GA), and learned feature extraction through deep neural network architectures.

5.2 Classical wavelength selection methods

Principal Component Analysis (PCA) projects spectral data onto orthogonal axes of maximum variance (Wold et al., 1987) and typically retains 95–99% of total variance in 5–15 components, but as an unsupervised method it does not optimize for class separation and is more often used for visual assessment of separability or as a step preceding supervised classification (Liu et al., 2023).

Competitive Adaptive Reweighted Sampling (CARS) has emerged as the benchmark supervised wavelength selection method across the TCM spectral literature (Li et al., 2009). CARS selects wavelength subsets iteratively, assigning higher selection probability to wavelengths with larger absolute PLS regression coefficients and progressively eliminating uninformative ones across Monte Carlo runs; Yin et al. (2019) recommended it, with SPA and genetic algorithms, as benchmark methods. The consistent finding across multiple species—including P. notoginseng (Liu et al., 2023), D. officinale (Yang et al., 2022), and Fritillaria spp. (Wu et al., 2024)—is that CARS achieves an effective balance between aggressive dimensionality reduction (typically retaining 10–30 wavelengths from 200 + original bands) and preservation of classification accuracy.

The Successive Projections Algorithm (SPA) selects wavelengths by maximizing the projection of each candidate variable onto the orthogonal subspace of previously selected variables, effectively minimizing collinearity among selected features (Araújo et al., 2001). SPA typically produces smaller wavelength subsets than CARS (5–15 variables) but may miss subtle spectral features that contribute to class separation through synergistic interactions. Genetic Algorithm (GA) approaches treat selection as combinatorial optimization, evolving wavelength subsets under a fitness function; they are more expensive than CARS or SPA but can find non-obvious combinations exploiting inter-band correlation (Leardi and Nørgaard, 2004). Interval PLS (iPLS) offers a complementary, interval-based alternative in which the spectrum is partitioned into contiguous windows and the most informative windows retained, a formulation that preserves the physical contiguity of absorption bands at the cost of fixed interval boundaries (Nørgaard et al., 2000).

The effect of wavelength selection is dataset- and model-dependent rather than consistently favorable. In a single Atractylodes macrocephala dataset containing 224 wavelengths, Li, P. et al. (2025) reported that CARS retained 36 wavelengths (16.1%) and matched the full-spectrum BLS accuracy of 95.56%, whereas SPA retained 28 wavelengths (12.5%) and achieved 94.44%. The same study’s post-hoc SHAP ranking retained 25 wavelengths (11.2%) and also achieved 94.44%. For D. officinale, Yang et al. (2022) reported that nonlinear SVC reached 100% recognition in calibration and prediction sets, while CARS was the best wavelength-selection method for quantitative constituent models; the available report does not establish a retained-feature ratio for the origin classifier. These results support compact, interpretable subsets in some settings but do not establish a general performance advantage over full-spectrum models.

5.3 Learned feature extraction and non-targeted anomaly detection

Deep learning offers an alternative in which feature extraction is learned end-to-end rather than prescribed. Convolutional neural networks applied to one-dimensional spectra learn hierarchical features through successive convolutional and pooling layers, bypassing explicit wavelength selection, while attention mechanisms assign adaptive weights across spectral regions and thereby emphasize discriminative wavelengths and suppress uninformative ones.

Zhang, Q.-Y. et al. (2025) developed MT-NIRNet, a multi-task deep learning framework for Polygonatum origin and variety classification that incorporates squeeze-and-excitation (SE) attention mechanisms. This architecture achieved 100% classification accuracy across 9 geographical origins and 4 varieties simultaneously, demonstrating that learned attention weights can match or exceed the performance of manual wavelength selection on that dataset. For Fritillaria species discrimination, Hu et al. (2023) applied an attention-enhanced 1D-CNN to HySpex VNIR and SWIR data, achieving 98.97 and 99.35% accuracy respectively, with the attention mechanism revealing that the model focused on spectral regions corresponding to starch (C–H stretching near 1700 nm) and alkaloid (N–H bending near 1,500 nm) absorption features—providing both high accuracy and a degree of chemical interpretability.

Explainable AI methods offer a complementary path toward identifying important spectral features post hoc. Li, P. et al. (2025) demonstrated that the top 25 wavelengths identified by SHapley Additive exPlanations (SHAP) analysis of a Broad Learning System (BLS) model for Atractylodes macrocephala authentication achieved 94.44% accuracy with minimal performance loss compared to full-spectrum models—suggesting that XAI-based feature importance ranking can serve as an effective and interpretable alternative to traditional wavelength selection methods.

A conceptually distinct approach abandons the supervised, closed-set formulation altogether. Discriminative models assign every input to one of the classes seen in training and are therefore structurally unable to flag a material that belongs to no known class—precisely the situation created by a novel adulterant or an unanticipated congeneric substitute. Huang et al. (2026) addressed this by training a deep spectral autoencoder on the hyperspectral signatures of authentic Fritillaria alone and using the reconstruction error as an anomaly score, so that adulterants and congeneric species are detected as departures from the learned manifold of authenticity rather than as members of a pre-enumerated fraud class. The appeal for regulatory screening is evident: a one-class model requires no examples of the fraud it is meant to catch. The corresponding weakness is that the decision threshold must be calibrated on authentic material of sufficient diversity, and that natural variability arising from harvest year, cultivation mode or climate may itself be scored as anomalous. This paradigm is at an early stage in TCM authentication and has not yet been evaluated on forest-sourced materials with the seasonal and canopy-driven variability characteristic of understory production.

The spectral regions most frequently reported as informative for TCM geographical discrimination are summarized schematically in Figure 3B. A systematic comparison of these wavelength selection methods is presented in Table 3.

Table 3. Comparison of wavelength selection methods for geographical authentication of forest-sourced medicinal materials. In-text numeric keys used in the previous version have been replaced by author–year citations to ensure traceability against the reference list.

CARS, competitive adaptive reweighted sampling; GA, genetic algorithm; iPLS, interval partial least squares; PCA, principal component analysis; PLS, partial least squares; SHAP, Shapley additive explanations; SPA, successive projections algorithm. Typical wavelength counts are indicative ranges compiled from the cited studies and are not benchmarks; the number retained depends on the original band count, the number of classes and the classifier.

5.4 Limitations and open questions

The comparative literature on wavelength selection is dominated by single-dataset comparisons in which several methods are applied to one set of spectra and the winner reported. Because the ranking of methods varies with species, spectral range and classifier, such results establish local rather than general superiority, and the frequently repeated statement that “CARS is the best method” is better read as “CARS is the most frequently reported method that performs adequately”. Two questions in particular are open. The first concerns stability: wavelength subsets selected by stochastic procedures such as CARS and GA vary between runs and between resampled training sets, yet selection stability is seldom quantified, and a chemically meaningful band should be recoverable across replicates. The second concerns the relation between selected wavelengths and known absorption features. Where attention weights or SHAP values coincide with documented metabolite bands the interpretation is reassuring, but coincidence has generally been asserted post hoc rather than tested, and no study has yet confirmed a selected band by independent spectroscopic assignment. Section 8.3 returns to this point.

6 Classification and modelling approaches

6.1 Traditional chemometric methods and the transition to machine learning

Partial Least Squares Discriminant Analysis (PLS-DA) has served as the default classification method for spectral authentication of TCM materials since the early 2000s (Barker and Rayens, 2003; Brereton and Lloyd, 2014). PLS-DA models the relationship between spectral data (X) and class membership (Y) through latent variable projection, handling the collinearity inherent in spectral data by constructing orthogonal components that maximize covariance between X and Y. For TCM origin classification, PLS-DA achieves typical accuracies of 80–92% depending on the species, number of origin classes, and spectral range (Pan et al., 2024). Its principal advantages are model transparency (latent variable loadings reveal which spectral regions drive classification), minimal hyperparameter tuning, and an established position in formal standards: PLS-DA is the method specified in SN/T 5727.2–2025, an entry–exit inspection and quarantine industry standard of the People’s Republic of China that applies NIR spectroscopy to P. notoginseng. It should be noted that SN/T standards are issued for import and export inspection and are not pharmacopoeial texts; no major pharmacopoeia currently recognizes any spectral method, chemometric or otherwise, as an official method for geographical origin determination (Section 9).

Soft Independent Modeling of Class Analogy (SIMCA) and Linear Discriminant Analysis (LDA) represent complementary chemometric approaches. SIMCA builds separate PCA models for each class and classifies new samples based on residual distance, providing a natural mechanism for detecting outliers and unknown adulterants (Wold, 1976). LDA seeks linear combinations of features that maximize between-class variance relative to within-class variance. Both methods perform well for two-class or few-class problems but show degraded performance when the number of geographical origins exceeds 5–6, particularly when inter-class spectral differences are subtle (Fu et al., 2017).

The displacement of these methods by machine learning was gradual and is not complete. Miao et al. (2025), reviewing VIS/NIR spectroscopy and HSI for staple food quality and safety, characterize the transition as driven less by any single algorithmic breakthrough than by the growth of dataset sizes and by the availability of implementations that made non-linear models routine. The same account applies to TCM authentication, with the qualification that dataset sizes in this field have grown far less than in food science, so that the conditions under which machine learning displaces chemometrics are met less often.

6.2 Machine learning methods

Support Vector Machines (SVM) with radial basis function (RBF) kernels have performed competitively in several TCM spectral-classification studies, particularly when combined with task-specific wavelength selection. SVM’s ability to construct nonlinear decision boundaries in high-dimensional spaces makes it well suited to separating spectrally similar origin classes (Cortes and Vapnik, 1995). For D. officinale, Yang et al. (2022) reported 100% recognition with nonlinear SVC; the available report does not establish that the origin classifier used a CARS-selected subset, while for Ganoderma lucidum, Fu et al. (2017) reported 93.17% accuracy across 25 groups (7 species × 15 producing areas) using SVM with an information-criterion-based one-versus-one strategy. SVM is therefore a useful comparator for small and medium datasets, although its apparent advantage depends on preprocessing, tuning and the validation partition.

Random Forest (RF) provides complementary strengths through its ensemble learning framework and built-in variable importance metrics (Breiman, 2001). For P. notoginseng geographical traceability, Li et al. (2018) showed that RF classification with mid-level and high-level data fusion of FT-MIR and NIR spectra outperformed single-spectrum models across five Yunnan production regions. The variable importance rankings generated by RF can identify discriminative wavelength ranges without requiring a separate feature selection step, offering interpretability that complements the accuracy of the ensemble. K-Nearest Neighbors (KNN) and gradient-boosted tree ensembles such as XGBoost (Chen and Guestrin, 2016) have also been applied, the latter increasingly favored for its regularized boosting formulation, which handles sparse inputs and feature interactions efficiently.

6.3 Deep learning architectures

Deep learning methods have become increasingly prominent in TCM spectral classification since 2020, and in direct comparisons on a single dataset they frequently, though not invariably, outperform traditional chemometrics; the size of that advantage is examined in Section 6.5. The most significant architectures for spectral data include one-dimensional convolutional neural networks (1D-CNNs), residual networks (ResNet), long short-term memory networks (LSTM), and increasingly, transformer-based models.

1D-CNNs treat spectra as one-dimensional sequences and have been reported to reach ceiling-level accuracy on datasets with more than 10 origin classes (Ma et al., 2023). LSTM networks, designed for long-range sequential dependence, suit spectra in which distant wavelength regions carry correlated chemical information. Cui et al. (2025) compared multiple machine-learning algorithms for NIR-based Gardeniae Fructus authentication; because the complete numerical results were not accessible for independent audit, this review does not use that study to quantify an algorithmic performance margin.

Transformer architectures, which replace recurrence and convolution with self-attention over the input sequence, have begun to enter the spectral domain. Ding et al. (2026) reported what appears to be the first application of a transformer-based hybrid to hyperspectral analysis of a medicinal material, using a ResLocalformer that combines local attention windows with residual paths to classify and quantify sodium metabisulfite in goji berry powder from 900 to 1,700 nm HSI data, reaching 97.22% classification accuracy over nine concentration levels. The relevance to geographical authentication is architectural rather than substantive: self-attention offers a principled mechanism for modelling interactions between non-adjacent absorption bands, which is exactly the structure that correlation-spectroscopy transforms attempt to expose by other means. Whether the substantial parameter counts of transformer models can be supported by the sample sizes available in origin-authentication studies is not yet established.

The 2DCOS-ResNet pipeline represents a prominent approach in this literature. Two-dimensional correlation spectroscopy (2DCOS) converts one-dimensional spectra into correlation images encoding synchronous and asynchronous relationships invisible in raw spectra (Noda, 1993; Liu et al., 2022), which are then processed through ResNet architectures whose skip connections permit deeper feature hierarchies without gradient degradation. In studies of G. elata and P. notoginseng, 2DCOS/3DCOS-ResNet pipelines have reached ceiling-level classification accuracy on internal test partitions (Li, G. et al., 2025; Zuo et al., 2025). The extension to three-dimensional correlation spectroscopy (3DCOS), when paired with residual network architectures, has also been reported to maintain ceiling-level classification without conventional preprocessing (Zuo et al., 2025). Because these reports cover few species and come from a small number of research groups, independent replication remains necessary before the pipeline can be regarded as an established benchmark.

Despite these advances, deep learning is not universally superior. Model choice must be established empirically for the sample size, class structure and validation design at hand rather than by a fixed sample-count threshold. Xu et al. (2025) provide a direct comparison for Anoectochilus roxburghii, an understory orchid, and its counterfeits under explicitly small-sample conditions. SVM reached 100% accuracy, and a CNN matched rather than exceeded it when spectra from both leaf surfaces were fused in a multi-view architecture. This result shows that richer input representation can matter as much as additional model capacity. Yang et al. (2024) used a multi-task deep network to combine geographical-origin classification with total-ginsenoside prediction for American ginseng (Panax quinquefolius), illustrating a task formulation that encourages the network to encode chemically relevant information alongside the class label.

Taken together, the evidence supports dataset-specific comparison rather than a universal model-selection rule. Deep learning can exploit learned representations, whereas SVM remains a computationally efficient comparator for modest datasets and PLS-DA provides a transparent baseline. PLS-DA is also the method specified in the one inspection standard identified in this review. The supported strengths, limitations and validation evidence for these three paradigms are summarized qualitatively in Figure 4B.

6.4 Data fusion strategies

Data fusion has provided improvements over single-range baselines in several studies, but the direction and magnitude are dataset-specific. Three levels are distinguished (Borràs et al., 2015): data-level fusion, concatenating raw spectral matrices; feature-level fusion, combining extracted features or latent variables; and decision-level fusion, aggregating predictions from independent models.

Feature-level fusion has outperformed single-range baselines in reported within-study comparisons, but the corpus does not establish a universal hierarchy among fusion levels. For P. notoginseng, Li et al. (2018) demonstrated that mid-level and high-level fusion of FT-MIR and NIR spectra with Random Forest classification outperformed single-spectrum models for five Yunnan production regions. For C. chinensis (Weilian), a 2025 study employing dual-lens HSI (VNIR + SWIR) with feature-level fusion and CNN modelling achieved 98.10% accuracy across 10 geographical regions, compared to 95.26% for single-range approaches (Yuan et al., 2025). Jiang et al. (2023) achieved 100% accuracy for Ganoderma lucidum adulteration detection through mid-level fusion of NIRS and HSI data.

The emerging frontier is attention-based cross-modal fusion. A multimodal CNN with cross-attention for wolfberry origin classification achieved 99.88% test accuracy (Li, B. et al., 2025), far exceeding single-modal CNN performance of 87.1%. For Pinellia ternata authentication, spectral-texture multimodal fusion with DD-SIMCA achieved 99.87% versus 92.47% for spectral data alone—a 7.4 percentage-point improvement that demonstrates the substantial complementary value of spatial texture information encoded in hyperspectral images (Zhang, H. et al., 2025).

All of the above combine spectral ranges, or spectral with spatial information, within a single measurement modality. A more ambitious form of fusion joins measurements that probe different physics. Zhang et al. (2026) fused laser-induced breakdown spectroscopy (LIBS) with HSI for the geographical tracing of ginseng under small-sample conditions, compressing both streams by PCA to 20 components retaining more than 90% of variance before fusion. The two modalities are genuinely complementary: LIBS reports elemental composition, which reflects soil geochemistry and is therefore a direct proxy for growing site, whereas HSI reports molecular composition, which reflects the plant’s biosynthetic response to that site. For forest-sourced materials this pairing is of particular interest, because the soil signal and the canopy-mediated metabolic signal are partially independent and may fail in different circumstances. The cost is that LIBS is micro-destructive and requires laboratory instrumentation, so cross-modal fusion of this kind belongs to confirmatory laboratory testing rather than to field screening. These results suggest that the full information content of hyperspectral data cubes—combining spectral and spatial dimensions, and potentially elemental information from orthogonal modalities—remains underexploited in current TCM authentication studies.

6.5 Interpreting reported performance: within-study versus cross-study evidence

Reported accuracies in this literature must be read with care, and two categories of comparison should be kept strictly separate.

Within-study comparisons apply several algorithms to a single dataset, holding species, sample set, instrument, spectral range, preprocessing and validation partition constant. Only these support inference about relative algorithmic merit. Their results are informative but heterogeneous. With second-order derivative plus normalization on the G. elata dataset of Ma et al. (2023), test accuracy ranged from 95.56% for PLS-DA to 100% for 1D-CNN. On the Dalbergia dataset of Chen et al. (2024), SVM and CNN both reached the reported ceiling under their respective optimal preprocessing. On the Anoectochilus dataset of Xu et al. (2025), SVM matched a CNN. These examples do not support a universal performance margin for any modelling paradigm.

Cross-study comparisons place accuracies from different publications side by side. These are not benchmarks. Studies differ in the number and geographical separation of origin classes, in sample size, in whether spectra or samples formed the unit of partitioning, in instrument and spectral range, in the form in which the material was presented to the sensor, and—most consequentially—in whether validation was internal or external. Aggregate figures of the form “deep learning outperforms PLS-DA by 10–25 percentage points”, which the present authors reproduced in the submitted version, are obtained by pooling across studies and are not defensible. The supportable conclusion is limited to the named within-study comparisons above, which show dataset-dependent outcomes rather than a general margin.

A third consideration concerns the ceiling itself. Accuracies of 100% are reported frequently in Table 2, and they should be interpreted as evidence that a task was saturated rather than that a method is exact. Two studies in the tabulated corpus reported an external test set. Li, G et al. (2025) reported 100% on the internal test set and 95.45% external-validation accuracy for Gastrodia, while Han et al. (2025) reported 100% for the training, internal test and external test sets for its best D. officinale fusion model. These few results are insufficient to establish cross-condition generalizability, and the predominance of internal validation remains the principal reason why laboratory accuracies should not be read as field performance.

6.6 Limitations and open questions

The modelling literature is methodologically homogeneous in ways that constrain what can be concluded from it. Validation is overwhelmingly internal; sample sizes are small relative to model capacity; negative results and failed architectures are not reported; and the partitioning unit is often left unspecified, so that leakage between replicate spectra of the same physical specimen cannot be excluded. Three questions follow. First, what is the realistic ceiling on external accuracy for a well-constructed model applied to an unseen harvest year? The two available external-test reports are not a basis for generalization. Second, is the reported superiority of correlation-image pipelines an architectural effect or a preprocessing effect, given that 2DCOS simultaneously changes the input representation and removes the conventional preprocessing step? A controlled comparison isolating the two has not been published. Third, how should models be evaluated when the class structure is open rather than closed—when the material presented may originate from a region absent from training? Conventional accuracy is undefined in that setting, and the anomaly-detection formulation discussed in Section 5.3 has not yet been combined with geographical classification.

7 Species-specific applications across forest-sourced resource classes

7.1 Overview and organization

The preceding sections evaluated preprocessing, feature extraction and classification on their methodological merits. The analysis now shifts from method to material. Species are grouped by the resource classes defined in Section 2.1—woody taxa, understory-cultivated herbs and forest-dwelling macrofungi—because the three differ in the tissue sampled, in the environmental variables that drive chemotypic variation, and in the authentication question of commercial importance. For each species Table 1 states the cultivation mode, the nature of its forest dependence, the dominant climatic and environmental drivers, the specific authentication challenge and the best-supported spectral approach, together with the strength of validation on which that support rests. Table 2 gives the corresponding methodological detail and reported accuracies.

7.2 Woody medicinal taxa

Woody taxa are the resource class most obviously within the remit of forest management and the class least studied by spectral authentication, largely because their drugs—bark, heartwood, fruit, flower bud—are chemically distinctive enough that species substitution has attracted more attention than origin discrimination.

The most directly forest-relevant study in this corpus concerns Dalbergia. Chen et al. (2024) applied VNIR HSI with machine learning to 10 Dalbergia species, a genus whose timber trade is regulated under CITES and in which substitution of low-value for high-value heartwood is both economically motivated and difficult to detect morphologically once the material is milled. Under optimal preprocessing, SVM and CNN both reached the ceiling on the test partition. The study is methodologically important beyond its taxon because it established the model-dependence of preprocessing choice (Section 4.2), and it demonstrates that heartwood, a tissue with no photosynthetic history and low moisture content, carries sufficient chemical signal for discrimination.

Among woody fruit and flower drugs, Gardeniae Fructus (Gardenia jasminoides) has been addressed by Cui et al. (2025), who combined NIR spectroscopy with LSTM modelling to quantify multiple quality markers and identify authenticity simultaneously. Lonicerae Japonicae Flos, the flower bud of the woody liana Lonicera japonica, was the subject of the handheld-instrument study of Peng et al. (2025), which is discussed in Section 8.1 as a deployment case. Wolfberry (Lycium barbarum), a woody shrub cultivated at scale in northwestern China, provided the dataset for the cross-attention multimodal fusion of Li, B. et al. (2025) described in Section 6.4. In each case the woody habit of the taxon has gone unremarked in the source publication, and no study has yet asked whether stand-level variables—density, rotation age, pruning regime—leave a detectable spectral trace in the harvested drug. For a review addressed to forest management this is a conspicuous omission, and it parallels at the woody-taxon level the understory-versus-open-field gap identified for herbaceous species.

7.3 Understory-cultivated herbaceous taxa

7.3.1 P. notoginseng

P. notoginseng possesses the most extensive spectral authentication literature among understory TCM species, reflecting both its high commercial value and the well-characterized geographic variation in its ginsenoside profile across Yunnan Province—the species’ primary production region. It is cultivated under shade structures or retained canopy on a three- to four-year cycle, and its saponin profile responds measurably to the climatic gradient across its production range. The evolution of analytical approaches for this species illustrates the broader trajectory of the field. Liu et al. (2023) advanced this significantly by demonstrating the environmental basis for spectral discrimination through correlating ginsenoside content (Rb1, Rg1, Rd., notoginsenoside R1) with climate factors, then achieving sensitivity, specificity, precision, and F1-score of 1.00 at both district and town levels using 2DCOS-ResNet.

The most practically significant development is the study of Zuo et al. (2025) combining FT-NIR with 2DCOS/3DCOS images and D1ResNet, which reached ceiling accuracy on the internal test partition across five production areas while reportedly removing the need for conventional preprocessing. The regulatory dimension is also noteworthy: SN/T 5727.2–2025 is, to the authors’ knowledge, the first Chinese entry–exit inspection standard to apply NIR spectroscopy to the authentication of understory-grown P. notoginseng for export inspection. Its scope is inspection rather than pharmacopoeial quality control, and it prescribes PLS-DA rather than any of the deep learning methods reviewed above; it should therefore be read as evidence that spectral authentication has entered regulatory practice at the inspection level, not as evidence of pharmacopoeial acceptance.

7.3.2 G. elata

G. elata research has advanced rapidly, driven by the species’ significant price variation across its geographically dispersed production areas. The species is an achlorophyllous mycoheterotroph dependent on Armillaria mellea and hence on woody substrate, and is cultivated almost exclusively in forest settings; its forest dependence is obligate rather than facultative. Zhaotong in Yunnan Province holds the Geographical Indication (GI) designation for premium Gastrodia, with materials from this region commanding substantially higher prices than those from Guizhou, Tibet, or Hubei origins. Ma et al. (2023) demonstrated that 1D-CNN achieved perfect F1 scores across 11 origins using Vis–NIR spectroscopy (350–2,500 nm), while simultaneously predicting gastrodin content—a key quality marker—with R²p = 0.91–0.95. Environmental factors strongly influence results: dry matter content varies significantly between Zhaotong and other origins, correlated with precipitation, altitude, temperature, and soil pH, which makes this species an informative test case for the climate-driven chemotypic drift discussed in Section 1.3.

Li, G. et al. (2025) extended the 2DCOS-ResNet paradigm to Gastrodia, achieving 100% test set accuracy and 95.45% external validation accuracy for four production regions. The external validation gap—4.55 percentage points between test and external sets—is one of the few directly reported external-validation results in this corpus and motivates the transfer learning approaches discussed in Section 8.2.

7.3.3 D. officinale

D. officinale authentication benefits from the species’ well-characterized polysaccharide and flavonoid profiles, which respond sensitively to cultivation environment. The species is epiphytic, cultivated on tree trunks or on rock under retained canopy, and is therefore among the most directly canopy-dependent of the focal taxa. Yang et al. (2022) achieved 100% recognition rates across multiple origins using CARS wavelength selection combined with nonlinear SVC, while simultaneously quantifying polysaccharides (R²p = 0.9586), polyphenols (R²p = 0.9670), and flavonoids (R²p = 0.8803). Han et al. (2025) fused FT-NIR and ATR-FTIR data with ResNet models built on 2DCOS images and again reached 100% accuracy. Critically for field deployment, Sedjoah et al. (2020) demonstrated that even handheld NIR spectrometers achieved >90% origin identification accuracy for Dendrobium from Zhejiang, Yunnan, and Anhui provinces, establishing the feasibility of portable authentication for this species.

7.3.4 Other understory herbs

For P. polyphylla, Pei et al. (2019) demonstrated that high-level data fusion of FT-MIR and NIR spectroscopy achieved 100% classification accuracy, with FT-MIR contributing more discriminative power than NIR due to its sensitivity to steroid saponin structural variations. Fritillaria species discrimination—a unique challenge due to multiple closely related species with similar morphology—has been addressed through attention-enhanced 1D-CNN applied to HySpex VNIR and SWIR data, achieving 98.97 and 99.35% accuracy for binary classification between genuine and non-genuine species (Hu et al., 2023). Wu et al. (2024) demonstrated portable NIR classification of six Fritillaria species from five GI-protected areas using novel FGSVD feature extraction, showing that even handheld instruments can discriminate among closely related TCM species. The genus has also become the testbed for non-targeted screening: Huang et al. (2026) applied a deep spectral autoencoder to detect adulterants and congeneric substitutes as anomalies rather than as enumerated classes (Section 5.3). C. chinensis, a shade-obligate herb whose berberine content increases under moderate canopy cover (Li, X. et al., 2025), was authenticated across 10 regions at 98.10% by Yuan et al. (2025) using dual-lens VNIR–SWIR feature fusion. A. roxburghii, an understory orchid subject to persistent counterfeiting, was separated from its substitutes at ceiling accuracy under small-sample conditions by Xu et al. (2025), whose comparison of SVM against CNN is discussed in Section 6.3.

7.4 Forest-dwelling macrofungi

Ganoderma lucidum is cultivated on hardwood logs or on sawdust substrate derived from forest thinnings, and its authentication problem combines species substitution within the genus with origin mislabeling and, in the spore-powder trade, adulteration with cheaper fillers. Ganoderma lucidum authentication has particularly benefited from multi-sensor data fusion. Jiang et al. (2023) achieved 100% accuracy for adulteration identification through mid-level fusion of NIRS and HSI data, with key HSI wavelength peaks at 4,200, 4,250, 5,600, 5,800, and 8,300 cm⁻¹ corresponding to polysaccharide and triterpenoid absorption features. Fu et al. (2017), working on a considerably harder problem—25 groups formed by seven species crossed with 15 producing areas—reported 93.17%, and this lower figure on a more finely divided class structure is a useful corrective to the ceiling accuracies reported elsewhere. Because fungal cell walls are chitinous rather than cellulosic and the diagnostic metabolites are triterpenoids and β-glucans, wavelength regions optimal for fungal authentication should not be assumed transferable from plant taxa, and the substrate on which the fungus was grown is a confounder with no analogue in the plant classes.

7.5 Limitations and open questions

Coverage across the three resource classes is markedly uneven. The understory herbaceous class dominates the literature, woody taxa are represented by a handful of studies in which the woody habit is incidental to the research question, and the fungal class rests on two or three publications. Within the herbaceous class, effort is concentrated on the four or five most valuable species, so that the reported performance patterns describe the species that are easiest to fund rather than the species where authentication failure causes most harm. Three questions are open. First, does any spectral model transfer across species within a resource class, or is each model an artefact of its taxon? No cross-species transfer experiment has been reported. Second, do the wavelength regions identified as discriminative differ systematically between woody, herbaceous and fungal materials in ways predictable from their dominant biopolymers? The data to answer this exist in the published studies but have never been assembled. Third, and most consequentially for this review’s argument, none of the studies in any resource class recorded canopy cover, stand density or light transmission as a metadata field, so the hypothesis that forest structure leaves a spectral trace in the harvested drug cannot be tested retrospectively on existing data. It requires new, purpose-designed acquisition.

8 Emerging frontiers

8.1 Portable and miniaturized devices

The transition from laboratory-based HSI systems to field-deployable devices represents a critical step toward practical implementation of spectral authentication along TCM supply chains, and is the point at which spectral authentication becomes usable at the forest-farm gate rather than in a central laboratory. Portable and handheld NIR spectrometers have been evaluated for medicinal-material discrimination (Sedjoah et al., 2020; Wu et al., 2024). The available studies differ in material, task and validation design, and do not provide a controlled benchtop-versus-portable accuracy or cost comparison. A 2025 study used a handheld NIR spectrometer for Lonicerae Japonicae Flos authentication and adulterant-ratio prediction (Peng et al., 2025). Consumer-grade NIR devices have also been evaluated for basic TCM discrimination tasks (Lin et al., 2020).

More advanced portable platforms are emerging. Compact video-rate hyperspectral sensors with integrated edge computing have been announced commercially, with manufacturer specifications reporting frame rates of the order of tens of hyperspectral frames per second and on-board inference capability (Living Optics, 2024); industrial HSI sorting systems are likewise marketed for throughputs of tens of items per second on conveyor belts. These figures are vendor specifications rather than peer-reviewed results, no independent validation on medicinal materials has been published, and they are cited here only to indicate the hardware envelope that is becoming available. The critical bottleneck for portable deployment remains cross-condition robustness: models trained on samples acquired with one instrument or under one set of environmental conditions may lose accuracy when tested under different conditions (Peng et al., 2025), motivating the transfer learning approaches discussed below.

8.2 Transfer learning and domain adaptation

Cross-batch, cross-instrument, and cross-year variability represents the single largest barrier to operational deployment of HSI-based TCM authentication. Models that achieve ceiling accuracy on internal partitions can lose performance when applied to new batches, instruments, or harvest years; the available studies do not provide a harmonized external dataset from which a common percentage loss can be estimated. For forest-sourced materials this problem has an ecological as well as an instrumental component: the harvest-year effect is not only a matter of drift in the sensor but of genuine chemotypic variation driven by the growing season’s temperature, precipitation and canopy condition (Section 1.3), and a domain-adaptation method that corrects instrument drift will not correct for a chemotype that has actually changed. Transfer-learning strategies have been proposed for cross-instrument adaptation, but the corpus does not contain a controlled comparison sufficient to identify a universally best strategy.

A particularly promising approach is the SISCW wavelength selection method developed by Shu et al. (2024), which enabled a single PLS model for total plant alkaloids to transfer across seven NIR devices over 7 years (2014–2020) without requiring standard samples—demonstrating that careful wavelength selection can substitute for complex transfer algorithms in some cases, and providing the longest multi-year validation currently available in the adjacent literature. Data augmentation through generative adversarial networks (GANs) is emerging as a complementary solution. Yang et al. (2025) developed NIR-GAN, a spectral data augmentation framework designed around the specific properties of spectral rather than image data—spectra are continuous and smooth, and conventional image upsampling distorts their structure—and reported 98.57% identification accuracy across 47 medicine-food homologous herb species from limited real samples. Few-shot class-incremental learning has expanded classification systems from 4 to 14 Chrysanthemum categories using only 30 samples per new category at 80.13% accuracy (Cai et al., 2023a), suggesting a practical path toward continually expanding authentication scope without complete retraining.

8.3 Explainable artificial intelligence

Regulatory acceptance of AI-driven authentication requires transparent model decisions, creating tension with the inherently opaque nature of deep learning models. Two complementary approaches are addressing this challenge. Model-intrinsic interpretability leverages attention mechanisms that reveal which spectral regions drive classification decisions. Cai et al. (2023b) used Grad-CAM++ to visualize CNN classification regions for Radix Paeoniae Alba origin identification, demonstrating that the model focused on spectral features corresponding to known paeoniflorin and albiflorin absorption bands—providing chemical validation of the model’s reasoning.

Post-hoc explanation methods apply perturbation-based analysis to trained models. Zone-based SHAP/LIME, introduced by Contreras et al. (2024), performs group perturbations on spectral zones rather than individual wavelengths, producing physically realistic explanations for correlated spectral data. SHAP analysis applied to rice quality prediction narrowed critical wavelengths to the 2000–2,500 nm range (Zhu et al., 2025), reducing the required spectral range by approximately 80% while maintaining prediction accuracy.

All of these methods are correlational. They identify the input regions on which a model’s output depends; they do not establish that those regions carry the chemical information the model is presumed to be using. A model may attend to a band because it encodes a diagnostic metabolite, or because that band happens to covary with an artefact of sample presentation, batch or instrument. Post-hoc coincidence between an attributed band and a documented absorption feature—the argument offered by most of the studies cited above—is suggestive but not probative, since the near-infrared region is dense with overlapping overtone and combination bands and coincidence is easy to obtain.

Establishing a causal link would require the attribution to be tested rather than observed, and the experimental tools to do so already exist within analytical chemistry. Four are directly applicable. Spiking and standard-addition experiments, in which a candidate metabolite is added to a matrix at graded concentrations, test whether the attributed band responds as the causal hypothesis predicts and whether model confidence tracks that response. Band assignment by orthogonal spectroscopy—matching the attributed NIR overtone to its fundamental in the MIR or to a Raman line—converts a wavelength interval into a specific bond vibration. Chemically informed ablation, in which the attributed interval is masked and the model retrained, distinguishes a band the model requires from one it merely uses. And targeted quantification of the implicated compound across the origin classes tests whether the metabolite in fact varies between origins in the direction the attribution implies. A model for which an attributed band has been assigned to a defined bond vibration, in a metabolite independently shown to vary with provenance, and whose removal measurably degrades performance, would rest on a mechanistic argument of the kind that pharmacopoeial committees are accustomed to evaluating. No published study in TCM spectral authentication has yet completed this chain, and doing so for even one species-and-band would be a more consequential contribution than a further increment in reported accuracy.

8.4 Limitations and open questions

The three frontiers reviewed here share an evidential weakness: each is supported by demonstrations rather than by comparative evaluation. Portable-device studies report the accuracy attained by one instrument on one species and rarely benchmark against a laboratory reference on the same specimens, so the accuracy cost of miniaturization is not quantified. Transfer-learning studies report success on the transfers attempted and are silent on those that failed. Explainability studies report attributions without a criterion for judging one attribution better than another. Two questions follow. First, what is the actual accuracy penalty for moving from benchtop to handheld on identical material, and does it scale with the number of origin classes? Second, are the wavelength regions attributed by XAI methods stable across model architectures trained on the same data? If two architectures achieving equal accuracy attribute to different bands, at most one attribution can reflect the underlying chemistry, and this test appears never to have been run.

9 Challenges and future perspectives

The key challenges and proposed research priorities discussed in this section are summarized in Figure 5 and Table 4. Several fundamental challenges constrain the current state of HSI-based forest-sourced medicinal material authentication and define the research priorities for the coming decade. Sample size remains a critical limitation: most studies use modest, single-campaign datasets from limited geographic regions, which cannot represent the full variation across harvest years, instruments and production environments. The absence of standardized HSI data acquisition protocols across research groups—different instruments, spectral ranges, resolutions, and illumination conditions—makes cross-study comparison nearly impossible (Section 6.5) and prevents the construction of community-wide spectral databases.

Table 4. Current challenges and proposed research directions for HSI-based authentication of forest-sourced medicinal materials.

GAN, generative adversarial network; SHAP, Shapley additive explanations; XAI, explainable artificial intelligence.

Regulatory acceptance lags behind technological capability. HSI-based methods are not yet recognized by major pharmacopoeias (Chinese, European, or United States Pharmacopoeias) as official testing methods for origin authentication, though the promulgation of SN/T 5727.2–2025 for understory P. notoginseng represents a first opening at the entry–exit inspection level. Deep learning models face particular regulatory barriers due to their opacity, and the XAI methods described in Section 8.3 have not yet been formally validated for spectral authentication applications in a regulatory context.

9.1 The understory versus open-field spectral gap

The most scientifically significant gap identified in this review is the complete absence of studies directly comparing spectral signatures of the same TCM species grown under understory versus open-field conditions. Given that understory cultivation fundamentally alters secondary metabolite profiles through shade stress, soil microbiome interactions, and humidity regulation (Zubay et al., 2021; Wang et al., 2024), the spectral consequences should be measurable and potentially substantial. Characterizing this understory-specific spectral signature would enable a new dimension of authentication: not merely identifying geographical origin, but verifying the cultivation mode that underlies Daodi quality claims. Addressing this gap would require controlled experiments in which the same TCM species—ideally from identical genetic stock—is cultivated simultaneously under measured canopy cover (e.g., 30, 50, 70% shade) and in open-field conditions, with HSI data acquired from both groups under standardized protocols, and with canopy transmittance, stand basal area and microclimate logged as covariates rather than treated as nuisance. Preliminary evidence from metabolomic studies suggests that the spectral differences should be detectable: C. chinensis accumulates significantly more berberine under moderate shade (Li, X. et al., 2025), and wild-tending cultivation modes yield chemical profiles measurably distinct from open-field production (Wang et al., 2024). If confirmed spectrally, such cultivation-mode signatures could enable a new tier of authentication—verifying not only where a medicinal material was produced, but how it was grown—which would convert spectral authentication from a quality-control tool into an instrument of forest-management verification, capable of substantiating claims made for understory cultivation as a conservation-compatible land use.

9.2 Community spectral infrastructure: precedents and requirements

Calls for shared spectral databases are routine in this literature and have produced nothing. It is therefore worth being specific about what has worked elsewhere and what such an infrastructure would actually require.

Four precedents are instructive. GenBank succeeded because sequence deposition became a condition of publication enforced by journals, which converted sharing from an act of generosity into a cost of doing business. The RRUFF project in mineralogy succeeded on a different model: a single funded group acquired reference spectra on characterized specimens under one protocol, producing internal consistency that a federated effort could not have achieved. MassBank and the GNPS platform in metabolomics demonstrated that heterogeneous spectra contributed by many laboratories become usable only when accompanied by mandatory acquisition metadata and a curation layer that flags non-conforming records. And within the spectroscopic community itself, the IEEE GRSS and Pavia/Indian Pines hyperspectral benchmarks show that even small, imperfect shared datasets reshape a field once they become the common basis of comparison.

Applied to forest-sourced medicinal materials, the requirements follow directly. A minimum reporting standard is the precondition for everything else: instrument make and spectral range, spectral and spatial resolution, illumination geometry, reference standards used for reflectance calibration, sample presentation form (intact, sliced or milled—Section 4.2), moisture content at acquisition, and the partitioning unit used in validation. Provenance metadata must be equally specific, and for this domain must extend beyond the geographical coordinate to the ecological variables that define the production system: canopy cover or transmittance, stand composition and age, altitude, aspect, soil type, cultivation mode, harvest date and growing-season climate summary. Voucher specimens must be deposited in a recognized herbarium or fungarium with accession numbers cited, since a spectral record whose taxonomic identity cannot be re-verified is of limited archival value. The database must be versioned and its records citable with persistent identifiers, so that a model can be tied to the exact data version on which it was trained. A held-out evaluation partition should be curated centrally and not released, to permit genuine benchmarking of the kind that Section 6.5 shows to be impossible at present. And because chemotypes drift with climate, the design must be longitudinal from the outset: a database that captures a single harvest year records a moving target at one instant, and its value for training durable models decays.

Institutionally, the RRUFF model appears the more realistic starting point for this field. A single well-resourced consortium—plausibly a forestry research institute working with provincial inspection authorities, which together control both the standing forest and the sampling access—could produce a coherent multi-year reference set for a small number of high-value forest-sourced species faster than an open federated repository could reach usable consistency, with federation added once the acquisition standard has been demonstrated.

9.3 Climate-resilient authentication and integration with smart forestry

Two implications for the research agenda follow from the climatic considerations set out in Section 1.3. The first is that model longevity should become a reported quantity. A calibration is an empirical claim about a chemotype–provenance relationship, and if that relationship drifts with growing-season climate then accuracy has a half-life which is currently unmeasured because almost no study revisits its origins in a later year. Multi-year designs, and the routine reporting of accuracy as a function of the interval between calibration and application, would make this tractable. The second is that climate covariates should be treated as model inputs rather than as unmodelled noise. Where growing-season temperature and precipitation are known for the harvest in question, a model conditioned on them may separate the component of spectral variation that is climatic from the component that is geographical—which is exactly the confound that limits cross-year transfer. This also suggests the natural point of integration with smart forestry: stand-level canopy and microclimate data of the kind routinely collected by forest inventory and remote sensing are precisely the covariates that spectral authentication models need, and are at present discarded at the point where the harvested material enters the analytical chain.

9.4 Research priorities

The gaps and emerging capabilities surveyed above converge on several research priorities that, if addressed in the coming years, could transform HSI authentication from a species-by-species laboratory exercise into a standardized field tool. The most urgent need is for large-scale, multi-year spectral databases with standardized acquisition protocols and ecological metadata along the lines set out in Section 9.2. Equally important is the development of cultivation-mode-aware models capable of simultaneously classifying geographical origin and cultivation conditions through multimodal fusion architectures. Cross-condition robustness—validated through transfer learning across instruments, batches, and harvest years—remains the single largest barrier to operational deployment and deserves dedicated benchmarking efforts, for which a centrally held evaluation partition is a prerequisite. On the hardware side, portable HSI systems must be validated along forest-sourced TCM supply chains, building on the SN/T 5727.2-2025 framework. Underpinning all of these, explainable AI must move from correlational attribution to tested causal assignment of bands to bond vibrations (Section 8.3)—the interpretive bridge that regulators and pharmacopoeia committees will require before granting formal recognition.

10 Discussion and conclusions

The convergence of hyperspectral imaging technology, advanced machine learning, and China’s rapidly expanding forest-based medicinal economy has created both urgent demand and increasingly mature technical capability for spectral-based geographical authentication. This review has systematically evaluated the methodological pipeline from preprocessing through classification across three classes of forest-sourced resource, revealing several clear conclusions.

In preprocessing, hybrid strategies combining scatter correction with derivative transformation can outperform simpler methods within a dataset, but optimal combinations remain species-, morphology- and model-specific, and the form in which a material is presented to the sensor can matter as much as the species it comes from. The 2DCOS transformation is a substantial simplification, serving as preprocessing and feature engineering at once, though the claim that it removes the need for conventional preprocessing rests on a small and closely related body of work.

For feature extraction, CARS remains widely used. In the directly verified A. macrocephala example, CARS, SPA, and SHAP retained 11.2–16.1% of the 224 wavelengths with zero to 1.12 percentage points of accuracy loss; this single result does not establish a universal retention threshold. Attention mechanisms integrate selection and classification end-to-end, and one-class formulations such as spectral autoencoders extend the problem from closed-set classification to open-set screening.

In classification, deep learning architectures—particularly 2DCOS/3DCOS combined with ResNet—have been reported to achieve 98–100% authentication accuracy for several major forest-sourced species under controlled laboratory conditions and predominantly internal validation. The directly verified within-study comparisons show dataset-dependent outcomes and do not support a universal margin over PLS-DA or SVM; accuracies compared across studies are not benchmarks. SVM remains a competitive comparator for small-sample datasets, and PLS-DA retains value as a transparent baseline and as the method specified in one inspection-level standard, not as evidence of pharmacopoeial acceptance. Feature-level fusion of multi-range spectral data has improved performance in several within-study comparisons, and cross-modal fusion with elemental methods is an open direction.

The field’s most consequential unanswered question—whether and how the understory growing environment imprints a distinctive spectral fingerprint on harvested medicinal materials—represents a research opportunity of both scientific and forest-management significance, since a positive answer would allow a harvested product to testify to the silvicultural conditions under which it was grown. Addressing this gap, along with the priority directions proposed in Section 9, and doing so under a climate that is actively shifting the production zones on which Daodi designations rest, would advance forest-sourced medicinal material authentication from a species-by-species research endeavor toward a unified, standardized, and field-deployable quality assurance system integrated with the sensing and data infrastructure of smart forestry.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. During the preparation and revision of this manuscript, the authors used Claude (Anthropic, San Francisco, CA, USA; exact model and version not recorded) and OpenAI Codex (OpenAI, San Francisco, CA, USA; GPT-5.6; desktop application version not recorded) for language editing, document consistency checks, and formatting assistance. The authors reviewed and edited all AI-assisted output and take full responsibility for the content of the published article.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Xian-Ge Hu, Zhejiang Agriculture and Forestry University, China

Reviewed by: Hao Liang, Zhejiang Agriculture and Forestry University, China

Yujia Dai, Zhejiang Agriculture and Forestry University, China