Skip to main content
NCBI home page
RSC Advances logoLink to RSC Advances
. 2026 Oct 16. Online ahead of print. doi: 10.1039/d6ra07316f

Single-step Raman spectroscopy for maple syrup quality control

Saba Bashir a, Talia Gélinas a, Jean-Francois Masson a,✉
PMCID: PMC13673142  PMID: 42891069

Abstract

Maple syrup quality evaluation involves multiple parameters, including color classification, flavor assessment, and the measurement of sugar concentration, which are influenced by the complex chemical composition of the syrup matrix. Conventional quality assessment relies on separate analytical and sensory approaches, creating a bottleneck in the industrial processing of maple syrup. As such, rapid techniques are needed to assess multiple quality control parameters from a single measurement. In this study, Raman spectroscopy combined with machine learning was evaluated to characterize maple syrup quality attributes using 721 samples representing different color grades (golden, amber, dark, and very dark) and flavor categories. After spectral preprocessing, principal component analysis (PCA) was performed to investigate natural spectral variations among samples. PCA revealed differences associated with syrup color and flavor categories; however, the overlap observed among several groups indicated that the relationship between Raman fingerprints and maple syrup properties involves complex spectral variations that are not fully resolved by linear unsupervised analysis. To improve discrimination, convolutional neural network (CNN) models were developed to extract nonlinear spectral features directly from Raman data. For color classification, the CNN achieved prediction accuracies of 91%, 68%, 73%, and 92% for golden, amber, dark, and very dark syrups, respectively. Flavor classification showed strong discrimination for several defect categories, including OK (83%), √ (76%), VR4 (73%), VR5 (78%), VR13 (100%), and CT (83%), while lower performance for VR1 (55%), VR2 (66%), and VR12 (64%) reflected the challenge of differentiating sensory defects with overlapping chemical characteristics. VR13 showed a high classification rate in the evaluated data; however, the small number of independent VR13 samples prevents a robust estimate of general performance for this class. In addition to Raman fingerprints, fluorescence background variations were evaluated as an indicator of optical properties and were found to correlate with maple syrup light transmittance values used for color grading. The quantitative capability of Raman spectroscopy was further assessed by predicting sucrose concentrations using standard sucrose solutions (45–67%, w/w). Concentration-dependent Raman features were extracted and used to develop a random forest regression (RFR) model, achieving coefficients of determination (R2) of 0.87. Application of the model to real maple syrup samples yielded sucrose predictions with an average bias of 0.35%, indicating that Raman-derived spectral information can be used to evaluate both compositional and quality-related characteristics of maple syrup. Overall, the results reported demonstrate the power of Raman spectroscopy for a multiparametric analysis of maple syrup.

Maple syrup quality evaluation using Raman spectroscopy provides information on multiple parameters, including color classification, flavor assessment, and the measurement of sugar concentration.graphic file with name d6ra07316f-ga.webp

1. Introduction

Maple syrup is one of the most economically and culturally significant natural products in Canada, representing a major component of the Canadian agri-food industry.1 Its commercial value depends heavily on product authenticity, sensory quality, and standardized grading.1,2 Maple syrup quality assessment involves the evaluation of multiple parameters, including sugar content, color grade, flavor characteristics, and chemical composition. Traditional maple syrup quality control relies on several independent analytical approaches, such as refractometry for measuring soluble solids, spectrophotometry for light-transmittance-based color grading, sensory panels for off-flavor detection, and chromatographic techniques for detailed compositional analysis. Although these methods provide reliable information, they require separate analyses for different quality parameters, increasing the time and resources needed for complete syrup evaluation. Moreover, sensory assessment of flavor defects depends on trained panel evaluation, where human perception can introduce variability.3 These limitations highlight the need for rapid and objective analytical approaches capable of providing complementary information on multiple maple syrup quality attributes from a single measurement.

One of the major challenges in maple syrup production is the occurrence of quality defects and adulteration. Due to its high market value, maple syrup is vulnerable to dilution or substitution with lower-cost sweeteners such as corn syrup or cane sugar, practices that may not always be easily detected through conventional inspection methods.4,5 In parallel, flavor defects caused by microbial activity, overheating, storage conditions, or processing failures can significantly alter syrup composition and sensory properties,6–10 while flavor profiles vary in the harvest season and may lead to the so-called “buddy” flavor defects for sap harvested late in the season.7 The classification system established by the Producteurs et Productrices Acéricoles du Québec (PPAQ), the world's largest consortium of maple syrup producers, categorizes syrups according to sensory evaluation performed by trained technicians. Syrups are classified as “OK” when no off-flavor is detected, while off-flavor samples are categorized according to both their intensity (√: trace/non-objectionable, VR: objectionable flavor, CT: strong defect) and defect family (classes 1–6), including naturally occurring, microbiological, chemical, unidentified, buddy, and texture-related defects.11 CT syrups are comparatively uncommon because they typically result from major microbial spoilage, severe processing failures, or extensive contamination, rendering the syrup unsuitable for conventional commercialization.

Within the VR category, different subclasses are associated with distinct chemical origins and chemical signatures (Table 1). VR1 is mainly linked to natural or processing-related off-flavors, including woody, sap-like, caramel, vegetal, or burnt notes. VR2 defects are mainly linked to microbial activity, resulting in moldy or fermented characteristics caused by metabolites such as ethanol and organic acids. VR4 represents mixed or unidentified technical defects (e.g. residual soap, oxidation, or combination effects), while VR5 corresponds to the late-season “buddy” defect. VR12 (subgroup under VR1) is associated with sap-related off-flavors, while VR13 (also under VR1) originates from excessive heating and caramelization reactions, producing compounds such as hydroxymethylfurfural (HMF), furans, carbonyl compounds, and aromatic pyrazines.1,3,11–13 The different origins of the flavor defects suggest variance in chemical composition should be present in the maple syrup.

Table 1. Organoleptic property classification for maple syrup samples by PPAQ.

Defect family √ (trace) VR (flavor defect) CT (strong defect) Sensory family
1 √1 VR1 CT1 Naturally occurring defects: Wood (11), sap (12), burnt/over-caramelized (13)
2 √2 VR2 CT2 Microbiological defects: Mold, fermented
3 √3 VR3 CT3 Chemical defects: Residues or suspected chemical contamination (washing/sanitizing products, etc.)
4 √4 VR4 CT4 Unidentified defects: Abnormal tastes such as metallic, antifoam, burnt/mixed defects
5 √5 VR5 CT5 Buddy defect (late-season bud flavor)
6 √6 VR6 CT6 Spinning syrup: Abnormal viscosity/stringy syrup
Open in a new tab

Among these categories, VR5 (buddy defect) is one of the most recognized and economically important defects in the maple industry. It develops near the end of the harvesting season during the physiological transition of maple trees toward bud formation.14,15 Changes in sap chemistry, including increased amino acids, sulfur-containing compounds, and microbial metabolites, produce characteristic cabbage-like, sulfurous, onion-like, or bitter sensory notes that negatively affect syrup quality and consumer acceptance.7 Compounds such as methionine, asparagine, dimethyl disulfide (DMDS), dimethyl trisulfide (DMTS),7 and Maillard reaction products have been associated with this defect.16 Previous studies have suggested that compounds such as methionine and asparagine, combined with microbial activity, play a critical role in the development of buddy flavor defects.12,17,18 Because VR5 occurs seasonally and can rapidly affect large syrup volumes near the end of production, it represents a major economic challenge for maple producers. Because these defects induce substantial variations in pH, viscosity, sugar composition, fluorescence background, volatile compounds, and other physicochemical properties, conventional grading methods based primarily on sensory inspection and color evaluation often remain insufficient for reliable defect discrimination.3 Consequently, there is growing interest in rapid, objective, and non-destructive analytical techniques capable of differentiating multiple maple syrup defect subclasses with improved sensitivity and reproducibility.

Conventional analytical techniques used for maple syrup characterization, including HPLC, GC/MS, and NMR, provide detailed chemical information5,19–21 but are often time-consuming, reagent-dependent, and unsuitable for rapid industrial deployment. Similarly, standard laboratory assays used to evaluate antioxidant or phenolic content require extensive sample preparation and specialized instrumentation.22,23 Although these methods are highly informative, their practical implementation for routine quality monitoring remains limited, particularly in high-throughput industrial environments.

Among emerging analytical technologies, Raman spectroscopy has gained considerable attention as a rapid, non-destructive, and label-free technique capable of providing detailed molecular fingerprints from complex food matrices.24 Recent advances in machine learning and deep learning have further expanded the capabilities of spectroscopic analysis by enabling automated extraction of subtle nonlinear spectral patterns that are often difficult to identify using traditional statistical methods.25–28 In Raman spectroscopy, recent developments include algorithmic and deep-learning approaches for noise reduction and recovery of weak spectral information.29 More broadly, deep-learning methods can model nonlinear relationships within spectral data and have been applied to classification, regression, and spectral imaging, although their performance remains dependent on data quality, representative validation, and model interpretability.30,31 Therefore, this study proposes a single-step Raman spectroscopy platform combined with machine learning for comprehensive maple syrup quality control (Fig. 1). By integrating rapid Raman spectral acquisition with advanced computational analysis, the proposed approach aims to enable Single-measurement assessment of syrup grading, off-flavor detection, percent transmittance prediction, and sucrose quantification from a single measurement. This strategy has the potential to provide a fast, objective, and scalable alternative to conventional maple syrup quality assessment methods while supporting improved product authenticity, industrial efficiency, and consumer confidence.

Fig. 1. Conceptual workflow illustrating proposed single-step Raman fingerprinting coupled with machine learning for comprehensive maple syrup analysis.

Fig. 1

Open in a new tab

2. Materials and methods

2.1. Maple syrup samples

A total of 721 maple syrup samples obtained from the PPAQ were analyzed in this study. The samples collected during the 2024 (n = 150) and 2025 (n = 571) production seasons were previously categorized by trained technicians according to standard maple syrup quality evaluation criteria, including color classification based on light transmittance (% T) and organoleptic evaluation based on flavor characteristics. Based on light transmittance grading, the dataset consisted of four maple syrup color classes: golden (n = 70), amber (n = 280), dark (n = 217), and very dark (n = 154).

In addition to color grading, samples were categorized according to their organoleptic quality classification provided by PPAQ. Regular maple syrup samples presenting characteristic maple flavor without noticeable flavor defects were labelled as OK (n = 162). Samples containing slight traces of atypical flavor were classified as “√” (crochet) samples (n = 101). Samples presenting more pronounced off-flavor characteristics were categorized under the VR classification, including VR1 (n = 85), VR2 (n = 105), VR4 (n = 78), and VR5 (n = 140). Additional VR subclasses, including VR12 (n = 14) and VR13 (n = 9), were also included. Samples associated with stronger off-flavor characteristics affecting commercial quality were categorized under the CT classification, including CT4 samples (n = 27). These samples include all usual categories produced on a yearly basis, but their sampling proportion does not represent production which is heavily dominated by OK and √ maple syrups. The category frequencies reflect the samples available in the collected dataset and occurrence in a yearly production. Categories were retained without merging, and the resulting imbalance was addressed during model training and considered when interpreting minority-class performance.

2.2. Raman spectral acquisition

Raman spectra were collected directly from maple syrup samples without chemical modification or sample preparation. For each measurement, approximately 2 mL of maple syrup was transferred into an Eppendorf tube, which was filled close to the top surface to minimize the distance between the sample and Raman probe and improve spectral acquisition efficiency. Raman measurements were performed using a BAC101-785 industrial-grade Raman probe (B&W Tek, USA) coupled SureLock™ BT Series Mini-Benchtop wavelength-stabilized 785 nm laser system (Coherent Corp., USA). Spectra were collected with an acquisition time of 3 s per spectrum. The standoff measurements were done by positioning the Raman probe directly above the syrup surface with a minimum working distance during acquisition. For each maple syrup sample, 30 consecutive Raman spectra were recorded under identical acquisition conditions, resulting in a complete dataset of 21 630 spectra (721 samples × 30 spectra/sample) (Table 2).

Table 2. Summary of maple syrup samples based on color grade and flavor profile.

Classification basis Sample categories (n) Total samples Total spectra
Color grade (% T) Golden (70), amber (280), dark (217), very dark (154) 721 21 630
Organoleptic quality OK (162), √ (101), VR1 (85), VR2 (105), VR4 (78), VR5 (140), VR12 (14), VR13 (9), CT4 (27) 721 21 630
Open in a new tab

2.3. Preparation of sucrose standard solutions

Sucrose standard solutions were prepared to establish a reference dataset for Raman-based sucrose quantification. Solutions with concentrations ranging from 45% to 67% (w/w) were prepared at 2% concentration intervals (45%, 47%, 49%, …, 67%) by dissolving appropriate amounts of sucrose in deionized water. The mixtures were vortexed and sonicated to ensure complete dissolution and obtain homogeneous solutions prior to analysis. Raman spectra of the sucrose standards were acquired using the same experimental setup and acquisition parameters as the maple syrup samples. The °Brix values of the prepared sucrose solutions and maple syrup samples were measured using an ATR-Series laboratory refractometer (SCHMIDT + HAENSCH, Germany) as the reference method for soluble solids content determination. For each concentration, 30 Raman spectra were collected for subsequent spectral analysis and model development.

2.4. Spectral preprocessing and chemometric analysis

Raw Raman spectra were exported in .asc format and processed into MATLAB R2025b (MathWorks, USA) for spectral processing and analysis. The spectral range was first truncated to the 350 and 1500 cm−1 fingerprint region containing the major vibrational information associated with maple syrup constituents while removing non-informative regions. Following truncation, all spectra were standardized by cubic spline interpolation at 1 cm−1 intervals, generating a uniform spectral axis from 350 to 1500 cm−1. Spectral smoothing was subsequently performed using a Savitzky–Golay filter with a second-order polynomial and a frame length of 15 points to reduce high-frequency noise while preserving the position, width, and intensity of Raman bands. Finally, rubber-band baseline correction was applied to compensate for broad background contributions and remove fluorescence-induced baseline variations before further spectral analysis.

Principal component analysis (PCA) was performed on the preprocessed spectra to explore natural spectral variation among maple syrup color grades and flavor categories. Convolutional neural network (CNN) models were developed to classify maple syrup color grades and flavor categories directly from the preprocessed Raman spectra. Each spectrum was treated as a one-dimensional input signal, allowing the model to learn local spectral patterns associated with variations in Raman band position, intensity, shape, and neighboring wavenumber relationships. The CNN architecture consisted of stacked one-dimensional convolutional layers for spectral feature extraction, followed by batch normalization and pooling layers to improve training stability and reduce dimensionality. The extracted spectral features were then passed through fully connected layers, and a final softmax layer was used to assign each spectrum to its corresponding class. Rectified linear unit (ReLU) activation functions were used in the hidden layers, and dropout regularization was applied to reduce overfitting. The models were trained using labelled spectra with categorical cross-entropy loss and an adaptive optimizer. For flavor classification, class-weighting was applied during training to reduce bias toward highly represented categories and improve learning from underrepresented flavor classes. The weight for each class was calculated from the training spectra using inverse class frequency:2.4.where Ntrain is the total number of training spectra, K is the number of classes, and nc is the number of training spectra in class c. These weights were applied directly to the categorical cross-entropy loss during CNN training, while validation and test metrics were calculated without weighting. In addition, leave-one-sample-out cross-validation was performed to provide a more robust evaluation of the minority categories. In each iteration, all replicate spectra from one syrup sample were withheld for testing, while the model was trained using the remaining syrup samples. This process was repeated until every independent sample had served once as the test sample. Class weights were recalculated using only the training data in each iteration, preventing information from the held-out sample from influencing model fitting. To benchmark the CNN against conventional classification approaches, we evaluated the same preprocessed Raman spectra using partial least squares discriminant analysis (PLS-DA) and support vector machine (SVM) models. Performance was assessed using accuracy, balanced accuracy, precision, recall, and F1 score. To examine the spectral basis of the CNN predictions, one-dimensional gradient-weighted class activation mapping (Grad-CAM) was applied to the final convolutional layer. For sucrose quantification, peak-to-dip intensity differences were calculated across the complete analysed Raman range of the sucrose standards by subtracting the intensity of each local minimum from adjacent local minima.ΔIj = I(ṽmax,j) − I(ṽmin,j)where ΔIj denotes the contrast feature for the jth maximum–minimum pair, and I(ṽmax,j) and I(ṽmin,j) are the Raman intensities at the corresponding local maximum and minimum, respectively. The complete set of calculated intensity differences was used as the predictor matrix for a random forest regression (RFR) model. The dataset was randomly divided into training (60%) and testing (40%) subsets. Model performance was evaluated using the coefficient of determination (R2) and root mean squared error (RMSE). Following evaluation on the test subset, the trained model was applied to the maple syrup spectra to estimate their sucrose concentrations.

3. Results and discussion

The average preprocessed Raman spectra of maple syrup samples categorized as golden, amber, dark, and very dark are presented in Fig. 2a. Although baseline correction was applied during preprocessing, noticeable differences in the 350–1500 cm−1 spectral background range were observed, which were particularly evident in darker syrup categories compared with golden and amber samples. This behavior can be attributed to the increased presence of fluorescent chromophores formed during maple syrup production. During thermal evaporation of maple sap, non-enzymatic browning reactions, mainly Maillard reactions between reducing sugars (glucose and fructose) and amino compounds, as well as caramelization reactions of sugars, lead to the formation of intermediate compounds such as hydroxymethylfurfural (HMF), furfural derivatives, and advanced brown polymeric structures known as melanoidins.32–36 In addition, naturally occurring phenolic compounds and their oxidation products contribute to the optical properties of maple syrup.37 The accumulation of these conjugated molecules in darker syrup grades increases light absorption and fluorescence emission, producing a broad background contribution that overlaps with Raman scattering.

Fig. 2. Average preprocessed Raman spectra of golden, amber, dark, and very dark maple syrup samples: (a) full fingerprint region (350–1500 cm−1) and (b) enlarged 1000–1500 cm−1 region highlighting characteristic Raman spectral differences among color grades.

Fig. 2

Open in a new tab

To further investigate the characteristic Raman features with reduced influence from fluorescence background, the 1000 to 1500 cm−1 region was magnified and is shown in Fig. 2b. Compared with the lower wavenumber region, this spectral window exhibited more distinct and well-resolved Raman bands, enabling clearer comparison among maple syrup color classes. Pronounced differences were observed around 1088–1092, 1156–1176, 1256, 1332–1361, and 1435–1460 cm−1, suggesting changes in the relative contribution of Raman-active compounds associated with syrup composition and processing. The reduced fluorescence influence in this region allowed these spectral variations to be more clearly observed compared with the full fingerprint range.

3.1. Principal component analysis

PCA was performed on the preprocessed Raman spectra (350–1500 cm−1) to evaluate the dominant spectral variability within the maple syrup datasets (Fig. 3a). The PCA score plot showed a gradual distribution of samples along PC1 rather than completely separated clusters. This behavior is scientifically meaningful because maple syrup color classification is based on percent light transmittance, which is a continuous optical property, not a discrete chemical boundary. Therefore, overlap between adjacent grades, particularly golden/amber and amber/dark, is expected. Golden and amber syrups were generally located toward positive PC1 scores, while dark and very dark syrups shifted progressively toward negative PC1 scores. This pattern suggests that Raman spectra capture systematic changes associated with syrup color intensity, although the overlap between neighboring grades indicates that color classes are not separated by abrupt chemical boundaries. The first principal component (PC1) explained most of the spectral variability, contributing 66.63% of the total variance, including broad intensity differences and changes across the Raman fingerprint.

Fig. 3. Principal component analysis (PCA) score plots of Raman spectra from golden, amber, dark, and very dark maple syrup samples: (a) the full fingerprint region (350–1500 cm−1) and (b) the selected 1000–1500 cm−1 region. (c) PCA score plots comparing Raman spectra of OK vs. VR samples, (d) PCA score plots comparing Raman spectra of OK vs. CT samples.

Fig. 3

Open in a new tab

The gradual arrangement of samples is also important from a quality-control perspective. Rather than forming isolated clusters, the samples follow a spectral continuum from lighter to darker syrup grades. This pattern supports the idea that Raman spectroscopy is sensitive to the same progressive chemical changes that influence syrup color development. At the same time, the overlap among neighboring classes shows why PCA alone is not sufficient as a final classification tool. Samples close to official transmittance boundaries are expected to have similar chemical and optical properties, and their Raman spectra may therefore occupy intermediate positions in PCA space. This explains why the unsupervised PCA plot provides useful exploratory information but does not fully resolve the four commercial color grades.

To evaluate whether the color-related separation was driven only by broad fluorescence/background effects, PCA was also performed using the selected 1000–1500 cm−1 region (Fig. 3b). This spectral window contains more resolved Raman bands and is less dominated by the broad low-wavenumber background (Fig. 2b). The color-dependent distribution was still retained, with golden and very dark samples distributed at opposite regions along PC1 and amber and dark samples positioned between them. This suggests that the differentiation among maple syrup color grades is not solely driven by fluorescence associated with color-active compounds but also originates from Raman-active molecular variations within the syrup matrix. However, because PCA is an unsupervised linear method, these score plots should be interpreted as evidence of spectral trends rather than definitive chemical identification of individual compounds.

PCA was further performed to assess the ability of Raman spectral fingerprints to differentiate maple syrup samples by organoleptic quality categories. Limited discrimination was observed between OK, √, VR, and CT samples, with substantial overlap in the PCA space (Fig. 3c, d and S1). This suggests that the spectral differences associated with these categories are relatively subtle compared with the dominant chemical variations present within maple syrup. This behavior is expected because the √ category represents a slight sensory deviation from the characteristic maple flavor, while VR samples include a broad range of flavor defects with different origins. Therefore, these groups do not represent a single chemical change, resulting in heterogeneous spectral responses.

CT samples showed a stronger tendency to separate from OK, √, and VR samples, primarily along PC1 (Fig. 3d and S1c, d), although the separation was incomplete and some CT samples overlapped with OK and other flavor categories. This partial separation suggests that CT defects are generally associated with larger spectral variation than lower-intensity flavor deviations, but that the distinction is not absolute. From an organoleptic grading perspective, CT represents strongly unpleasant flavor characteristics that substantially affect product quality; therefore, these samples are more likely to contain pronounced compositional or matrix-related differences that can be detected by Raman spectroscopy.

Pairwise PCA comparisons among VR subclasses (VR1, VR2, VR4, VR5, VR12, and VR13) showed partially different but overlapping spectral distributions, indicating that different types of flavor defects introduce different levels of variation within the Raman fingerprints (Fig. S3). This behavior is consistent with the nature of maple syrup flavor defects, where VR classifications represent diverse defect origins rather than a single chemical alteration. According to maple syrup quality evaluation, off-flavors can originate from multiple sources, including changes in sap quality before processing, microbial activity, thermal processing conditions, and contamination during production or storage. These mechanisms can modify the syrup matrix through different chemical pathways, such as sugar degradation, formation of organic acids, production of microbial metabolites, and changes in phenolic or volatile compounds.1,3 Consequently, each VR subclass may present a different combination of chemical contributors, explaining the observed spectral variability among defect categories.

Interestingly, VR13 samples displayed distinct sample-dependent distributions despite the limited number of samples (n = 9) (Fig. S3e and f). Since VR13 is associated with burnt/over-caramelized characteristics, this variability may reflect differences in the extent of thermal transformation during processing among different maple syrups. Different degrees of caramelization and non-enzymatic browning reactions can modify the relative abundance of sugar degradation products and other heat-generated compound constituents associated with color and flavor development. These differences may contribute to variations in the Raman spectral fingerprints observed within the VR13 group. A similar clustering behavior was observed for VR12 samples. The PCA plot showed that VR12 and VR13 samples formed subclusters (Fig. S3f). Although all samples were assigned to the same sensory category, the presence of distinct subclusters suggests measurable compositional differences among individual syrups. These results indicate that a common organoleptic classification does not necessarily correspond to a single chemical profile. The origin of this variability may be related to differences in sap composition, processing conditions, or other production-related factors; however, identifying the specific chemical contributors would require complementary analytical techniques such as GC-MS or LC-MS.

Independent component analysis (ICA) was performed to complement the PCA results by separating statistically independent sources of variation within the Raman spectra (Fig. S4). Unlike PCA, which maximizes the explained variance through orthogonal components, ICA seeks components that are statistically independent and may better resolve overlapping spectral contributions. ICA performed on both the full fingerprint region (350–1500 cm−1) and the selected fingerprint region (1000–1500 cm−1) produced sample distributions consistent with the trends observed in PCA. In both spectral regions, a gradual transition from golden to very dark maple syrup samples were observed, confirming that the spectral differences among color grades are preserved irrespective of the dimensionality reduction method. The comparable clustering patterns obtained using PCA and ICA further support the robustness of the preprocessing strategy and indicate that the observed discrimination arises from intrinsic spectral differences rather than methodological artifacts.

3.2. Convolutional neural networks

The exploratory PCA analysis demonstrated that Raman spectra contain relevant information associated with maple syrup color classification; however, complete separation among all grades was not achieved. This limitation is mainly related to the nature of PCA as an unsupervised linear dimensionality reduction method. PCA identifies directions explaining the largest variance in the dataset without considering class information; therefore, the extracted components do not necessarily correspond to the spectral features that provide optimal discrimination between maple syrup categories. Moreover, Raman spectra contain highly correlated variables, where neighboring wavelengths often carry related chemical information, making subtle class-specific variations difficult to isolate using linear projections.

To overcome these limitations, a convolutional neural network (CNN) was implemented to extract more complex spectral patterns directly from Raman fingerprints. Unlike PCA, CNNs use supervised learning, where model optimization is guided by the known sample categories. The convolutional layers analyze local regions of the spectra and learn relationships among neighboring Raman bands, allowing detection of small variations in peak intensity, shape, and combined spectral features that may not dominate the overall variance captured by PCA. The CNN model produced higher classification performance, achieving accuracies of 91%, 68%, 73%, and 92% for golden, amber, dark, and very dark maple syrup samples, respectively (Fig. 4a). The highest accuracy observed for very dark syrups aligns with PCA results, which showed these samples exhibited the most distinct spectral distribution.

Fig. 4. Row-normalized confusion matrices illustrating classification performance of CNN models for maple syrup samples: (a) color grade classification (golden, amber, dark, and very dark) and (b) flavor grade classification (OK, √, CT, and VR subcategories). Results were obtained using leave-one-sample-out cross-validation, in which each syrup served once as the independent test sample and all replicate spectra from the held-out syrup were excluded from model training. Values represent the proportion of samples assigned to each predicted category.

Fig. 4

Open in a new tab

In contrast, most misclassifications occurred between adjacent color grades, particularly golden-amber and amber-dark samples. This behavior reflects the continuous nature of maple syrup classification based on percent light transmittance, which represents a continuous physicochemical property rather than discrete chemical groups. Therefore, samples located near the transmittance boundary between two color classes may present highly similar compositions and Raman responses, making their discrimination more challenging.

Following color classification, the CNN model was further applied to evaluate Raman-based discrimination of maple syrup flavor categories. Compared with color grades, flavor classification represents a more complex problem because off-flavors can originate from different chemical pathways and may not always produce large variations in the overall syrup composition. The confusion matrix (Fig. 4b) showed that the CNN model successfully differentiated several flavor categories, with classification accuracies of 83%, 76%, 73%, 78%, 100% and 83% obtained for OK, CROC (√), VR4, VR5, VR13 and CT samples, respectively. The 83% accuracy obtained for OK samples indicates that their Raman fingerprints could be distinguished from several defective categories, although some overlap remained, particularly with √ samples.

Lower accuracies were observed for VR1 (55%), VR2 (66%), and VR12 (64%). VR1 was most frequently confused with VR2 and VR5, with 16% assigned to each category, while VR12 was primarily misclassified as VR2 (18%), with smaller proportions assigned to VR1 and VR13 (9% each). This overlap is reasonable because VR1 is a broad sensory family that includes woody, sap-like, vegetal, burnt, and over-caramelized defects, whereas VR12 and VR13 represent sap-related and over-caramelized subclasses within this family. Consequently, VR1 contains samples with different chemical origins and does not represent a uniform spectral population. In that case, the relevant spectral changes are likely to occur within C–C, C–O, C–O–C, and C–H vibrational regions that are already dominated by sucrose and other carbohydrates. Fermentation-related VR2 samples may contain ethanol and organic acids, whose C–C and C–O vibrations occur within approximately 850–1150 cm−1 and overlap strongly with sugar-ring and glycosidic vibrations. Sap-related VR12 differences may involve minor organic acids, amino acids, or phenolic constituents, but their Raman signals are expected to be weak relative to the sucrose-dominated spectrum. Buddy-related VR5 has been associated with amino acids and sulfur-containing compounds; potential C–S and S–S vibrations occur in the lower fingerprint region, but these compounds are present at low concentrations and may not be resolved directly in the maple syrup matrix. Overheating and caramelization associated with VR13 may alter furan, carbonyl, and conjugated structures; however, several prominent C Created by potrace 1.16, written by Peter Selinger 2001-2019 O and conjugated vibrations occur near or above the upper limit of the analysed range. However, the apparently strong VR13 result must still be interpreted cautiously because only nine independent VR13 samples were available. Since Raman spectroscopy primarily detects changes in the molecular composition of the syrup matrix and was not combined with targeted compound identification, therefore, defects caused by low-concentration sensory compounds may be more challenging to differentiate. These findings are consistent with the PCA results, where some VR subclasses exhibited clear separation while others showed overlapping distributions. Although class weighting reduced bias toward the more highly represented categories, the classification results for VR12 and VR13 should be considered preliminary and exploratory because of their limited sample numbers. Their class-specific performance is therefore sensitive to individual predictions and cannot yet be regarded as a robust estimate of classification accuracy.

The CNN was further compared with PLS-DA and SVM using the same classification tasks. The CNN achieved an overall accuracy of 0.74, balanced accuracy of 0.75, precision of 0.78, recall of 0.75, and F1 score of 0.76 (Fig. S6). The corresponding values were 0.59, 0.65, 0.61, 0.65, and 0.62 for SVM and 0.53, 0.35, 0.45, 0.35, and 0.36 for PLS-DA, respectively (Fig. S5–S7). The CNN provided comparatively strong discrimination for √, VR1, VR4, VR5, and CT, although SVM produced comparable or slightly higher scores for VR12 and VR13 (Fig. S5); however, these categories contained substantially fewer samples, and their class-specific estimates are therefore more sensitive to individual predictions.

Unlike PLS-DA, which mainly evaluates linear relationships between the spectra and class labels, the CNN can learn more complex relationships involving the position, shape, intensity, and combined behavior of neighboring Raman bands. SVM can also model complex class boundaries, but it does not directly show which spectral regions contribute to each class prediction. The one-dimensional Grad-CAM analysis (Fig. S7) was performed to examine whether the spectral regions emphasized by the CNN corresponded to chemically meaningful Raman features. For color grading, golden and amber samples showed strong attribution around 740–850 cm−1, encompassing the carbohydrate C–C/C–O stretching region near 740 cm−15 and the dominant sucrose-related band near 835 cm−1.38 Both classes also showed attribution near 1450–1500 cm−1, which includes the CH2 deformation band at approximately 1460 cm−1.39 Very dark samples placed greater weight on the approximately 950–1070 cm−1 interval, overlapping carbohydrate C–O and coupled C–H/C–O–H vibrations,40 while dark samples showed a broader distribution across these shared carbohydrate regions.

A similar interpretation applies to flavor classification. The CNN repeatedly emphasized regions containing known sugar-ring, glycosidic, C–O, C–O–H and CH2 vibrations, but assigned different relative importance to them among flavor categories. VR13 and CT showed notable attribution around approximately 980–1070 cm−1, overlapping C–O and C–H/C–O–H modes.40 VR2 showed greater attribution around approximately 1200–1270 and 1430–1490 cm−1,39 corresponding to mixed carbohydrate deformation and CH2 deformation regions. Several highlighted regions coincide with reported carbohydrate-associated vibrational regions, including bands near 835, 923–1067, 1127, 1265, and 1460 cm−1. Overall, the Grad-CAM maps indicate that the CNN relied on spectrally and chemically plausible regions. However, Grad-CAM identifies regions contributing to a prediction but does not establish that these regions are unique molecular markers or identify the compounds responsible for individual flavor defects. The maps should therefore be interpreted as evidence of class-dependent and partially chemically interpretable spectral learning, rather than definitive identification of the molecular origins of color or flavor.

An important consideration for future practical implementation is the quantification of predictive uncertainty. Recent deep-learning studies in spectroscopy have shown that uncertainty estimates can be reported alongside model predictions using approaches such as deep ensembles, Monte Carlo dropout, or evidential learning.41,42 In a practical maple syrup quality-control workflow, uncertainty could be estimated from repeated probabilistic predictions or an ensemble of independently trained models and then evaluated against classification errors using independent syrup samples. Predictions exceeding a predefined uncertainty threshold could be flagged for repeat measurement or confirmation by conventional grading and sensory assessment, whereas consistently low-uncertainty predictions could be used for routine screening. Importantly, such thresholds would need to be established and validated before deployment, and uncertainty estimates would also require calibration to determine whether they reliably reflect prediction errors. This is particularly relevant for samples close to class boundaries or samples that differ from those represented in the training data. Such a framework would also be important for underrepresented flavor categories, such as VR12 and VR13, for which the present performance estimates remain preliminary. Incorporating calibrated uncertainty estimates therefore represents an important next step toward reliable implementation of Raman-based maple syrup quality control.

3.3. Raman background analysis for % transmittance prediction

In addition to Raman spectral fingerprints, the relationship between the fluorescence background contribution and maple syrup percentage transmittance (% T) was investigated. During Raman acquisition, clear differences in baseline contribution were observed among maple syrup color grades, with golden samples exhibiting the lowest background intensity and very dark samples the highest (Fig. 5a and b). To evaluate this relationship, the Raman background was subtracted from the original spectra, and the average background intensity was calculated for each sample. The extracted background values were then compared with the measured % T, which is used for maple syrup color grading. A clear inverse relationship was observed between Raman background intensity and % T, with darker syrups exhibiting lower light transmission and higher background contributions. This behavior is consistent with compositional differences associated with the development of maple syrup color. During evaporation and processing, changes in the concentrations of chromophoric compounds and other minor constituents responsible for color formation increase light absorption, thereby reducing transmittance. These compositional changes also influence the fluorescence contribution observed during Raman measurements, explaining the progressive increase in background intensity from golden to very dark syrups.3,11

Fig. 5. Fluorescence background analysis of Raman spectra from different maple syrup color grades: (a) representative raw Raman spectra with estimated baseline contribution and (b) average baseline intensity comparison among golden, amber, dark, and very dark samples. (c) Fluorescence-based regression analysis between measured and predicted percent light transmission (b) confusion matrix showing predicted % T based on extracted average Raman background.

Fig. 5

Open in a new tab

A regression model was developed using the mean Raman background intensity against % T values of maple syrup samples (Fig. 5c). The classification results based on predicted transmittance values showed good agreement with official color grades, with accuracies of 84%, 94%, 81%, and 83% for golden, amber, dark, and very dark samples, respectively (Fig. 5d). Most classification errors occurred between neighboring color classes, as expected, because maple syrup color grading is based on continuous % T ranges rather than discrete chemical boundaries.

These results indicate that the fluorescence background observed during Raman acquisition reflects important differences associated with maple syrup color properties. While baseline correction is required to isolate Raman vibrational features for molecular analysis, the background variation itself showed a strong relationship with light transmittance. Therefore, Raman measurements provide two complementary sources of information: the corrected Raman spectra, which describe molecular fingerprints, and the background response, which reflects color-related changes in the syrup matrix.

3.4. Sucrose quantification

Sucrose solutions were prepared in water at concentrations ranging from 45% to 67% (w/w) with increments of 2%. Raman spectra acquired from these standard solutions showed concentration-dependent changes, where the intensity of sucrose-related Raman signatures increased progressively with increasing sucrose concentration. This controlled dataset was used to develop a regression model for sucrose prediction, which was subsequently applied to estimate sucrose concentration in real maple syrup samples.

However, directly applying a model developed from pure sucrose solutions to maple syrup samples is challenging due to the complexity of the maple syrup matrix. As discussed in the previous sections, Raman spectra of maple syrup, particularly within the 350–1000 cm−1 region, were strongly influenced by a broad fluorescence background. This contribution increased from golden to very dark samples, reflecting differences associated with color intensity and matrix composition. In addition, variations in light transmittance values and flavor-related compositional differences introduce further spectral variability. Therefore, using absolute Raman intensities from pure sucrose solutions may not accurately represent sucrose variation in real maple syrup samples, where matrix-dependent effects contribute significantly to the recorded spectra. To minimize the influence of these background variations, spectral features based on relative Raman responses were extracted instead of using the complete raw spectral intensity profile. Specifically, peak-to-dip differences from the entire Raman spectra were calculated and used as input features. These features were first evaluated against known sucrose concentrations to confirm their relationship with concentration changes. Features showing concentration-dependent behavior were then used to train a random forest regression (RFR) model.

The sucrose spectral dataset was divided into training and test sets (60 : 40) to develop and validate the random forest regression (RFR) model. The model showed strong agreement between measured and predicted sucrose concentrations, indicating that the Raman signal captured concentration-dependent sucrose information across the calibration range. For the selected regression model, the coefficient of determination was R2 = 0.87, with a root mean squared error of RMSE = 2.2% (Fig. 6a). The calibration bias, calculated as the mean difference between the Raman-predicted and refractometer-measured sucrose concentrations across all maple syrup samples, was 0.35%, indicating minimal systematic over- or underestimation by the model. These results confirm that Raman spectroscopy can provide quantitative information for sucrose estimation under controlled standard-solution conditions.

Fig. 6. Performance of the random forest regression model for sucrose quantification. (a) The correlation between measured and predicted sucrose concentrations using Raman-derived features. (b) Distribution of Raman-predicted sucrose concentrations in maple syrup samples using the random forest regression model, with calculated prediction bias relative to reference measurements.

Fig. 6

Open in a new tab

After validation using standard solutions, the developed RFR model was applied to maple syrup samples using the extracted peak-to-dip features to estimate sucrose-equivalent concentrations from their Raman spectra. Maple syrup contains sucrose as the dominant sugar, but also contains glucose, fructose, organic acids, minerals, phenolic compounds, and color-associated products that can alter Raman intensity and background response. Therefore, the transfer from pure sucrose solutions to real maple syrup is more difficult than calibration alone, because the syrup matrix introduces additional spectral variation that is not present in aqueous sucrose standards.

4. Conclusions

In conclusion, this study demonstrates the potential of Raman spectroscopy as a rapid, reagent-free, and single-step analytical approach for maple syrup quality assessment. The Raman spectra captured meaningful variations associated with color grade, flavor category, fluorescence/background behavior, and sucrose-related composition, showing that a single spectral measurement can contain multiple layers of quality information. Unsupervised PCA revealed broad trends associated with maple syrup color and flavor categories, but the overlap among several groups showed that the chemical variation in maple syrup is complex and not fully resolved by linear dimensionality reduction alone. This was especially evident for flavor classification, where OK, √, and several VR subclasses showed partial spectral overlap because flavor defects can arise from different chemical origins and may involve low-abundance sensory-active compounds. Supervised machine-learning analysis improved class discrimination by extracting more subtle spectral patterns from the Raman fingerprints. The CNN-based classification results supported the ability of Raman spectroscopy to distinguish several maple syrup quality categories, although misclassification among related VR subclasses indicated that some flavor defects remain chemically similar or spectrally heterogeneous.

The fluorescence/background contribution, often treated as an analytical interference in Raman spectroscopy, also provided useful information for maple syrup evaluation. Its systematic increase from lighter to darker syrup classes and its relationship with percent transmittance indicate that the background signal reflects meaningful matrix-dependent optical properties. Therefore, Raman measurements can provide two complementary sources of information: corrected vibrational fingerprints for molecular classification and background-derived features associated with color intensity and light transmittance. For sucrose estimation, Raman spectra of aqueous sucrose standards showed a clear concentration-dependent response and enabled calibration of regression models. However, transfer of the model to real maple syrup samples was limited by matrix complexity, since Brix reflects total soluble solids rather than sucrose alone and maple syrup contains additional sugars, organic acids, minerals, phenolics, and color-associated compounds. Overall, the results show that Raman spectroscopy combined with chemometric and machine-learning analysis can support rapid, multi-parameter maple syrup quality evaluation, but future work should include matrix-matched calibration standards, a larger representation of rare flavor defects, and external validation across production seasons and instruments to improve robustness for industrial deployment.

Conflicts of interest

There are no conflicts to declare.

Supplementary Material

RA-OLF-D6RA07316F-s001

Acknowledgments

The authors acknowledge the financial support from the Natural Science and Engineering Research Council (NSERC) of Canada, the Canadian Foundation for Innovation (CFI) and the Producteurs and Productrices Acéricoles du Québec (Quebec Maple Syrup Producers). S.B. thanks the financial support of the Fonds de Recherche du Québec – Nature et Technologies (FRQ-NT) and of Université de Montréal. JFM benefitted from support provided by CIRCA (FRQ-S grant: https://doi.org/10.69777/294004), from the RQMP (FRQ-NT grant: https://doi.org/10.69777/309032) and from the CQMF (FRQ-NT grant: https://doi.org/10.69777/341666). The authors disclose the use of artificial intelligence for language improvement and sentence structure. All texts were initially written by the authors and subsequently proof-read by the authors.

Data availability

The website https://github.com/jfmassonUdeM/Data.git contains all raw data used in this article.

Supplementary information (SI): figures are available to support the PCA, ICA, SVM and CNN data presented in the main text. See DOI: https://doi.org/10.1039/d6ra07316f.

References

  1. Perkins T. D. and van den Berg A. K.Chapter 4 Maple Syrup—Production, Composition, Chemistry, and Sensory Characteristics, in Advances in Food and Nutrition Research, Academic Press, 2009, Vol. 56, pp. 101–143 [DOI] [PubMed] [Google Scholar]
  2. Stuckel J. G. Low N. H. The chemical composition of 80 pure maple syrup samples produced in North America. Food Res. Int. 1996;29(3):373–379. doi: 10.1016/0963-9969(96)00000-2. [DOI] [Google Scholar]
  3. Filteau M. Lagacé L. LaPointe G. Roy D. Maple sap predominant microbial contaminants are correlated with the physicochemical and sensorial properties of maple syrup. Int. J. Food Microbiol. 2012;154(1):30–36. doi: 10.1016/j.ijfoodmicro.2011.12.007. [DOI] [PubMed] [Google Scholar]
  4. Mohammed F. Warland J. Guillaume D. A comprehensive review on analytical techniques to detect adulteration of maple syrup. Microchem. J. 2021;164:105969. doi: 10.1016/j.microc.2021.105969. [DOI] [Google Scholar]
  5. Paradkar M. Sakhamuri S. Irudayaraj J. Comparison of FTIR, FT-Raman, and NIR spectroscopy in a maple syrup adulteration study. J. Food Sci. 2002;67(6):2009–2015. doi: 10.1111/j.1365-2621.2002.tb09493.x. [DOI] [Google Scholar]
  6. Porter W. Buch M. Willits C. Maple sirup. IV. Effect of heating sirups under conditions of high temperature and low water content: Some physical and chemical changes. J. Food Sci. 1952;17(1-6):475–481. doi: 10.1111/j.1365-2621.1952.tb16792.x. [DOI] [Google Scholar]
  7. Camara M. Cournoyer M. Sadiki M. Martin N. Characterization and removal of buddy off-flavor in maple syrup. J. Food Sci. 2019;84(6):1538–1546. doi: 10.1111/1750-3841.14618. [DOI] [PubMed] [Google Scholar]
  8. van den Berg A.; Perkins T.; Isselhardt M.; Godshall M. and Lloyd S.Metabolism off-flavor in maple syrup part II: Remediation of metabolism off-flavor in maple syrup, National Maple Syrup Digest, 2009, 21A, pp. 11–18 [Google Scholar]
  9. Ramadan M. F. Gad H. A. Farag M. A. Chemistry, processing, and functionality of maple food products: An updated comprehensive review. J. Food Biochem. 2021;45(8):e13832. doi: 10.1111/jfbc.13832. [DOI] [PubMed] [Google Scholar]
  10. Martin N., Optimisation du contrôle du moussage dans les casseroles de l’évaporateur en production acéricole - PARTIE 1, Centre ACER, Saint-Norbert d'Arthabaska, QC, 2017. https://www.agrireseau.net/documents/96935/optimisation-du-contr%C3%B4le-du-moussage-dans-les-casseroles-de-l_evaporateur-en-production-acericole-partie-1 [Google Scholar]
  11. ACER P. e. p. a. d. Q. P. C., Maple Syrup: A World of Flavours – Quality Control and Flavour Classification of Maple Syrup, Longueuil, Québec, Canada, 2021, https://ppaq.ca/app/uploads/2021/07/PPAQ_DepliantACER_Infographics_2021_ANG.pdf [Google Scholar]
  12. Lagacé L. Pitre M. Jacques M. Roy D. Identification of the Bacterial Community of Maple Sap by Using Amplified Ribosomal DNA (rDNA) Restriction Analysis and rDNA Sequencing. Appl. Environ. Microbiol. 2004;70(4):2052–2060. doi: 10.1128/AEM.70.4.2052-2060.2004. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. van den Berg A. K., Godshall M. A., Lloyd S. W. and Isselhardt M. L., Metabolism Off-Flavor in Maple Syrup, 2009, https://api.semanticscholar.org/CorpusID:15523588
  14. Garcia E. J. McDowell T. Ketola C. Jennings M. Miller J. D. Renaud J. B. Metabolomics reveals chemical changes in Acer saccharum sap over a maple syrup production season. PLoS One. 2020;15(8):e0235787. doi: 10.1371/journal.pone.0235787. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Holgate K. C., Changes in the Composition of Maple Sap during the Tapping Season, 1950 [Google Scholar]
  16. van Boekel M. A. Formation of flavour compounds in the Maillard reaction. Biotechnol. Adv. 2006;24(2):230–233. doi: 10.1016/j.biotechadv.2005.11.004. [DOI] [PubMed] [Google Scholar]
  17. Fabian F. Buskirk H. Aerobacter aerogenes as a cause of ropiness in maple sirup. Ind. Eng. Chem. 1935;27(3):349–350. doi: 10.1021/ie50303a027. [DOI] [Google Scholar]
  18. Lagacé L. Girouard C. Dumont J. Fortin J. Roy D. Rapid prediction of maple syrup grade and sensory quality by estimation of microbial quality of maple sap using ATP bioluminescence. J. Food Sci. 2002;67(5):1851–1854. doi: 10.1111/j.1365-2621.2002.tb08734.x. [DOI] [Google Scholar]
  19. Bashir S. Beniani I. Stevenson K. Montiel-Leon J. M. Forest S. Chisanga M. Masson J.-F. Untargeted profiling and classification of maple syrup quality from the chromatographic and mass spectrometric analysis of volatile and nonvolatile chemical fractions. ACS Food Sci. Technol. 2025;5(5):1941–1959. doi: 10.1021/acsfoodscitech.5c00118. [DOI] [Google Scholar]
  20. Abou-Zaid M. M. Nozzolillo C. Tonon A. Coppens M. Lombardo D. A. High-performance liquid chromatography characterization and identification of antioxidant polyphenols in maple syrup. Pharm. Biol. 2008;46(1–2):117–125. doi: 10.1080/13880200701735031. [DOI] [Google Scholar]
  21. Gad H. A. Ramadan M. F. Farag M. A. Authentication and quality control determination of maple syrup: A comprehensive review. J. Food Compos. Anal. 2021;100:103901. doi: 10.1016/j.jfca.2021.103901. [DOI] [Google Scholar]
  22. Thériault M. Caillet S. Kermasha S. Lacroix M. Antioxidant, antiradical and antimutagenic activities of phenolic compounds present in maple products. Food Chem. 2006;98(3):490–501. doi: 10.1016/j.foodchem.2005.05.079. [DOI] [Google Scholar]
  23. Legault J. Girard-Lalancette K. Grenon C. Dussault C. Pichette A. Antioxidant activity, inhibition of nitric oxide overproduction, and in vitro antiproliferative effect of maple sap and syrup from Acer saccharum. J. Med. Food. 2010;13(2):460–468. doi: 10.1089/jmf.2009.0029. [DOI] [PubMed] [Google Scholar]
  24. Li-Chan E. C. Y. The applications of Raman spectroscopy in food science. Trends Food Sci. Technol. 1996;7(11):361–370. doi: 10.1016/S0924-2244(96)10037-6. [DOI] [Google Scholar]
  25. Geddis A. Williams H. Bashir S. Malenfant J. Dubois C. Hamlet L. Masson J.-F. Artificial intelligence and machine learning for plasmonic and surface-enhanced sensing. Chem. Soc. Rev. 2026;55(6):3599–3644. doi: 10.1039/D5CS01522G. https://dx.doi.org/10.1039/d5cs01522g [DOI] [PubMed] [Google Scholar]
  26. Lussier F. Thibault V. Charron B. Wallace G. Q. Masson J.-F. Deep learning and artificial intelligence methods for Raman and surface-enhanced Raman scattering. TrAC, Trends Anal. Chem. 2020;124:115796. doi: 10.1016/j.trac.2019.115796. [DOI] [Google Scholar]
  27. Coca-Lopez N. Alcolea-Rodriguez V. Bañares M. A. Brockhauser S. Gorenflot J. Henderson A. Hildebrandt R. Jeliazkova N. Kochev N. Lozano Diz E. Artificial intelligence-powered Raman spectroscopy through open science and FAIR principles. ACS Nano. 2025;19(44):38189–38218. doi: 10.1021/acsnano.5c09165. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Liu Y. Chen S. Xiong X. Wen Z. Zhao L. Xu B. Guo Q. Xia J. Pei J. Artificial intelligence guided Raman spectroscopy in biomedicine: Applications and prospects. J. Pharm. Anal. 2025:101271. doi: 10.1016/j.jpha.2025.101271. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Fang S. Wu S. Chen Z. He C. Lin L. L. Ye J. Recent progress and applications of Raman spectrum denoising algorithms in chemical and biological analyses: A review. TrAC, Trends Anal. Chem. 2024;172:117578. doi: 10.1016/j.trac.2024.117578. [DOI] [Google Scholar]
  30. Liu X. An H. Cai W. Shao X. Deep learning in spectral analysis: Modeling and imaging. TrAC, Trends Anal. Chem. 2024;172:117612. doi: 10.1016/j.trac.2024.117612. [DOI] [Google Scholar]
  31. Zhan H. Huang Y. Chen Z. Recent progress in artificial intelligence enabled NMR spectroscopy: Methodologies, implementations, quality assessments, and prospects. Applied Physics Reviews. 2026;13(1):011322. doi: 10.1063/5.0277355. [DOI] [Google Scholar]
  32. Unno T. Antioxidant activity of different grades of maple syrup as determined by the hydrophilic oxygen radical absorbance capacity method. Food Sci. Technol. Res. 2015;21(3):495–498. doi: 10.3136/fstr.21.495. [DOI] [Google Scholar]
  33. Singh M. Zhang M. Espinal-Ruiz M. Rathnayake S. Xue J. Shi J. Liu X. Hanner R. Corradini M. G. Maple syrup adulteration: Fluorescence fingerprints as a source of information for enhanced detection. J. AOAC Int. 2026;109(1):10–21. doi: 10.1093/jaoacint/qsaf029. [DOI] [PMC free article] [PubMed] [Google Scholar]
  34. Saraiva A. Carrascosa C. Ramos F. Raheem D. Lopes M. Raposo A. Maple syrup: chemical analysis and nutritional profile, health impacts, safety and quality control, and food industry applications. Int. J. Environ. Res. Public Health. 2022;19(20):13684. doi: 10.3390/ijerph192013684. [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Singh A. S. Jones A. M. P. Saxena P. K. Variation and correlation of properties in different grades of maple syrup. Plant Foods Hum. Nutr. 2014;69(1):50–56. doi: 10.1007/s11130-013-0401-x. [DOI] [PubMed] [Google Scholar]
  36. Ball D. W. The chemical composition of maple syrup. J. Chem. Educ. 2007;84(10):1647. doi: 10.1021/ed084p1647. [DOI] [Google Scholar]
  37. Li L. Seeram N. P. Maple syrup phytochemicals include lignans, coumarins, a stilbene, and other previously unreported antioxidant phenolic compounds. J. Agric. Food Chem. 2010;58(22):11673–11679. doi: 10.1021/jf1033398. [DOI] [PubMed] [Google Scholar]
  38. Salvador L. Guijarro M. Rubio D. Aucatoma B. Guillén T. Vargas Jentzsch P. Ciobotă V. Stolker L. Ulic S. Vásquez L. Exploratory monitoring of the quality and authenticity of commercial honey in Ecuador. Foods. 2019;8(3):105. doi: 10.3390/foods8030105. [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Pierna J. A. F. Abbas O. Dardenne P. Baeten V. Discrimination of Corsican honey by FT-Raman spectroscopy and chemometrics. Base. 2011;15(1):75–84. [Google Scholar]
  40. Paradkar M. M. Sivakesava S. Irudayaraj J. Discrimination and classification of adulterants in maple syrup with the use of infrared spectroscopic techniques. J. Sci. Food Agric. 2002;82(5):497–504. doi: 10.1002/jsfa.1067. [DOI] [Google Scholar]
  41. Zhan H. Huang Z. Xing C. Chen X. Huang Y. Chen Z. Uncertainty exploration of deep learning enabled fast multidimensional NMR spectroscopy of proteins. Anal. Chem. 2026;98(17):12362–12370. doi: 10.1021/acs.analchem.5c07201. [DOI] [PubMed] [Google Scholar]
  42. Kakita V. M. R. Hansen D. F. Deep learning assisted proton pure shift NMR spectroscopy. J. Am. Chem. Soc. 2026;148(9):9226. doi: 10.1021/jacs.5c22860. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

RA-OLF-D6RA07316F-s001
RA-OLF-D6RA07316F-s001.pdf (2.9MB, pdf)

Data Availability Statement

The website https://github.com/jfmassonUdeM/Data.git contains all raw data used in this article.

Supplementary information (SI): figures are available to support the PCA, ICA, SVM and CNN data presented in the main text. See DOI: https://doi.org/10.1039/d6ra07316f.

Articles from RSC Advances are provided here courtesy of Royal Society of Chemistry

RESOURCES