Identification of common buckwheat (Fagopyrum esculentum Moench) adulterated in Tartary buckwheat (Fagopyrum tataricum (L.) Gaertn) flour based on near-infrared spectroscopy and chemometrics

Abstract

Near-infrared spectroscopy (NIRS) presents great potential in the identification of food adulteration due to its advantages of nondestructive, simple, and easy to operate. In this paper, a method based on NIRS and chemometrics was proposed to predict the content of common buckwheat (Fagopyrum esculentum Moench) flour in Tartary buckwheat (Fagopyrum tataricum (L.) Gaertn) flour. Partial least squares regression (PLSR) and support vector regression (SVR) models were used to analyze the spectrum data of adulterated samples and predict the adulteration level. Various preprocessing methods, parameter-optimization methods, and competitive adaptive reweighted sampling (CARS) wavelength-selection methods were used to optimize the model prediction accuracy. The results of PLSR and SVR modeling for predicting of Tartary buckwheat adulteration content were satisfactory, and the correlation coefficients of the optimum identification models were above 0.99. In conclusion, the combinations of NIRS and chemometrics indicated excellent predictive performance and applicability to analyze the adulteration of common buckwheat flour in Tartary buckwheat flour. This work provides a promising method to identify the adulteration of Tartary buckwheat flour and results obtained can give theoretical and data support for adulteration identification of agro-products.

Graphical abstract

Highlights

• •Adulteration of Tartary buckwheat flour was identified by NIRS and chemometrics.
• •Quantitative model of the adulterated degree of Tartary buckwheat was developed.
• •Common buckwheat content could be accurately predicted by PLSR and SVR modeling.
• •Correlation coefficients of the optimum identification models all exceeded 0.99.

1. Introduction

Although there are wide varieties of buckwheat, there are two cultivated species that are of agricultural significance, and one is common buckwheat (Fagopyrum esculentum Moench), the other is Tartary buckwheat (Fagopyrum tataricum (L.) Gaertn) (He et al., 2019). Tartary buckwheat flour has similar appearance to common buckwheat flour, but the flavonoid content in common buckwheat is much lower than that of Tartary buckwheat (Li et al., 2019). Owing to the high edible value and nutritional value of Tartary buckwheat, it is favored by consumers with a large market demand. However, to reduce costs and obtain higher profits, some unscrupulous merchants add common buckwheat to Tartary buckwheat or directly mimic Tartary buckwheat. The addition of low-quality substances including common buckwheat reduces the nutritional value of Tartary buckwheat and disturbs the normal market order, thereby negatively affecting the interest of consumers (Zuo et al., 2014).

Ultraviolet and visible spectroscopy (UV–Vis) (Popa et al., 2020), real-time quantitative polymerase chain reaction (RT-PCR) (Li et al., 2021), metabolomics (Li et al., 2022), chromatography (Dou et al., 2022), gas chromatography-mass spectrometry (GC-MS) (Shi et al., 2021), and nuclear magnetic resonance spectroscopy (Gunning et al., 2023) are extensively used to identify food adulteration. A multiplex RT-PCR assay has been developed to distinguish Tartary buckwheat from common buckwheat (Kim et al., 2023). A metabolomics approach based on ultra-high performance liquid chromatography (UPLC) coupled to triple quadrupole mass spectrometry was proposed to identify the metabolites in Tartary buckwheat and common buckwheat seeds (Li et al., 2022). The chromatographic fingerprint of Tartary buckwheat flour is established by UPLC to distinguish adulterated Tartary buckwheat from pure Tartary buckwheat (Wang et al., 2016). The key aroma compounds in Tartary buckwheat were characterized by GC-MS combined with means of sensory-directed flavor analysis (Shi et al., 2021). The adulteration of saffron was identified by 60 MHz nuclear magnetic resonance hydrogen spectrum (Gunning et al., 2023). The results of these methods are relatively direct and accurate, but they also have the shortcomings of complicated operation, high cost, large workload, and poor timeliness. Therefore, establishing a rapid, accurate, simple, and efficient analytical technique to identify the adulteration of Tartary buckwheat flour is urgent.

Near-infrared spectroscopy (NIRS) has the advantages of rapidity, nondestructive nature, easy operation, and high repeatability. NIRS-related techniques combined with chemometrics are extensively used in grain, spices, meat products, tea, and other fields (Lima et al., 2020; Firmani et al., 2019; Liu et al., 2019). Partial least squares regression (PLSR) can solve the multicollinearity of variables in chemometrics. It is suitable for the problem that the number of variables is greater than the sample size. PLSR realizes regression modeling and data-structure simplification by combining model and epistemic methods as the most commonly used quantitative analysis method (Leng et al., 2021). Portable NIRS combined with PLSR has also been adopted to detect the adulteration level in quinoa flour (Wang et al., 2022).

Moreover, support vector regression (SVR) maps nonlinear data to high-dimensional space through kernel function (radial basis function, RBF) and enable linear separability of data. Then, the principle of minimizing structural risk is adopted to process the data (Park et al., 2015). For the prediction problem of small samples, the SVR model has better prediction effect. The quantitative prediction of cyclic adenosine phosphate content in jujube by NIRS has been reported, and results showed that NIRS combined with SVR could greatly improve the prediction performance and stability of the quantitative model (Chen et al., 2019).

More studies on the adulteration identification of Tartary buckwheat by NIRS are necessary and have excellent scientific significance and promising applications. In the present study, NIRS combined with PLSR and SVR algorithms were used to construct quantitative prediction models for the adulteration of common buckwheat flour in Tartary buckwheat flour. Meanwhile, the effects of various preprocessing methods, parameter-optimization methods, and competitive adaptive reweighted sampling (CARS) feature variables selection on the model's prediction accuracy were explored. An accurate and simple method of identifying Tartary buckwheat adulteration was established, providing new insights into the quality control of Tartary buckwheat.

2. Materials and methods

2.1. Sample preparation and spectral collection

2.1.1. Samples preparation

The experimental samples were Tartary buckwheat (Fagopyrum tataricum (L.) Gaertn) from Inner Mongolia, Sichuan, and Shanxi and common buckwheat (Fagopyrum esculentum Moench) from Shaanxi, for a total of four representative buckwheat samples (provided by Zhengzhou Duofuduo Co., Ltd.). The buckwheat rice was cleaned and smashed through an 80-mesh sieve, and then the sample was placed in an electrothermal constant-temperature blast-drying oven at 60 °C to dry until constant weight. The buckwheat flour was directly passed through an 80-mesh sieve, sealed in a bag with a number, and stored at −20 °C.

After grinding and sieving the above Tartary buckwheat samples, they were fully mixed with different proportions of common buckwheat samples as follows: 0% (pure Tartary buckwheat), 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, and 100% (pure common buckwheat) (Wang et al., 2018). Each proportion of samples was configured with 50 copies for a total of 1800 samples (Chen et al., 2018; Wu et al., 2017). They were sealed in a bag and stored at −20 °C for later use. Before spectrum collection, the sample was put into a centrifugal tube and homogenized for 2 min by point-motion mixing instrument (LC-Vortex-P1, Shanghai Lichen Bangxi Instrument Technology Co., Ltd.) to minimize the possible dispersion effect inherent in particle size and ensure sample uniformity.

2.1.2. Spectrum collection

The samples were collected using an FX 2000 near-infrared spectrometer (room temperature, 25 °C; humidity, 60%). The blank collection was used as the measurement background. The samples were weighed and placed in a sample cup to avoid gaps. Morpho software was used to collect the spectral data of the samples. The instrument process parameters were set. The wavelength-acquisition range was 900–1700 nm, the resolution was 7.8 nm, and the integration time was set to 20 ms. Three parallel spectra were collected for each sample, and each scan was repeated 32 times. The NIRS data of pure Tartary buckwheat samples and adulterated Tartary buckwheat samples were collected by the same method. After collecting three stable spectra, the average spectra were taken as the final spectra. In this experiment, the set partitioning based on joint x-y distance algorithm was used to randomly select the calibration set and prediction set of the spectral data of the Tartary buckwheat samples according to the ratio of 5:1 (Zhou et al., 2020; Khamsopha et al., 2021). The calibration-set samples were used to establish the model, and the prediction set samples were used to verify the model accuracy. The NIRS data of the samples were analyzed and processed by MATLAB program (R2018b).

2.2. Preprocessing of spectral data

As a secondary analysis technology, NIRS spectrum contains rich information of the analytes. Furthermore, the characteristics of serious band overlap, poor specificity of spectral information, and low signal-to-noise ratio necessitate spectrum preprocessing before establishing the model to eliminate the influence of spectral offset or baseline change on the model.

The spectral data were averaged and preprocessed by MATLAB (R2018b). The following preprocessing methods were selected: Standard normal variate (SNV), multiplicative scatter correction (MSC), min–max normalization (MMN), Savitzky–Golay filtering (SG), SG filtering first derivative (SG-1st), SG filtering second derivative (SG-2nd), and SNV transformation and detrending (SNV-DT) (Li et al., 2020; Liu et al., 2019). Appropriate preprocessing methods can improve the stability of the quantitative model.

2.3. Selection of feature wavelengths

The appropriate variable-selection method can eliminate the redundant variables in the spectrum and simplify the model to improve its accuracy. The feature-wavelength-selection method was used to eliminate ineffective information in the spectral data that affect the prediction ability of the model. The effective information was then extracted from the complex spectral information as modeling variables to establish an efficient and stable mathematical model. CARS algorithm can effectively reduce the influence of collinear variables while removing non-informative variables. CARS method was further used to screen certain wavelengths representing the main information of the raw spectrum and preprocessed data, and then a quantitative analysis model was constructed according to the feature variables.

2.4. Model establishment

PLSR and SVR were used as modeling methods to predict the adulteration content of Tartary buckwheat flour. To evaluate the accuracy, applicability, and stability of the quantitative prediction model more accurately and intuitively, the determination coefficient of calibration (R²c), determination coefficient of prediction (R²p), root-mean-square error of calibration (RMSEC), and root-mean-square errors of prediction (RMSEP) were regarded as the main evaluation indices. The larger the R²p was, the closer it was to 1, indicating that the correlation between the measured value and the predicted value was better, and the higher the accuracy of the model. The RMSEC and RMSEP values evaluate the accuracy of the prediction results of the calibration set and prediction set, and the smaller the values were, the closer the predicted value of the sample was to the real value, the higher the prediction accuracy of the model was. The performance of the model was evaluated on the basis of fitting degree and accuracy, and the linear-fitting degree of the regression model was evaluated by comparing the real adulteration value with the predicted one (Mahgoub et al., 2020). When the correlation coefficient R²p of the model exceeded 0.8, it was considered suitable for accurate prediction (Zhang et al., 2014). The R²p was higher than 0.99, the slope was close to 1, and the intercept value was close to 0, indicating that the prediction performance of the model was better and the linearity was higher (Santos et al., 2016).

2.5. Parameter optimization

Based on SVR algorithm modeling, RBF kernel function was selected, including penalty parameter C and kernel parameter g. The set range was (0.1, 1024) The parameters greatly influence the process of modeling, and the prediction ability of the model built by different parameters differs. When using SVR algorithm for modeling, the parameters set or obtained were probably not optimum or close to the optimum, and too large parameters can easily lead to overfitting of the model. Too small also led to model underfitting, so optimizing the parameters was necessary to obtain a better prediction model (Sun et al., 2021). In this experiment, cross-validation (CV), genetic algorithm (GA), and particle swarm optimization algorithm (PSO) were selected to optimize the parameter combination (C, g), determine the optimum parameter combination, and establish a SVR model with strong prediction ability.

3. Results and discussion

3.1. NIRS acquisition and data preprocessing

Fig. 1(a), S1a, and S2a are the raw spectral curves of 1800 samples of three kinds of Tartary buckwheat mixed with 12 proportions of Shaanxi common buckwheat. The characteristic absorption peaks of the spectral curves were basically the same, and the overlap was relatively large. The NIRS primarily revealed the composition and content of the sample. The raw NIRS produced baseline drift and noise owing to the interference of various factors (light, particle size, density, surface texture, and other physical factors). Therefore, appropriate spectral preprocessing was necessary to highlight the differences caused by dopants.

Fig. 1.

Raw and preprocessing spectra of Tartary buckwheat in Inner Mongolia: (a) RAW, (b) SNV, (c) MSC, (d) MMN, (e) SG, (f) SG-1st, (g) SG-2nd, and (h) SNV-DT.

The NIRS curves of the seven preprocessing methods of SNV, MSC, MMN, SG, SG-1st, SG-2nd and SNV-DT are shown in Fig. 1(b–h), S1(b–h), and S2(b–h). Many overlaps existed in the spectral curves after preprocessing, and distinguishing the spectral differences between pure and adulterated Tartary buckwheat samples was difficult. Therefore, the purity of Tartary buckwheat must be further quantified by regression model. Accordingly, PLSR and SVR were used to establish quantitative models, and the influence of spectral preprocessing methods on modeling results was analyzed.

3.2. PLSR model performance

3.2.1. Full-spectrum PLSR model

After preprocessing the NIRS data, the PLSR full-spectrum model was constructed. The model prediction results are shown in Table 1. In addition to the poor modeling effect of SG, SG-1st, and SG-2nd for Sichuan Tartary buckwheat samples, the R²p of the full-spectrum PLSR model established by other preprocessing methods reached more than 0.99, and the prediction performance of the model was enhanced. Notably, the prediction performance of the model after SG, SG-1st, and SG-2nd preprocessing decreased, which may be due to the preprocessing-induced loss of effective information in the spectral data or to the noise in the test area causing baseline offset. Ultimately, the model's prediction performance decreased (Liu et al., 2019).

Table 1.

Prediction results of PLSR models under different preprocessing of Tartary buckwheat from Inner Mongolia, Sichuan and Shanxi.

Preprocessing methods	Number of variables	Inner Mongolia					Sichuan					Shanxi
		Optimal factor number	Calibration set		Prediction set		Optimal factor number	Calibration set		Prediction set		Optimal factor number	Calibration set		Prediction set
			R2c	RMSEC	R2p	RMSEP		R2c	RMSEC	R2p	RMSEP		R2c	RMSEC	R2p	RMSEP
RAW	256	13	0.9987	0.0163	0.9984	0.0183	13	0.9932	0.0382	0.9910	0.0407	11	0.9972	0.0244	0.9977	0.0223
SNV	256	11	0.9989	0.0154	0.9986	0.0170	11	0.9922	0.0393	0.9902	0.0448	9	0.9967	0.0255	0.9973	0.0251
MSC	256	10	0.9988	0.0160	0.9986	0.0168	11	0.9922	0.0393	0.9901	0.0449	9	0.9967	0.0255	0.9973	0.0250
MMN	256	11	0.9988	0.0160	0.9987	0.0158	13	0.9926	0.0385	0.9918	0.0403	10	0.9967	0.0257	0.9980	0.0219
SG	256	13	0.9987	0.0169	0.9982	0.0192	12	0.9922	0.0409	0.9907	0.0420	11	0.9969	0.0255	0.9977	0.0224
SG-1st	256	13	0.9988	0.0163	0.9991	0.0140	14	0.9934	0.0363	0.9851	0.0487	8	0.9967	0.0252	0.9937	0.0332
SG-2nd	256	212	0.9992	0.0127	0.9974	0.0236	156	0.9954	0.0299	0.9759	0.0606	73	0.9984	0.0180	0.9966	0.0267
SNV-DT	256	10	0.9988	0.0158	0.9986	0.0168	11	0.9922	0.0392	0.9874	0.0489	9	0.9967	0.0254	0.9975	0.0241

In the adulteration analysis of Inner Mongolia Tartary buckwheat, the model obtained by SG-1st algorithm had the highest accuracy, with R²p and RMSEP reaching 0.9991 and 0.014, respectively. In the adulteration analysis of Sichuan Tartary buckwheat, the model obtained by MMN algorithm was better than that of the PLSR model established by the raw spectrum, and the accuracy was the highest, with R²p and RMSEP reaching 0.9918 and 0.0403, respectively. Similarly, in the adulteration analysis of Shanxi Tartary buckwheat, the PLSR model obtained by MMN preprocessing surpassed the PLSR model established by the raw spectrum, with the highest accuracy R²p of 0.9980 and RMSEP of 0.0219. These results indicated that PLSR model can be used to determine the purity of Tartary buckwheat.

Fig. 2, S3, and S4 show the distribution of predicted residuals sum of squares (PRESS) in the PLSR model under different preprocessing methods of the calibration set of Tartary buckwheat samples. These samples were of various proportions and originated from Inner Mongolia, Sichuan, and Shanxi. Fig. 2(f) shows the distribution of PRESS in the SG-1st-PLSR model of Tartary Buckwheat calibration set in Inner Mongolia. With increased number of factors, the PRESS decreased continuously. When the number of factors was 13, the PRESS reached the minimum 0.2225. With increased number of factors, the value of PRESS tended to stabilize, showing that the number of factors under the minimum PRESS was the optimum (Zhou et al., 2020). In the process of establishing a PLSR model, the number of factors selected by the model greatly influenced the model accuracy. A fewer number of factors corresponded with less impact of the model, but it may have also led to a decline in model accuracy. A greater number of factors corresponded with more comprehensive points calculated by the model and greater closeness to the actual situation. However, a large number of factors may not only increase the computational complexity of the model, but also increase the number of variables with low or no correlation, leading to overfitting of analysis results (Yan et al., 2019). The optimum factor number of the model after SG-2nd preprocessing was large, which may lead to the phenomenon of model overfitting, as reflected in the large R²c. Conversely, SG-2nd was no longer used in the subsequent modeling because R²p was too small.

Fig. 2.

Changes in PRESS value of the PLSR model under different preprocessing methods of Inner Mongolia Tartary buckwheat:(a) RAW, (b) SNV, (c) MSC, (d) MMN, (e) SG, (f) SG-1st, (g) SG-2nd, and (h) SNV-DT.

Compared with previous results, ours performed better. A real-time quantitative detection method was established for adulterated corn oil, rapeseed oil, and sunflower seed oil in camellia oil based on NIRS and chemometrics (Du et al., 2021). After optimization with different preprocessing methods such as SNV, MSC, SG, and MMN, the quantitative prediction model of PLSR was constructed. R²p was greater than 0.995, RMSEC and RMSEP were less than 6.79 and 4.98, respectively, and the prediction performance of adulteration level of camellia oil was better. PLSR model was also performed to detect the adulteration ability of coconut sugar with different concentrations in palm sugar by using MSC preprocessing spectrum. The model R²p was 0.91 and RMSEP was 9.13%. These results demonstrated the potential of NIRS for food adulteration identification (Rismiwandira et al., 2021). However, many redundant variables existed among the 256 variables in the full spectrum, and the construction of PLSR model took a long time. Therefore, further optimizing the prediction model and extracting feature variables was necessary to improve the discriminant accuracy and modeling efficiency.

3.2.2. Non-full-spectrum PLSR model

To eliminate redundant variables in the NIRS and improve the stability and accuracy of the PLSR prediction model, CARS algorithm was used to simplify the model. According to the NIRS data after different preprocessing methods, a non-full-spectrum PLSR model was established. The model parameters and prediction results are shown in Table 2. After CRAS screening, only SNV, MSC, MMN, and SNV-DT combined with PLSR model showed significant improvement in prediction correlation coefficient, among which the model established after MSC preprocessing had better prediction ability and was superior to the full-spectrum MSC-PLSR model. The MMN-CARS-PLSR model showed higher accuracy than other preprocessing models in the adulteration analysis of Tartary buckwheat in Inner Mongolia, with an R²p of 0.9988 and RMSEP of 0.0152. MSC-CARS-PLSR model showed better performance in the adulteration analysis of Tartary buckwheat from Sichuan, with an R²p of 0.9924 and RMSEP of 0.04. For the adulteration analysis of Shanxi Tartary buckwheat, the SNV-DT-CARS-PLSR model showed better prediction performance, with an R²p of 0.9976 and RMSEP of 0.0235. The selection of feature wavelengths improved the generalization ability of the model, indicating that the CARS method had a certain effect on extracting the effective information of spectral data. CARS-PLSR was used to conduct quantitative analysis of sesame oil adulteration, and found that the model based on variable selection of CARS was superior to the all-spectrum model after comparing the RMSE values of the models (Chen et al., 2018). It was also reported that PLSR was presented to establish a quantitative analysis model for two low-cost adulterants (maltodextrin and starch) in hawthorn fruit flour. Results showed that the PSO-CARS-PLSR model showed good predictive performance, with R²p reaching 0.993 and RMSEP 0.65 (Sun et al., 2021). CARS-PLS model was also the optimum model for quantifying the content of adulterated lotus stamens and corn stigmas in saffron (Li et al., 2020). The above studies confirmed the feasibility of CARS for screening feature variables when using PLSR to construct models.

Table 2.

Prediction results of PLSR models of Tartary buckwheat from Inner Mongolia, Sichuan and Shanxi under different preprocessing after CARS screening.

Preprocessing methods	Inner Mongolia						Sichuan						Shanxi
	Number of variables	Optimal factor number	Calibration set		Prediction set		Number of variables	Optimal factor number	Calibration set		Prediction set		Number of variables	Optimal factor number	Calibration set		Prediction set
			R2c	RMSEC	R2p	RMSEP			R2c	RMSEC	R2p	RMSEP			R2c	RMSEC	R2p	RMSEP
RAW	26	9	0.9986	0.0173	0.9984	0.0182	35	10	0.9923	0.0406	0.9879	0.0480	39	9	0.9970	0.0251	0.9975	0.0234
SNV	35	8	0.9988	0.0160	0.9985	0.0174	37	8	0.9919	0.0399	0.9913	0.0424	39	7	0.9967	0.0254	0.9970	0.0263
MSC	29	6	0.9987	0.0166	0.9987	0.0167	35	8	0.9917	0.0404	0.9924	0.0400	39	7	0.9967	0.0256	0.9976	0.0240
MMN	32	7	0.9987	0.0168	0.9988	0.0152	35	10	0.9916	0.0408	0.9900	0.0438	35	8	0.9966	0.0260	0.9978	0.0227
SG	32	10	0.9986	0.0174	0.9775	0.0669	39	10	0.9921	0.0411	0.9903	0.0430	29	9	0.9968	0.0257	0.9975	0.0234
SG-1st	44	7	0.9984	0.0186	0.9633	0.0868	47	27	0.9908	0.0427	0.9805	0.0550	42	5	0.9965	0.0263	0.9924	0.0373
SNV-DT	41	7	0.9987	0.0163	0.9984	0.0175	55	8	0.9914	0.0412	0.9902	0.0432	51	8	0.9969	0.0249	0.9976	0.0235

3.3. SVR model performance

3.3.1. Full-spectrum SVR model

Based on SVR algorithm modeling, RBF kernel function was selected. CV, GA, and PSO were also selected to optimize the parameter combination (C, g) and determine the optimum parameter combination. Thus, an SVR model with strong prediction ability was established. The prediction results of Tartary buckwheat purity by full-spectrum SVR model are shown in Table 3. The R²p of the full-spectrum SVR model established by the three parameter-optimization methods combined with different preprocessing methods was found to exceed 0.96. Among them, the prediction result of the SVR model established by SNV, MSC, MMN, and SNV-DT was better than that of the SVR model established by the raw spectrum, and even better than the PLSR model in the adulteration identification analysis of Sichuan and Shanxi Tartary buckwheat. For the full-spectrum CV-SVR model, the R²p of the MSC preprocessing model was higher, but the kernel parameter g was 1024, which was prone to overfitting and poor generalization ability (Tu et al., 2015). In the adulteration analysis of Tartary buckwheat in Inner Mongolia, the prediction accuracy of MSC-GA-SVR model was higher, i.e., R²p was 0.9985 and RMSEP was 0.0002. In Sichuan Tartary buckwheat adulteration analysis, the prediction accuracy of SNV-DT-CV-SVR model was higher, i.e., R²p was 0.9957 and RMSEP was 0.0004. In Shanxi Tartary buckwheat adulteration analysis, the prediction accuracy of MMN-PSO-SVR model was higher, i.e., R²p was 0.9982 and RMSEP was 0.0002. These results showed that the three parameter-optimization algorithms were feasible to establish the prediction model for SVR.

Table 3.

Prediction results of SVR models under different preprocessing of Tartary buckwheat from Inner Mongolia, Sichuan and Shanxi.

Modeling method	Preprocessing methods	Number of variables	Inner Mongolia						Sichuan						Shanxi
			Parameters		Calibration set		Prediction set		Parameters		Calibration set		Prediction set		Parameters		Calibration set		Prediction set
			C	g	R2c	RMSEC	R2p	RMSEP	C	g	R2c	RMSEC	R2p	RMSEP	C	g	R2c	RMSEC	R2p	RMSEP
CV-SVR	RAW	256	128	36.7583	0.9993	0.0001	0.9970	0.0003	776.0469	4.5948	0.9895	0.0012	0.9827	0.0016	48.5029	73.5167	0.9980	0.0002	0.9958	0.0005
	SNV	256	8	55.7152	0.9993	0.0001	0.9982	0.0003	6.0629	42.2243	0.9991	0.0001	0.9949	0.0006	0.1436	42.2243	0.9960	0.0004	0.9960	0.0006
	MSC	256	36.7583	1024	0.9993	0.0001	0.9983	0.0002	48.5029	1024	0.9977	0.0002	0.9905	0.0011	0.6598	1024	0.9951	0.0005	0.9964	0.0005
	MMN	256	6.9644	168.8970	0.9993	0.0001	0.9983	0.0002	1.5157	168.8970	0.9946	0.0005	0.9939	0.0007	0.1895	194.0117	0.9960	0.0004	0.9974	0.0003
	SG	256	111.4305	84.4485	0.9993	0.0001	0.9971	0.0003	337.7940	10.5561	0.9880	0.0013	0.9820	0.0018	48.5029	73.5167	0.9967	0.0003	0.9964	0.0004
	SG-1st	256	3.4822	1024	0.9963	0.0004	0.9967	0.0004	675.5881	1024	0.9905	0.001	0.9805	0.0018	64	1024	0.9946	0.0005	0.9825	0.0031
	SNV-DT	256	4.5948	147.0334	0.9993	0.0001	0.9969	0.0005	2.6390	64	0.9991	0.0001	0.9957	0.0004	0.1250	64	0.9960	0.0004	0.9926	0.0011
GA-SVR	RAW	256	64.1679	0.0649	0.9683	0.0034	0.9847	0.0017	94.9643	0.3252	0.9578	0.0046	0.9691	0.0029	62.5853	0.3233	0.9867	0.0014	0.9937	0.0007
	SNV	256	0.4503	4.1247	0.9972	0.0003	0.9982	0.0002	41.7828	0.0620	0.9778	0.0022	0.9796	0.0026	4.7297	10.9282	0.9986	0.0001	0.9979	0.0003
	MSC	256	7.7540	471.3669	0.9990	0.0001	0.9985	0.0002	76.8093	5.9071	0.9782	0.0022	0.9801	0.0025	51.3329	1000	0.9992	0.0001	0.9973	0.0003
	MMN	256	15.9755	128.4962	0.9993	0.0001	0.9983	0.0002	80.8023	0.2985	0.9788	0.0022	0.9809	0.002	26.1010	76.8977	0.9992	0.0001	0.9982	0.0003
	SG	256	63.6873	0.0592	0.9642	0.0039	0.9828	0.019	87.6362	0.4330	0.9560	0.0048	0.9671	0.0032	28.1486	0.7420	0.9860	0.0015	0.9931	0.0008
	SG-1st	256	82.8784	32.5375	0.9944	0.0006	0.9956	0.0005	75.0867	273.8486	0.9762	0.0024	0.9634	0.0045	18.6157	496.9807	0.9929	0.0007	0.9863	0.0031
	SNV-DT	256	32.4694	3.8949	0.9993	0.0001	0.9981	0.0002	74.5807	0.0430	0.9785	0.0022	0.9757	0.003	0.9395	12.0984	0.9971	0.0003	0.9982	0.0002
PSO-SVR	RAW	256	14.7129	0.1834	0.9603	0.0043	0.9799	0.0022	28.3583	1.1515	0.9597	0.0044	0.9701	0.0028	22.1067	1.0983	0.9874	0.0013	0.9939	0.0007
	SNV	256	22.6567	3.6196	0.9993	0.0001	0.9984	0.0002	12.4957	0.2170	0.9783	0.0022	0.9819	0.0023	3.0771	10.6916	0.9983	0.0002	0.9980	0.0003
	MSC	256	5.3255	470.7442	0.9988	0.0001	0.9985	0.0002	55.3193	9.4121	0.9789	0.0021	0.9813	0.0023	10.0883	1000	0.9986	0.0001	0.9978	0.0003
	MMN	256	12.0768	118.8129	0.9993	0.0001	0.9983	0.0001	16.7824	1.5821	0.9796	0.0021	0.9836	0.0018	11.1221	77.1472	0.9992	0.0001	0.9982	0.0002
	SG	256	12.4957	0.217	0.9600	0.0043	0.9793	0.0022	21.7417	1.6115	0.9558	0.0049	0.9675	0.0031	29.2601	1.084	0.9871	0.0014	0.9935	0.0007
	SG-1st	256	69.6259	44.2865	0.9949	0.0006	0.9959	0.0005	73.3634	279.5391	0.9762	0.0024	0.9634	0.0045	34.6422	264.9562	0.9929	0.0007	0.9864	0.0031
	SNV-DT	256	14.0111	3.9591	0.9992	0.0001	0.9982	0.0002	10.6049	0.3308	0.9793	0.0021	0.9804	0.0025	21.6685	11.4546	0.9992	0.0001	0.9979	0.0003

3.3.2. Non-full-spectrum SVR model

To eliminate the redundant variables in the NIRS and improve the stability and accuracy of the SVR prediction model, the CARS algorithm was used to simplify the model. Fig. 3 shows the process of CARS optimization of Shanxi Tartary buckwheat after SNV-DT preprocessing. In the CARS algorithm, the number of Monte Carlo sampling runs was set to 100 times, and the wavelength variable corresponding with the minimum root-mean-square error of cross-validation (RMSECV) model was selected by five-fold cross validation. Fig. 3(a) shows the relationship between the number of Monte Carlo sampling times and the number of sampling variables. In the first five times of sampling, the number of sampling variables decreased rapidly, which reflected mostly the process of eliminating non-informative variables. After the decrease to a certain extent, it tended to be flat. Fig. 3(b) shows that when the number of sampling times reached 13, the corresponding RMSECV was at least 0.02711. Fig. 3(c) shows that the variable coefficients changed with different sampling times. “*" was used in the figure to describe the optimum feature wavelength corresponding with the minimum RMSECV. Fig. 3(d) shows the variable distribution map obtained by screening. The number of feature variables was reduced from 256 to 51, which effectively eliminated the redundant variables in the spectrum.

Fig. 3.

CARS variable-selection optimization process: (a) changes in the number of selected variables, (b) changes in RMSECV, (c) regression coefficients of each variable during the calculations of CARS algorithm, and (d) distribution of selected feature variables by CARS.

The CARS algorithm was used to screen the characteristics of the other two kinds of Tartary buckwheat samples under preprocessing. The prediction results of adulteration degree of non-full-spectrum SVR model are shown in Table 4. For Tartary buckwheat samples from Inner Mongolia and Sichuan province, the overall effect of the model established using CARS screening was poor and not as good as the full-spectrum SVR model. This finding may be due to the absence of effective information or insufficient extraction of effective information in the process of feature-wavelength screening, resulting in poor modeling effect (Tu et al., 2015). For Tartary buckwheat sample from Shanxi, the prediction performance of the model established by SNV-DT preprocessing was the optimum and better than that of the full-spectrum model. The R²p of SNV-DT-CARS-PSO-SVR model reached 0.9987, and the parameter value was (10.0567, 48.8665). On the premise of ensuring the prediction accuracy of the model, the CARS algorithm improved the generalization ability and modeling efficiency of the model, and the effect of the effective information extraction of spectral data was relatively ideal.

Table 4.

Prediction results of SVR models of Tartary buckwheat from Inner Mongolia, Sichuan and Shanxi under different preprocessing after CARS screening.

Modeling methods	Preprocessing methods	Inner Mongolia							Sichuan							Shanxi
		Number of variables	Parameters		Calibration set		Prediction set		Number of variables	Parameters		Calibration set		Prediction set		Number of variables	Parameters		Calibration set		Prediction set
			C	g	R2c	RMSEC	R2p	RMSEP		C	g	R2c	RMSEC	R2p	RMSEP		C	g	R2c	RMSEC	R2p	RMSEP
CV-SVR	RAW	26	445.7219	256	0.9987	0.0001	0.9972	0.0003	35	294.0668	128	0.9887	0.0012	0.9734	0.0025	39	194.0117	675.5881	0.9984	0.0002	0.9938	0.0007
	SNV	35	3.4822	891.4438	0.9993	0.0001	0.9981	0.0002	37	1.3195	388.0234	0.9943	0.0006	0.9917	0.0009	39	0.0947	588.1336	0.9953	0.0005	0.9867	0.0025
	MSC	29	111.4305	1024	0.9980	0.0002	0.9982	0.0002	35	55.7152	1024	0.9851	0.0015	0.9879	0.0015	39	6.9644	1024	0.9941	0.0006	0.9960	0.0006
	MMN	32	13.9288	891.4438	0.9989	0.0001	0.9980	0.0002	35	3.0314	1024	0.9910	0.0009	0.9922	0.0008	35	0.1649	1024	0.9949	0.0005	0.9971	0.0004
	SG	32	776.0469	222.8609	0.9991	0.0001	0.9611	0.0092	39	337.7940	64	0.9874	0.0014	0.9807	0.0019	29	55.7152	675.5881	0.9962	0.0004	0.9963	0.0004
	SG-1st	44	776.0469	1024	0.9973	0.0003	0.9978	0.0072	47	1024	1024	0.9815	0.0019	0.9646	0.0034	255	1024	1024	0.9933	0.0007	0.9871	0.0017
	SG-2nd	34	1024	1024	0.9923	0.0008	0.9907	0.001	54	0.001	0.001	0	0.0986	0	0.1634	51	1024	1024	0.9865	0.0014	0.9844	0.0025
	SNV-DT	41	5.278	512	0.9993	0.0001	0.9975	0.0002	55	6.0629	55.7152	0.9904	0.001	0.9932	0.0008	51	0.2176	776.0469	0.9981	0.0002	0.9973	0.0003
GA-SVR	RAW	26	99.0023	0.5494	0.9825	0.0019	0.992	0.0009	35	67.6401	3.9387	0.9582	0.0046	0.9679	0.003	39	74.1291	1.4134	0.9861	0.0015	0.9931	0.0008
	SNV	35	23.6285	170.1336	0.9992	0.0001	0.9978	0.0002	37	38.1195	0.5332	0.9813	0.0019	0.9816	0.0023	39	0.0948	40.4282	0.9912	0.0009	0.9930	0.0013
	MSC	29	5.2268	211.3114	0.9970	0.0003	0.9976	0.0003	35	69.1133	52.9404	0.9813	0.0019	0.9817	0.0023	39	23.6267	999.9924	0.9946	0.0005	0.9964	0.0005
	MMN	32	44.7944	969.4309	0.9992	0.0001	0.9977	0.0002	35	18.4855	6.8551	0.9781	0.0022	0.9769	0.0029	35	16.4793	436.5378	0.9976	0.0002	0.9984	0.0002
	SG	32	79.6892	0.4616	0.9775	0.0024	0.9892	0.0013	39	82.7218	2.7132	0.9588	0.0045	0.9676	0.0031	29	55.8240	2.4634	0.9858	0.0015	0.9931	0.0008
	SG-1st	44	91.5884	310.8626	0.9962	0.0004	0.9972	0.0097	47	99.7173	858.0904	0.9747	0.0026	0.9518	0.0061	255	87.5227	515.6899	0.9929	0.0007	0.9892	0.0015
	SG-2nd	34	99.5260	995.9393	0.9895	0.0011	0.9909	0.001	54	99.6802	998.6448	0.9446	0.0059	0.6128	0.0548	51	99.8743	996.2616	0.9793	0.0022	0.9610	0.0096
	SNV-DT	41	71.9222	0.105	0.9969	0.0003	0.9970	0.0003	55	63.0586	0.2518	0.9791	0.0021	0.9760	0.0031	51	79.5164	48.8940	0.9987	0.0001	0.9981	0.0003
PSO-SVR	RAW	26	22.3709	3.1012	0.9885	0.0013	0.9942	0.0006	35	48.0679	5.4254	0.9581	0.0046	0.9679	0.003	39	58.8203	2.3974	0.9869	0.0014	0.9934	0.0007
	SNV	35	2.8113	102.0651	0.9985	0.0002	0.9983	0.0002	37	9.9052	1.7547	0.9808	0.0019	0.9814	0.0024	39	9.5623	40.6473	0.9971	0.0003	0.9986	0.0002
	MSC	29	15.8356	210.1449	0.9975	0.0003	0.9978	0.0002	35	39.7044	83.7515	0.9811	0.0019	0.9813	0.0023	39	15.5483	1000	0.9945	0.0005	0.9962	0.0005
	MMN	32	21.7638	954.3434	0.999	0.0001	0.9980	0.0002	35	21.4624	8.4421	0.9801	0.002	0.9808	0.0023	35	14.4740	436.6765	0.9975	0.0002	0.9984	0.0002
	SG	32	17.7000	2.1661	0.9812	0.002	0.9899	0.0012	39	22.4437	9.6709	0.9598	0.0044	0.9683	0.003	29	63.0797	2.2894	0.9859	0.0015	0.9932	0.0008
	SG-1st	44	86.5797	346.9178	0.9962	0.0004	0.9972	0.0102	47	97.2623	1000	0.9753	0.0025	0.9509	0.0062	255	77.0715	579.9529	0.9929	0.0007	0.9892	0.0015
	SG-2nd	34	99.4144	1000	0.9895	0.0011	0.9909	0.001	54	100	999.6203	0.9446	0.0059	0.6130	0.0547	51	100	1000	0.9793	0.0022	0.9610	0.0096
	SNV-DT	41	32.1123	0.8504	0.9976	0.0003	0.9975	0.0003	55	17.5431	1.0815	0.9802	0.002	0.9811	0.0025	51	10.0567	48.8665	0.9977	0.0002	0.9987	0.0002

Table 1, Table 2, Table 3, Table 4 showed that the prediction accuracy of SG-1st-PLSR model was the highest in the adulteration identification of Tartary buckwheat in Inner Mongolia, with a correlation coefficient R²p of 0.9991 and RMSEP of 0.014. In the adulteration identification of Tartary buckwheat in Sichuan, the prediction accuracy of SNV-DT-CV-SVR model was the highest, with an R²p of 0.9957 and RMSEP of 0.0004. The SNV-DT-CARS-PSO-SVR model had the optimum prediction performance in the adulteration identification of Tartary buckwheat in Shanxi with an R²p of 0.9987 and RMSEP of 0.0002.

A linear-fitting diagram of predicted and true values was prepared for the optimum discrimination models of the three samples, as shown in Fig. 4 and S5. Fig. 4(a) showed the linear-fitting diagram of the predicted and actual values of the SNV-DT-CARS-PSO-SVR model of Shanxi Tartary buckwheat. The slope of the fitting line was close to 1, and the R²p reached 0.9987, indicating a high degree of fitting, which meant that the model had a high potential for adulteration detection. For convenient comparison, Fig. 4(b) showed the linear-fitting diagram of predicted and actual values of Tartary buckwheat in Shanxi based on the full-spectrum raw spectral PSO-SVR model. Compared with Fig. 4(a), the data points on the line were looser, especially the points related to the prediction set. Under the same modeling algorithm, the variable-selection model based on the CARS was superior to the full-spectrum model. Consistent with the results of previous studies, a quality identification model of Cabernet Sauvignon grape with NIRS and CARS-SVR has supported the optimum prediction performance (Luo et al., 2021). It was verified that the CARS-SVR model was an effective method to detect the adulterated concentration of Panax notoginseng flour (Zhang et al., 2022). CARS-SVR model was also established to realize the rapid detection of water content in lettuce leaves (Sun et al., 2017). Therefore, CARS combined with SVR had great potential applications in the quantitative analysis of food quality.

Fig. 4.

Linear-fitting diagram of Shanxi Tartary buckwheat adulteration identification model: (a) SNV-DT-CARS-PSO-SVR model, and (b) RAW-PSO-SVR model.

The prediction effect of the models separately built by PLSR and SVR algorithms on the adulteration degree of Tartary buckwheat were compared. Results showed that the two algorithms were all suitable for predicting the adulteration degree of Tartary buckwheat. However, the prediction performance of SVR model was better in the adulteration analysis of Tartary buckwheat in Sichuan and Shanxi, and the R²p of the optimum model were above 0.99. This finding was due to the ability of the SVR method to solve the problems of small sample, nonlinear, multiple dimensions, and local minimum well and its good generalization ability. It can maximize the reliability of prediction in the case of small sample and obtain the global optimum solution. PLSR was used to establish correction and prediction models by using the full spectral information of samples, which can also obtain higher correlation coefficients and better prediction results. The model built by PLSR was suitable for the adulteration identification of Tartary buckwheat in Inner Mongolia, but the prediction results of this study were relatively poor in the identification of Tartary buckwheat from Sichuan and Shanxi.

3.4. Performance of preprocessing methods and selection of feature wavelengths

The preprocessing methods and the variable screening methods applied to spectral data affected the model performance, as well as the preprocessing methods. In the analysis of Tartary buckwheat in Inner Mongolia, the R²p of the PLSR model established by using SG-1st algorithm was higher than that by other spectral preprocessing methods, and the RMSEP was smaller. SG-1st preprocessing showed excellent modeling effects in most studies (Rukundo et al., 2020), but poor modeling effects in Sichuan Tartary buckwheat and Shanxi Tartary Buckwheat. Derivation in preprocessing can eliminate the influence of baseline offset and gentle background interference and provide higher resolution. However, it also amplified noise and reduced the signal-to-noise ratio. Additionally, for the adulteration identification of Tartary buckwheat from Sichuan and Shanxi, the R²p of the model constructed by SNV-DT was higher than that of other preprocessing methods, and the RMSEP was also smaller. This finding was due to the ability of SNV-DT to effectively eliminate the drift of the spectral curve caused by the distance difference between the optical fiber probe and the sample, consistent with the experimental results of the previous application of preprocessing methods to improve the model's prediction effect (Yi et al., 2017; Bala et al., 2022).

The selection of feature variables also influenced the model's performance. Studies have shown that CARS can effectively extract feature variables to optimize the model, reduce redundant wavelength variables, and improve modeling efficiency (Chen et al., 2019; Basri et al., 2017; Li et al., 2023). However, in previous research, the modeling constructed under the non-full spectrum was not as good as the full spectrum model (Zhao et al., 2019). The preprocessing may have caused the loss of effective information in the spectral data, or the CARS method may not have completely extracted effective information, resulting in insufficient effective information involved in the final modeling.

4. Conclusions

This study used NIRS combined with chemometrics for the quantitative analysis of adulterated common buckwheat in Tartary buckwheat. By collecting the NIRS information of 12 proportions of adulterated Tartary buckwheat samples, the quantitative analysis model of the adulterated degree of Tartary buckwheat was constructed by PLSR and SVR through seven preprocessing methods, CARS feature-variable screening, and three parameter-optimization algorithms. The constructed SG-1st-PLSR, SNV-DT-CV-SVR and SNV-DT-CARS-PSO-SVR model showed excellent prediction accuracy than other models for the adulteration identification of Tartary buckwheat from Inner Mongolia, Sichaun, and Shanxi, respectively. The results of PLSR and SVR modeling for the prediction of Tartary buckwheat adulteration content were satisfactory, and the correlation coefficients of the optimum identification models all exceeded 0.99, indicating that the model of NIRS to determine Tartary buckwheat adulteration degree presented high accuracy and performance. In conclusion, the quantitative model of Tartary buckwheat adulteration established by NIRS and chemometrics can accurately determine the content of adulterated common buckwheat. These results indicated that the method had strong predictive performance and applicability and can be developed to rapidly detect adulterated cereal products. Indeed, future research should improve the universality and practicality of the identification models for Tartary Buckwheat flour adulteration. It is also highlighted to develop new feasible models for practical applications, as well as the intelligent and portable detector based on NIRS discriminant models.

CRediT authorship contribution statement

Yinghui Chai: Data curation, Data collection, Software, Writing – original draft, draft writing and revision. Yue Yu: Methodology, Software, Writing – review & editing, manuscript reviewing, Supervision. Hui Zhu: Data curation, Data collection, Software. Zhanming Li: Supervision, Software, Writing – review & editing, manuscript reviewing. Hao Dong: Software, Writing – review & editing, manuscript reviewing. Hongshun Yang: Software, Writing – review & editing, manuscript reviewing.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Handling Editor: Dr. Maria Corradini

Appendix A. Supplementary data

The following is the Supplementary data to this article:

Data availability

Data will be made available on request.