Integrating decoupled Raman scattering and intrinsic autofluorescence spectral profiles with machine learning enables portable spectrometers to differentiate newly harvested green tea from visually indistinguishable one-year-aged leaves with high statistical accuracy. In a methodological study evaluating Jiuhuafo green tea harvested across two consecutive production cycles, researchers demonstrated that merging molecular vibrational data with electronic transition profiles compensates for the physical noise and optical resolution limitations inherent to compact field instruments. The dual-modal spectroscopic framework achieved a peak five-fold cross-validation accuracy of 97.97 percent and an external test accuracy of 95.63 percent on a smartphone-based handheld spectrometer using a one-dimensional convolutional neural network. The same deep learning architecture secured an accuracy of 96.25 percent on a high-resolution laboratory confocal micro-Raman system. The findings illustrate a viable pathway for transitioning food authenticity screening from destructive chemical assays to rapid, non-invasive diagnostic evaluations conducted directly at agricultural receiving docks and commercial distribution points.

The commercial incentive to verify the production timeline of green tea arises from distinct consumer preferences and market pricing structures in major producing regions such as China. Early-spring harvests, particularly tender flushes gathered prior to the Qingming Festival known as Mingqian tea, command substantial financial premiums due to their delicate sensory profiles, vibrant green hue, and rich concentrations of natural active constituents such as theanine and polyphenolic catechins. However, green tea is an unfermented commodity whose bioactive molecules undergo gradual oxidative degradation over extended storage periods, inevitably reducing product quality and flavor vibrancy. Because properly stored one-year-aged leaves preserve a macroscopic visual appearance virtually identical to freshly harvested tea, unscrupulous vendors sometimes blend or mislabel aged stocks as current-year production. Traditional verification has depended heavily on human sensory inspection, which introduces subjectivity, inconsistent reproducibility, and a reliance on specialized personnel that cannot keep pace with commercial supply chains.

To establish an objective and non-destructive testing methodology, the research team designed an experimental investigation centered on eight hundred individual leaves of Special Grade Jiuhuafo green tea grown in Qingyang County, Anhui Province, China. The cohort was derived from standardized harvests gathered in early April of 2023 and early April of 2024, representing aged and fresh batches respectively. All samples underwent uniform industrial fixation, rolling, and drying before being hermetically sealed in nitrogen-flushed, aluminum-laminated composite bags and stored at a constant five degrees Celsius and six percent relative humidity. This strict preservation protocol eliminated uncontrolled environmental variables, such as moisture fluctuation and thermal degradation, thereby isolating the intrinsic biochemical signatures that accompany temporal aging. A stratified random sampling technique provided four hundred leaves per vintage, ensuring an evenly balanced dataset capable of accounting for natural leaf-to-leaf biological variance.

Optical analysis was conducted on individual leaves using two contrasting hardware platforms to test whether portable field instruments could replicate or exceed the diagnostic utility of benchtop systems. The primary reference unit was a high-resolution laboratory confocal micro-Raman spectrometer operating with a 785-nanometer excitation laser, an integration time of two seconds across thirty accumulations, and an optical resolution of 3.5 inverse centimeters across the spectral interval from 200 to 1700 inverse centimeters. In parallel, investigators utilized a portable handheld Raman device measuring 159 by 78.9 by 27 millimeters, driven by an identical 785-nanometer excitation wavelength but operating with a broader optical resolution of approximately 10 inverse centimeters and a five-second integration time. Every tea leaf was measured without destructive chemical extraction, grinding, or repeated surface scanning, maintaining a single-acquisition protocol intended to replicate high-throughput on-site screening operations.

Rather than treating the intense background emission present in the raw optical spectra as unwanted noise to be discarded, the analytical approach systematically decoupled the underlying autofluorescence baseline from discrete Raman scattering peaks. Raw Raman spectra acquired from complex plant matrices typically feature an autofluorescence background several orders of magnitude stronger than inelastic Raman scattering. The team employed a modified tenth-order iterative polynomial fitting algorithm equipped with an asymmetric refitting mechanism, penalizing intensities above the curve to dynamically decouple the baseline across sweeping iteration thresholds up to one hundred cycles. The baseline-corrected signal retained the Raman vibrational fingerprint between 200 and 1700 inverse centimeters, while the fitted curved baseline was isolated as the endogenous autofluorescence emission spectrum across 797 to 906 nanometers. Both extracted single-modality signals were subsequently normalized using the standard normal variate transformation to mitigate physical scattering from leaf surface irregularities.

Statistical evaluation of the extracted spectral features revealed stark mechanistic differences between the analytical signals captured by the two modalities. In the Raman spectra collected on the benchtop instrument, significant differences between 2023 and 2024 samples appeared as discrete, localized vibrational bands. Pronounced variations were detected in the 200 to 400 inverse centimeter region, associated with collective skeletal vibrations and backbone deformations, where false discovery rate corrected significance metrics exceeded values of 48 alongside large effect sizes. Further discriminative bands appeared between 1000 and 1100 inverse centimeters, representing carbon-carbon and carbon-nitrogen stretching in theanine, alongside polyphenolic catechin ring bending near 1030 inverse centimeters and chlorophyll macrocycle stretching from 1150 to 1170 inverse centimeters. These localized features exhibited an overall enhancement in aged leaves, reflecting oxidative polymerization and altered microenvironmental rigidity.

In contrast to the fragmented distribution of Raman peaks, the autofluorescence spectra exhibited continuous broadband divergence across virtually the entire measured window. On the benchtop instrument, 91.1 percent of the autofluorescence variables displayed large statistical effect sizes, driven by the progressive oxidative breakdown of green tea polyphenols and the gradual conversion of chlorophyll into pheophytin derivatives through magnesium loss. This structural degradation expands conjugated molecular systems, causing widespread baseline shifts and long-wavelength optical tailing. When evaluated on the handheld spectrometer, the Raman modality experienced substantial signal attenuation due to the wider entrance slit and lower resolution of the portable optics, resulting in zero features reaching a large effect size. However, the macroscopic autofluorescence emission proved remarkably resilient to optical downsampling, preserving large effect sizes across 54.9 percent of its spectral range.

Data-level fusion concatenated the standardized Raman and autofluorescence feature vectors for each sample, yielding combined vectors of 4276 variables for the benchtop platform and 1502 variables for the portable device. Five distinct machine learning classification algorithms were constructed and tuned across stratified five-fold cross-validation using 640 training samples: support vector machines, partial least squares discriminant analysis, gradient boosting decision trees, adaptive boosting, and a custom one-dimensional convolutional neural network. The neural network utilized one-dimensional convolutional layers with rectified linear units, max-pooling, dropout layers, and a cosine annealing learning rate schedule within an AdamW optimization framework to extract hierarchical nonlinear features directly from the concatenated spectral inputs, while an independent external test partition of 160 unseen samples provided final performance validation.

Cross-validation outcomes demonstrated that dual-modal fusion delivered substantial performance gains across both hardware platforms, with the most dramatic improvements occurring on the handheld device. On the benchtop instrument, five-fold cross-validation accuracy reached 95.47 percent for support vector machines and 96.25 percent for the convolutional neural network under fused conditions, representing modest improvements over the best single-modality baselines. On the portable handheld spectrometer, however, dual-modal concatenation increased the average accuracy across all five classifiers by an average of 2.38 percentage points, more than double the 0.97 percentage point gain observed on the benchtop unit. Fused handheld cross-validation accuracies reached 95.94 percent for support vector machines, 97.66 percent for partial least squares discriminant analysis, and 97.97 percent for the one-dimensional convolutional neural network.

To investigate why the portable spectrometer benefited disproportionately from spectral concatenation, the researchers mapped feature redundancy using Jaccard similarity coefficients and Venn diagrams across statistically significant variables. For the benchtop instrument, the Jaccard coefficient between significant Raman and autofluorescence variables was 0.835, with autofluorescence contributing only 15.0 percent unique significant features due to excessive collinearity captured by the high-density grating. On the handheld spectrometer, the Jaccard coefficient plummeted to 0.612, and the unique significant feature contribution of autofluorescence jumped to 34.2 percent. The physical widening of the optical slit in the portable device effectively acted as a low-pass filter that reduced collinearity in fine Raman bands while leaving macroscopic fluorophore emissions intact, creating a broader, non-redundant information space that machine learning algorithms readily exploited.

The computational workflow proved exceptionally rapid, supporting real-time operational deployment. Beyond the initial five-second optical integration required by the handheld device, spectral decoupling, normalization, and neural network inference consumed just 0.33 milliseconds per leaf on standard computing hardware. On the independent external validation set of 160 tea leaves, the optimal fused handheld model produced only seven classification errors, misidentifying four fresh leaves as aged and three aged leaves as fresh, while achieving an area under the receiver operating characteristic curve of 0.9917. While McNemar tests indicated that lightweight classical models such as partial least squares discriminant analysis achieved statistically comparable test set generalization, the convolutional architecture provided the highest overall classification stability. The researchers noted that future industrial translation will require testing diverse tea cultivars, variable processing styles, multi-year harvests, and moisture conditions, as well as calibrating algorithms across different portable devices.

Subject of Research: Nondestructive identification of fresh and aged green tea via dual-modal spectral fusion and machine learning

Article Title: Identification of fresh and aged green tea based on Raman-autofluorescence spectral fusion and machine learning: from laboratory analysis towards field application

Article References: He, Z., Yang, G., He, Z., Cao, F., Liu, Y., Zhu, S., & Dong, D. (2026). Identification of fresh and aged green tea based on Raman-autofluorescence spectral fusion and machine learning: from laboratory analysis towards field application. Food Chemistry: X, 40, Article 104771. https://doi.org/10.1016/j.fochx.2026.104771

Image Credits: AI Generated

DOI: 10.1016/j.fochx.2026.104771

Keywords: green tea, Raman spectroscopy, autofluorescence, spectral data fusion, machine learning, convolutional neural network, food fraud detection, portable spectrometers, Jiuhuafo tea, partial least squares discriminant analysis, theanine, catechins, chlorophyll degradation, iterative polynomial fitting, standard normal variate, optical sensing, chemometrics