Next Article in Journal
The Effects of Tea Polyphenol on Chicken Protein Digestion and the Mechanism under Thermal Processing
Previous Article in Journal
Effect of Trehalose/OEO/Tween 80/Tween 20 Addition on Physical Stability of Edible Packaging during Storage in Different Humidity Conditions
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Foods
Volume 12
Issue 15
10.3390/foods12152904
foods-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Rapid Prediction of Adulteration Content in Atractylodis rhizoma Based on Data and Image Features Fusions from Near-Infrared Spectroscopy and Hyperspectral Imaging Techniques

by
Zhiwei Jiang
1,2,†,
Aimin Lv
1,2,†,
Lingjiao Zhong
1,2,
Jingjing Yang
1,2,
Xiaowei Xu
3,
Yuchan Li
1,
Yuchen Liu
1,
Qiuju Fan
1,
Qingsong Shao
1,2 and
Ailian Zhang
1,2,*
1
State Key Laboratory of Subtropical Silviculture, Zhejiang A&F University, Hangzhou 311300, China
2
Zhejiang Provincial Key Laboratory of Resources Protection and Innovation of Traditional Chinese Medicine, Zhejiang A&F University, Hangzhou 311300, China
3
Wenzhou Forestry Technology Promotion and Wildlife Protection Management Station, Wenzhou 325027, China
*
Author to whom correspondence should be addressed.
†
These authors contributed equally to this work.
Foods 2023, 12(15), 2904; https://doi.org/10.3390/foods12152904
Submission received: 4 July 2023 / Revised: 27 July 2023 / Accepted: 28 July 2023 / Published: 30 July 2023
(This article belongs to the Section Food Analytical Methods)
Browse Figures
Versions Notes

Abstract

:
Atractylodis rhizoma (AR) is an herb and food source with great economic, medicinal, and ecological value. Atractylodes chinensis (DC.) Koidz. (AC) and Atractylodes lancea (Thunb.) DC. (AL) are its two botanical sources. The commercial fraud of AR adulterated with Atractylodes japonica Koidz. ex Kitam (AJ) frequently occurs in pursuit of higher profit. To quickly determine the content of adulteration in AC and AL powder, two spectroscopic techniques, near-infrared spectroscopy (NIRS) and hyperspectral imaging (HSI), were introduced. The partial least squares regression (PLSR) algorithm was selected for predictive modeling of AR adulteration levels. Preprocessing and feature variable extraction were used to optimize the prediction model. Then data and image feature fusions were developed to obtain the best predictive model. The results showed that if only single-spectral techniques were considered, NIRS was more suitable for both tasks than HSI techniques. In addition, by comparing the models built after the data fusion of NIRS and HSI with those built by the single spectrum, we found that the mid-level fusion strategy obtained the best models in both tasks. On this basis, combined with the color-texture features, the prediction ability of the model was further optimized. Among them, for the adulteration level prediction task of AC, the best strategy was combining MLF data (at CARS level) and color-texture features (C-TF), at which time the R2T, RMSET, R2P, and RMSEP were 99.85%, 1.25%, 98.61%, and 5.06%, respectively. For AL, the best approach was combining MLF data (at SPA level) and C-TF, with the highest R2T (99.92%) and R2P (99.00%), as well as the lowest RMSET (1.16%) and RMSEP (2.16%). Therefore, combining data and image features from NIRS and HSI is a potential strategy to predict the adulteration content quickly, non-destructively, and accurately.

1. Introduction

Atractylodis rhizoma (AR) is the rhizome of the famous perennial herb Atractylodes lancea (Thunb.) DC. (AL) or Atractylodes chinensis (DC.) Koidz. (AC), which has high economic, medicinal, and ecological value. Wild AR plants grow mainly on mountain slopes in grasslands, forest understories, thickets, and rock crevices. Nowadays, AR is widely cultivated in East and Southeast Asian countries such as China, Japan, Thailand, and Korea. Besides its distinctive aroma and capacity to treat digestive, respiratory, and neurological system illnesses, AR is frequently employed as a spice and flavoring agent in the food industry [1,2,3,4]. Therefore, it is regarded as an important cash crop.
Due to the destruction and overexploitation of its natural habitat, AR is facing a scarcity of wild resources, leading to an increase in its economic value and triggering a series of illegal practices of AR fraud in the market. Atractylodes japonica Koidz. ex Kitam. (AJ) is one of its common counterfeits, which is primarily cultivated in the northeastern provinces of China and sold as AR despite not being listed in the Chinese Pharmacopoeia. AJ and AR have different chemical compositions, prices, and efficacy concerns [5]. Without the right detection technology to curb this fraud, it will inevitably cause confusion in market management and affect the quality and efficacy of AR.
Conventional identification methods such as thin-layer chromatography, high-performance liquid chromatography, and DNA molecular labeling techniques may be able to determine whether AR is adulterated, but they have the disadvantage of being time-consuming, cumbersome, and operator-demanding [6,7,8]. Therefore, in recent years, the search for rapid non-destructive testing techniques has become more vocal. Near-infrared spectroscopy (NIRS) and hyperspectral imaging (HSI) techniques are typical of rapid non-destructive testing techniques, and they are widely used in the fields of food and medicine. NIRS combined with chemometrics can predict the adulteration level of ginger powder, Ganoderma lucidum spore powder, and honeysuckle [9,10,11]. And HSI combined with chemometrics has also been reported to successfully predict the adulteration concentration of oregano, turmeric powder, and saffron [12,13,14]. Additionally, the concepts of data fusion and image feature fusion have been proposed in the fields of food and medicine. Some studies have indicated that data fusion or combining image features can identify the origin and variety of rice, predict the polyphenol content and antioxidant capacity of peppercorns, and detect the soluble solids content in red globe grapes [15,16,17]. The application of NIRS combined with chemometrics has been reported to successfully identify the authenticity, botanical origin, and geographical origin of AR decoction pieces, as well as predict the atractylodin content of bran-fried AR [18,19]. However, there has been no study on the use of NIRS and HSI to predict the adulteration content in AR.
In previous studies, research has typically focused on a single spectral technique, and image features are generally combined with HSI data. In this study, the focus has been extended to data fusion using two spectral techniques and adding image features based on them. Data fusion provides a method to combine NIRS and HSI spectral data based on complementary or synergistic effects, which can usually be divided into low-level fusion (LLF), mid-level fusion (MLF), and high-level fusion (HLF) strategies depending on the data structure [20]. Because HLF is susceptible to severe information loss when forecasting model performance, LLF and MLF were taken into consideration in this work [21]. HSI offers the possibility of extracting image features in addition to spectral data. Color and texture features are generally valuable image features, which has also been confirmed in previous studies [22]. In the present study, based on the superior performance of the data and image feature fusions of NIRS and HSI in previous studies, we speculated that these may also have the potential to be applied to the prediction of adulteration content in AR. To test this hypothesis, we mainly performed the following work: (i) establishing a partial least squares regression (PLSR) prediction model for the adulteration concentration of AC and AL powder doped with AJ; (ii) optimizing the model using pretreatment (SGS, SNV, MSC, 1 Der and 2 Der) and feature variable extraction (CARS, SPA, and GA); (iii) introducing a data and image feature fusion strategy to make the model performance further optimized. Finally, this study establishes a new method for rapid nondestructive detection of the amount of AJ doped in AR, which provides new ideas for the quality evaluation of herbs and food and a theoretical basis for the improvement of the quality evaluation system of AR and is conducive to the benign development of Chinese herbal medicine.

2. Materials and Methods

2.1. Sample Collection and Processing

AC decoction pieces were mainly collected from China province: Anhui (1), Gansu (3), Guangxi (1), Hebei (1), Heilongjiang (5), Jilin (3), Liaoning (4), Inner Mongolia (12), Shaanxi (1), Sichuan (2), Yunnan (1), and Zhejiang (1). AL decoction pieces were collected from Anhui (7), Guangxi (1), Henan (1), Hubei (8), Jiangsu (3), Jiangxi (1), Shaanxi (2), and Zhejiang (1). AJ decoction pieces were collected from Liaoning. All samples were cleaned of impurities, dried, numbered, and powdered, and passed through a 50-mesh sieve in accordance with the adulteration content of 0, 10%, 20%, 30%, 40%, 50%, 70%, and 90% to make a total weight of 2 g. For each level, one sample was prepared, and a sum of 35 pure and 245 adulterated AC powders and 24 pure and 168 adulterated AL powders were obtained.

2.2. Spectral Acquisition and Image Feature Extraction

All samples were scanned by the NIR spectrometer (Antaris™ II, Thermo Fisher Scientific Co., Ltd., Waltham, USA) and hyperspectral imager (GaiaField-N17E, Sichuan Shuangli Hepu Technology Co., Ltd., Chengdu, China). The parameters of the NIR spectrometer were set as follows: the number of spectral scans was 32, and the resolution was 8 cm−1. The parameters of the hyperspectral imager were set as follows: The moving speed and distance of the carrier platform were 1.4 cm s-1 and 30 cm, respectively; the vertical distance between the sample and the lens was 42 cm; and the exposure time was 7 ms. For NIRS, the scan was repeated three times for each sample and averaged for subsequent data analysis. As for HSI, it was scanned once, as in previous studies [21]. Finally, 1557 NIRS data from 4000–10,000 cm−1 and 512 HSI data from 900–1700 nm were acquired.
The raw NIRS data was entered into the Unscrambler X 10.4 software to extract the corresponding spectral values. As shown in Figure 1, before HSI data could be used for further analysis, the following four processes had to be completed: (i) The original hyperspectral image was calibrated to create a calibrated hyperspectral image using Formula (1); (ii) The calibrated hyperspectral image was input and applied to the mask image to identify the region of interest (ROI) and remove outlier pixels; (iii) Two expansion and one erosion operations were performed; (iv) Each image was cropped into sub-images and averaged the reflectance values of all pixels in the sub-images to obtain the average spectrum for each sample. In addition, the study also considered the combination of color and texture features to improve the prediction ability. Color features (CF) were selected from first-order moments, second-order moments, and third-order moments of the three channels of RGB, while texture features (TF) included energy, homogeneity, contrast, and the correlation of each pixel in the Gray-Level Co-occurrence Matrix (GLCM) at three bands (47, 197, 378) and four directions (0°, 45°, 90°, 135°). Masked and filtered binary images were used to calculate the morphological features. A total of 48 texture features were extracted from the new hypercube by the GLCM algorithm, and 9 color features were calculated from the three RGB channels (Table S1) [23,24].
R C = R O − R B R W − R B
Note: RB = the all-black background image; RC = the corrected image; RO = the original image; RW = the all-white calibration plate image.

2.3. Preprocessing and Feature Variable Extraction

To remove systematic noise and outside environmental influences during sample collection, as well as to reduce the substantial amount of redundant and useless spectral data information in the full wavelength, simplify the model, and enhance its predictive power, pretreatment and feature variable extraction were required [25]. As previous studies, five pretreatment methods such as Savitzky-Golay smoothing (SGS), standard normalized variate (SNV), multiplicative scatter correction (MSC), the first derivative (1 Der), and the second derivative (2 Der) were chosen. Among them, SGS is used for smoothing filtering, which reduces the interference of noise on the sampled signal. The inhomogeneity of the sample causes light scattering, which leads to errors in the sample spectrum, and SNV can remove additive and multiplicative effects from the spectrum. MSC is mainly applied to eliminate scattering effects caused by inhomogeneity in particle distribution and particle size. Interference caused by baseline drift or soft background can be eliminated by 1 Der and 2 Der algorithms, as can overlapping peaks, to improve resolution and sensitivity [26].
In addition, three feature variable extraction methods, such as competitive adaptive reweighted sampling (CARS), successive projection algorithms (SPA), and genetic algorithms (GA), were selected [15]. CARS is an efficient wavelength selection algorithm based on the principle of “survival of the fittest.” CARS is designed to select critical wavelengths through a rigorous and computationally efficient procedure. Two consecutive wavelength selection steps are performed: in the first step, wavelengths with relatively small PLSR coefficients are forcibly removed using an exponential decay function. Next, variable sampling with adaptive reweighting is used to further eliminate wavelengths in a competitive manner. SPA is a forward variable selection algorithm used for multivariate calibration to select wavelengths with minimal redundancy. It performs simple projection operations in vector space to obtain a subset of useful variables with minimal covariance. The principle of variable selection via SPA is that the newly selected variable is the one with the largest projection on the orthogonal subspace of the previously selected variable. The optimal initial variables and the number of variables can be determined based on the minimum root-mean-square error of cross-validation. GA is a global optimization search method inspired by Darwin’s theory of natural selection. It selects variables that are better suited to the fitness function through the manipulation of genetic processes, such as reproduction, mutation, and selection, and successive genetic iterations. In this study, the parameter values of the GA were set based on preliminary tests: population size (30), window width (3), penalty slope (0), maximum generations (100), mutation rate (0.01), crossover probability (0.5), and replicate runs (100) [27].

2.4. Data and Image Feature Fusions

The NIRS and HSI data are directly stitched together to obtain the LLF dataset, and the LLF is usually used in conjunction with the feature variable extraction method because more information is introduced to interpret the sample. The features extracted by the same feature filtering method are integrated into a new MLF dataset, which effectively avoids the drawbacks of LLF by integrating data from two sources without causing a large increase in data [28]. NIRS, HSI, and LLF would go through five pretreatments and three feature variable selections to filter out the best combination of their respective methods, whereas MLF built an optimal PLSR model by fusing the NIRS and HSI feature variables extracted by the same feature variable extraction methods based on the best pretreatment (Figure 2). CF, TF, or color-texture features (C-TF) would be added to the best PLSR models based on NIRS, HSI, LLF, or MLF to obtain the optimal model.

2.5. Data Set Partitioning and Quantitative Analysis Methods

The joint x-y distance (SPXY) algorithm uses a partitioning algorithm that accepts the variability of x-space and y-space. In contrast to partitioning systems based just on x information or random sampling, the multidimensional space can be covered more effectively with this approach [29]. Therefore, SPXY was proposed to divide the dataset. The dataset was divided into training and prediction sets in the ratio of 4:1. For AC doped concentration prediction, the numbers of samples in the training and prediction sets were 672 and 168, respectively, while for AL, the numbers were 461 and 115, respectively. PLSR is a common modeling method for quantitative spectral analysis that effectively solves the issue of multiplicity correlation between variables and incorporates the principles and characteristics of principal component analysis, multiple linear regression, and conventional correlation analysis. Because of its forceful predictive power, the PLSR algorithm was chosen as the quantitative prediction model [30,31]. The choice of principal component number affects the modeling effect of PLSR; therefore, 10-fold cross-validation was proposed to select the optimal principal component number to minimize the root mean square error (RMSE). In this study, the performance of the model was assessed mainly by the correlation coefficient of training sets (R2T), the correlation coefficient of prediction sets (R2P), the root mean square error of training sets (RMSET), and the root mean square error of prediction sets (RMSEP). A good model ought to have a low RMSE and a high R2. In addition, the principal component number should be as small as possible, because a large principal component number may introduce some irrelevant information and cause overfitting of the model [32].

2.6. Software

The Unscrambler X 10.4 was used for NIRS data format conversion, and SpecView was used for hyperspectral image capture and calibration. The rest of the data processing was conducted on MATLAB R2022a.

3. Results and Discussion

3.1. Sample and Spectral Analysis

The color, texture, and size of the powder particles were relatively similar between pure and adulterated AR powder, so it was challenging to quickly identify them from their appearance (Figure 3). Unsurprisingly, the raw spectrograms did not seem to work well either. The curve trends of pure and adulterated samples were similar, and the peaks appeared at the same position and height, whether in the HSI or NIRS spectrogram (Figure 4), which was consistent with previous studies [33,34]. Therefore, it was necessary to perform specific processing on the raw spectra to achieve a fast and accurate prediction of the adulteration level.

3.2. Quantitative Analysis Based on NIRS Data and Image Features

For predicting the adulteration concentration of AC samples, the PLSR model based on the original data obtained high R2T (95.32%) and R2P (95.00%) and acceptable RSMET (8.63%) and RMSEP (8.89%) (Table 1). All five preprocessing methods improved the model’s performance to different degrees. Among them, 1 Der obtained the best result since the model had higher R2T (99.77%) and R2P (97.61%) as well as lower RMSET (1.94%) and RMSEP (6.60%). 2 Der was not considered because its R2P (87.45%) and RMSEP (15.53%) were suboptimal, even though its R2T (99.98%) and RMSET (0.60%) were excellent, while the feature variable extraction degraded the model’s performance to different degrees. Although CARS only degraded the model slightly, the modeling efficiency increased substantially; hence, CARS was chosen as the best feature variable extraction method, at which time the model’s R2T, R2P, RMSET, and RMSEP were 99.06%, 3.87%, 98.50%, and 5.14%, respectively. The principal component score selection was obtained by plotting the RMSE of Y with the principal components. In Figure 5, the smallest RMSE was obtained by cross-validation when the number of principal components was 20 (when the number of principal components was between 10 and 20, the RMSE value showed a downward trend and was basically stable after 20, so 20 was selected). Since previous studies have proved that HSI and color feature fusion had good modeling efficacy for whole wheat flour samples with different DON levels and that introducing a large number of features without reducing the data dimension may complicate the model, CF, TF, and C-TF were added after the feature variables were extracted [34,35]. It was obvious that the best performance of the model was obtained by combining CT. At this time, R2T, RMSET, R2P, and RMSEP were 99.06%, 3.87%, 98.50%, and 5.14%, respectively. Only 85 features were combined; compared with the original data, the features were reduced by 94.54%, while R2T and R2P were improved by 3.74% and 3.50%, respectively, and RMSET and RMSEP were reduced by 4.76% and 3.75%, respectively.
For predicting the adulteration level of AL, the PLSR model built from the raw NIRS data had R2T, RMSET, R2P, and RMSEP of 97.81%, 5.93%, 92.45%, and 11.73%, respectively (Table 1). The preprocessing did not optimize the model in all cases, especially with 2 Der, where the R2P was reduced by 24.29% and the RMSEP increased by 9.36%. While previous studies found significant performance gains in models built by choosing 2 Der as preprocessing, our study reached the opposite results [9]. A reasonable explanation was that the signal-to-noise ratio decreases as the derivative increases, and the spectral information may be lost [36]. The best pretreatment choice was SGS, where the R2T, RMSET, R2P, and RMSEP were 99.66%, 2.24%, 97.60%, and 7.08%, respectively. The best feature selection method was SPA, where 95 feature variables were filtered, a reduction of 1462 compared to the full wavelength. However, neither feature variable selection nor combining image features led to improved model performance, presumably because some of the useful information was eliminated and the algorithm utilized different information with different efficiency [37].

3.3. Quantitative Analysis Based on HSI Data and Image Features

The models based on HSI data performed worse than those based on NIRS. For predicting the adulteration content of the adulterated AC, the R2T, RMSET, R2P, and RMSEP of the model using the original data were only 81.92%, 16.39%, 79.99%, and 17.24%, and after preprocessing with 2 Der, the R2 could exceed 88% (Table 1). The best feature variable extraction method was SPA, at which time the R2T could reach 90.06%. Although R2P was only 85.59%, it was still the highest among the three methods. We tried to combine image features on the basis of 2 Der + SPA and found that the model results combining three image features were in the order of good to bad: 2 Der + SPA + CF > 2 Der + SPA + T-CF > 2 Der + SPA + TF. Similar results were found in the classification task [38], and it might be that the TF carries some useful information and the others may not contribute much to the modeling. Although the R2 of the model constructed by 2 Der + SPA + CF was higher than 86%, it was still far from the best PLSR model based on NIRS data.
Interestingly, in predicting the adulteration level in AL, the model built with original data gave better results than that built with pretreatments, except for 2 Der. After 2 Der processing, the R2T, RMSET, R2P, and RMSEP of the model were 99.12%, 3.69%, 91.66%, and 11.58%, respectively, which changed by 9.46%, 8.56%, 2.89%, and 4.03% compared to the original data (Table 1). In addition, we found that 2 Der was the best choice for predicting the adulteration level of either AC or AL based on HSI data. In the detection of metanil yellow adulteration in chickpea flour based on the HSI full spectrum, the best pretreatment was also 2 Der, and our findings were consistent with this [33]. This might be related to the ability of 2 Der to separate overlapping peaks, which may lead to the success prediction [39]. The best feature variable selection method was then SPA, at which point it filtered to 94 feature wavelengths. Combining TF on the basis of this seemed to be the best option. However, in general, feature variable extraction and combining TF did not lead to a significant improvement in model performance. Furthermore, the model built from HSI data seemed to be worse than the one built from NIRS, which was explainable because HSI shows its unique spatial resolution at the expense of spectral resolution compared to conventional NIRS spectroscopy [40]. However, when NIRS is combined with the C-TF method extracted from HSI, the complementary advantages of the two enable the model performance to be further optimized, which gave us the possibility to think about whether a better model can be obtained by fusing the two data and then combining the image features.

3.4. Quantitative Analysis Based on LLF Data and Image Features

The LLF dataset of 2069 features was obtained by integrating 1557 variables in the NIRS and 512 variables in the HSI. For the adulteration content prediction of adulterated AC, 1 Der was the best pretreatment choice, at which time R2T, RMSET, R2P, and RMSEP were 99.92%, 1.15%, 97.71%, and 6.73%, respectively (Table 1). 117 feature variables were obtained after SPA extraction of feature variables, which was 94.35% less compared to total variables, but there was a slight decrease in the performance of the model. There was further degradation in the performance of the model if image features were introduced.
For predicting the adulteration level of AL, its combination with the original data could obtain 98.16%, 5.47%, 91.95%, and 12.29% for R2T, RMSET, R2P, and RMSEP (Table 1). After pretreatment, SGS, SNV, MSC, and 1 Der could optimize the model, while 2 Der did the opposite. Among them, the model performed best after SGS treatment compared with the original data, with R2T and R2P increasing by 1.65% and 5.24%, respectively, while RMSET and RMSEP decreasing by 3.72% and 5.92%, respectively. However, similar to the AC prediction task, both feature variable selection and combining image features made the modeling worse. LLF did not seem to be suitable for combining image features to accomplish the above two prediction tasks. In addition, the models using LLF exhibited similar or poorer performance compared to models built with a single spectrum. However, in the previous study, the LLF strategy had better predictive ability in predicting TVB-N content in chicken than the optimal model for single-spectral data. Our results yielded different conclusions from this, which could be attributed to the large amount of data introduced by LLF, leading to the computational complexity and uncertainty of the PLSR method [21,41].

3.5. Quantitative Analysis Based on MLF Data and Image Features

Based on the best pretreatment, the MLF dataset was obtained by integrating the feature variables obtained by the same feature variable extraction method. For predicting the adulteration content of AC, the model built at the CARS level was the best, with R2T, RMSET, R2P, and RMSEP of 99.15%, 3.61%, 98.17%, and 6.55%, respectively (Table 1). If image features were considered, C-TF was the best choice because the R2T, RMSET, R2P, and RMSEP were 99.85%, 1.25%, 98.61%, and 5.06%, respectively, which was superior to the best model built based on NIRS data.
For predicting the adulteration content in AL, the model built at the SPA level was the best, with R2T, RMSET, R2P, and RMSEP of 99.62%, 2.37%, 96.22%, and 11.00%, respectively (Table 1). Similarly, the combination of C-TF on this basis also achieved the best results, with R2T, RMSET, R2P, and RMSEP reaching 99.92%, 1.16%, 99.00%, and 2.16%, better than the best models built with single spectral or LLF data. In general, the MLF strategy achieved very good results, which was consistent with the findings of previous studies [15].

3.6. Analysis of Feature Variables

The characteristic variables selected for the MLF strategy were analyzed to find the reasons for the success of the adulterated content prediction. In Figure 6A, the feature variables were mainly distributed in the ranges of 4200–5300 cm−1 and 8200–9800 cm−1. Unfortunately, there is little information in the scientific literature about the NIRS of AR. However, an attempt would be made to identify the compounds that may be present in the NIRS of AR. We speculated that the feature variables selected between 4200 and 5300 cm−1 may be related to the absorption valleys near 4200, 4600, and 5000 cm−1. The variables near 4600 and 5000 cm−1 may be connected to the combined C-H and N-H bonding vibrations found in proteins. Those near 4200 cm−1 could be attributed to the C-H and C-C bonding vibrations [42]. The region between 8200 and 9800 cm−1 was associated with the second overtone vibrations of C-H and N-H [43]. In Figure 6B, the characteristic variables were mainly distributed between 1140 and 1500 nm, which may be related to the first-order octave vibrations of C-H and N-H.
Intriguingly, the feature variables in Figure 6C were primarily dispersed between the ranges of 5200–7300 and 8200–9800 cm−1. The absorption band of 5200–7300 cm−1 might be due to the amide combination band vibration of CONH2 and the N-H stretching of the protein [44]. When it came to feature variables of HSI data processed by 2 Der and SPA (Figure 6D), selected feature variables concentrated in 1000–1180, 1220–1340, and 1450–1600 nm could be observed. Feature variables in the 1000–1180 nm range may be influenced by the second overtone of N-H or O-H [36]. In summary, proteins may affect the effectiveness of model modeling, and the specific substances still need further in-depth study.

4. Conclusions

To predict the adulteration content of AC and AL powder quickly and nondestructively, two spectroscopic techniques, NIRS and HSI, were developed. Preprocessing, feature variable selection, data fusion, and image feature fusion strategies were introduced to obtain the optimal prediction model. The results suggested that a single NIRS might be more appropriate than a single HSI, but if higher prediction results were pursued, adding C-TF data to the MLF would be a better choice. To be specific, for the task of predicting the adulteration level of AC, the best strategy was combining MLF data (at CARS level) and C-TF, at which time the R2T, RMSET, R2P, and RMSEP of the model were 99.85%, 1.25%, 98.61%, and 5.06%, respectively. For another task, the best approach was adding C-TF to the basic MLF data (at SPA level), where the R2T, RMSET, R2P, and RMSEP were 99.92%, 1.16%, 99.00%, and 2.16%, respectively. In addition, we found that proteins may be one of the factors affecting successful modeling, but the exact substance still needs further study. The instruments used in this study are suitable for analytical testing in the laboratory; therefore, the use of low-cost and convenient spectroscopic equipment for quality assessment in the field or considering applying a demixing algorithm to obtain a better model may be the focus of future work. In conclusion, the data and image feature fusions based on NIRS and HSI can predict the level of adulteration of AR powder quickly, nondestructively, and accurately, which is beneficial to safeguarding the quality and efficacy of herbs as well as providing a theoretical basis and new ideas for quality evaluation of Chinese herbs.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/foods12152904/s1.

Author Contributions

Conceptualization, A.L. and A.Z.; methodology, Z.J.; software, A.L. and L.Z.; validation, A.L. and L.Z.; formal analysis, Z.J.; investigation, Z.J.; resources, A.Z.; data curation, Z.J.; writing—original draft preparation, Z.J. and A.L.; writing—review and editing, A.L., L.Z., J.Y., X.X., Y.L. (Yuchan Li), Y.L. (Yuchen Liu) and Q.F.; visualization, Z.J., A.L. and L.Z.; supervision, A.L. and A.Z.; project administration, A.Z.; funding acquisition, Q.F. and Q.S. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the Zhejiang Provincial Key Laboratory of Resources Protection and Innovation of Traditional Chinese Medicine (2021E10013) and the Student Research Training Project of Zhejiang Agricultural and Forestry University (2023KX141).

Data Availability Statement

The data in this study were available from the following sources: the corresponding authors. These data are not publicly available due to the requirement to fund research projects.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Liu, C.; Song, C.; Wang, Y.; Xiao, Y.; Zhou, Z.; Cao, G.; Sun, X.; Liu, Y. Deep-fried Atractylodes lancea rhizome alleviates spleen deficiency diarrhea–induced short-chain fatty acid metabolic disorder in mice by remodeling the intestinal flora. J. Ethnopharmacol. 2023, 303, 115967. [Google Scholar] [CrossRef] [PubMed]
  2. Zhang, Q.; Yang, C.; Ma, S.; Guo, S.; Hu, X.; Zhou, Z.; Liu, Y.; Zhang, X.; Jiang, R.; Zhang, Z.; et al. Shiwei Qingwen decoction regulates TLR4/NF-κB signaling pathway and NLRP3 inflammasome to reduce inflammatory response in lipopolysaccharide-induced acute lung injury. J. Ethnopharmacol. 2023, 313, 116615. [Google Scholar] [CrossRef]
  3. Li, C.; Wang, C.; Guo, Y.; Wen, R.; Yan, L.; Zhang, F.; Gong, Q.; Yu, H. Research on the effect and underlying molecular mechanism of Cangzhu in the treatment of gouty arthritis. Eur. J. Pharmacol. 2022, 927, 175044. [Google Scholar] [CrossRef]
  4. Xu, W.; Lu, R.; Li, J.; Xia, M.; Chen, G.; Li, P. Comparative plastome analyses and evolutionary relationships of all species and cultivars within the medicinal plant genus Atractylodes. Ind. Crops Prod. 2023, 201, 116974. [Google Scholar] [CrossRef]
  5. Liu, Y.; Zhang, B.; Cai, Q. Study on the pharmacodynamics and metabolomics of five medicinal species in Atractylodes DC. on rats with rheumatoid arthritis. Biomed. Pharmacother. 2020, 131, 110554. [Google Scholar] [CrossRef] [PubMed]
  6. Gu, X.; Jin, Y.; Dong, F.; Cai, Y.; You, Z.; You, J.; Zhang, L.; Du, S. Toward rapid analysis, forecast and discovery of bioactive compounds from herbs by jointly using thin layer chromatography and ratiometric surface-enhanced Raman spectroscopy technique. J. Pharm. Biomed. Anal. 2018, 153, 9–15. [Google Scholar] [CrossRef] [PubMed]
  7. Zhao, Y.; Chang, X.; Gu, X.; Li, Y.; Zheng, Y.; Fang, H. Predictive analysis of quality markers of Atractylodis Rhizoma based on fingerprint and network pharmacology. J. AOAC Int. 2023, 106, qsad059. [Google Scholar] [CrossRef]
  8. Xin, T.; Xu, Z.; Jia, J.; Leon, C.; Hu, S.; Lin, Y.; Ragupathy, S.; Song, J.; Newmaster, S.G. Biomonitoring for traditional herbal medicinal products using DNA metabarcoding and single molecule, real-time sequencing. Acta Pharm. Sin. B 2018, 8, 488–497. [Google Scholar] [CrossRef]
  9. Yu, D.; Guo, S.; Zhang, X.; Yan, H.; Zhang, Z.; Chen, X.; Chen, J.; Jin, S.; Yang, J.; Duan, J. Rapid detection of adulteration in powder of ginger (Zingiber officinale Roscoe) by FT-NIR spectroscopy combined with chemometrics. Food Chem. X 2022, 15, 100450. [Google Scholar] [CrossRef]
  10. Shi, X.; Gan, X.; Wang, X.; Peng, J.; Li, Z.; Wu, X.; Shao, Q.; Zhang, A. Rapid detection of Ganoderma lucidum spore powder adulterated with dyed starch by NIR spectroscopy and chemometrics. LWT 2022, 167, 113829. [Google Scholar] [CrossRef]
  11. Yang, H.; Bao, L.; Liu, Y.; Luo, S.; Zhao, F.; Chen, G.; Liu, F. Identification and quantitative analysis of salt-adulterated honeysuckle using infrared spectroscopy coupled with multi-chemometrics. Microchem. J. 2021, 171, 106829. [Google Scholar] [CrossRef]
  12. Steene, J.; Ruyssinck, J.; Fernandez-Pierna, J.; Vandermeersch, L.; Maes, A.; Langenhove, H.; Walgraeve, C.; Demeestere, K.; Meulenaer, B.; Jacxsens, L.; et al. Authenticity analysis of oregano: Development, validation and fitness for use of several food fingerprinting techniques. Food Res. Int. 2022, 162, 111962. [Google Scholar] [CrossRef]
  13. Hashemi-Nasab, F.S.; Talebian, S.; Parastar, H. Multiple adulterants detection in turmeric powder using Vis-SWNIR hyperspectral imaging followed by multivariate curve resolution and classification techniques. Microchem. J. 2023, 185, 108203. [Google Scholar] [CrossRef]
  14. Hashemi-Nasab, F.S.; Parastar, H. Vis-NIR hyperspectral imaging coupled with independent component analysis for saffron authentication. Food Chem. 2022, 393, 133450. [Google Scholar] [CrossRef] [PubMed]
  15. Steene, J.; Ruyssinck, J.; Fernandez-Pierna, J.; Vandermeersch, L.; Maes, A.; Langenhove, H.; Walgraeve, C.; Demeestere, K.; Meulenaer, B.; Jacxsens, L.; et al. Fingerprinting methods for origin and variety assessment of rice: Development, validation and data fusion experiments. Food Control 2023, 151, 109780. [Google Scholar] [CrossRef]
  16. Esquerre, C.A.; Achata, E.M.; García-Vaquero, M.; Zhang, Z.; Tiwari, B.K.; O’Donnell, C.P. Use of an NIR MEMS spectrophotometer and visible/NIR hyperspectral imaging systems to predict quality parameters of treated ground peppercorns. LWT 2020, 131, 109761. [Google Scholar] [CrossRef]
  17. Ye, W.; Xu, W.; Yan, T.; Yan, J.; Gao, P.; Zhang, C. Application of near-infrared spectroscopy and hyperspectral imaging combined with machine learning algorithms for quality inspection of grape: A Review. Foods 2023, 12, 132. [Google Scholar] [CrossRef]
  18. Jiang, Z.; Jin, K.; Zhong, L.; Zheng, Y.; Shao, Q.; Zhang, A. Near-infrared spectroscopy combined with machine learning for rapid identification of Atractylodis rhizoma decoction pieces. Ind. Crops Prod. 2023, 197, 116579. [Google Scholar] [CrossRef]
  19. Lei, L.; Ke, C.; Xiao, K.; Qu, L.; Lin, X.; Zhan, X.; Tu, J.; Xu, K.; Liu, Y. Identification of different bran-fried Atractylodis Rhizoma and prediction of atractylodin content based on multivariate data mining combined with intelligent color recognition and near-infrared spectroscopy. Spectrochim. Acta. A Mol. Biomol. Spectrosc. 2021, 262, 120119. [Google Scholar] [CrossRef] [PubMed]
  20. Wang, H.P.; Chen, P.; Dai, J.W.; Liu, D.; Li, J.Y.; Xu, Y.P.; Chu, X.L. Recent advances of chemometric calibration methods in modern spectroscopy: Algorithms, strategy, and related issues. TrAC Trends Anal. Chem. 2022, 153, 116648. [Google Scholar] [CrossRef]
  21. Li, X.; Cai, M.; Li, M.; Wei, X.; Liu, Z.; Wang, J.; Jia, K.; Han, Y. Combining Vis-NIR and NIR hyperspectral imaging techniques with a data fusion strategy for the rapid qualitative evaluation of multiple qualities in chicken. Food Control 2023, 145, 109416. [Google Scholar] [CrossRef]
  22. Jiang, X.; Tian, J.; Huang, H.; Hu, X.; Han, L.; Huang, D.; Luo, H. Nondestructive visualization and quantification of total acid and reducing sugar contents in fermented grains by combining spectral and color data through hyperspectral imaging. Food Chem. 2022, 386, 132779. [Google Scholar] [CrossRef]
  23. Rady, A.M.; Adedeji, A.; Watson, N.J. Feasibility of utilizing color imaging and machine learning for adulteration detection in minced meat. J. Agric. Food Res. 2021, 6, 100251. [Google Scholar] [CrossRef]
  24. ElManawy, A.; Sun, D.; Abdalla, A.; Zhu, Y.; Cen, H. HSI-PP: A flexible open-source software for hyperspectral imaging-based plant phenotyping. Comput. Electron. Agric. 2022, 200, 107248. [Google Scholar] [CrossRef]
  25. He, H.; Chen, Y.; Li, G.; Wang, Y.; Ou, X.; Guo, J. Hyperspectral imaging combined with chemometrics for rapid detection of talcum powder adulterated in wheat flour. Food Control 2023, 144, 109378. [Google Scholar] [CrossRef]
  26. Dong, J.E.; Li, J.; Liu, H.; Wang, Y.Z. Machine learning and deep learning based on the small FT-MIR dataset for fine-grained sampling site recognition of boletus tomentipes. Food Res. Int. 2023, 167, 112679. [Google Scholar] [CrossRef] [PubMed]
  27. Xu, S.; Zhao, Y.; Wang, M.; Shi, X. Determination of rice root density from Vis–NIR spectroscopy by support vector machine regression and spectral variable selection techniques. Catena 2017, 157, 12–23. [Google Scholar] [CrossRef]
  28. Yu, H.D.; Qing, L.W.; Yan, D.T.; Xia, G.; Zhang, C.; Yun, Y.H.; Zhang, W. Hyperspectral imaging in combination with data fusion for rapid evaluation of tilapia fillet freshness. Food Chem. 2021, 348, 129129. [Google Scholar] [CrossRef]
  29. Wu, S.; Wang, L.; Zhou, G.; Liu, C.; Ji, Z.; Li, Z.; Li, W. Strategies for the content determination of capsaicin and the identification of adulterated pepper powder using a hand-held near-infrared spectrometer. Food Res. Int. 2023, 163, 112192. [Google Scholar] [CrossRef] [PubMed]
  30. Wang, C.; Sun, Y.; Zhou, Y.; Cui, Y.; Yao, W.; Yu, H.; Guo, Y.; Xie, Y. Dynamic monitoring oxidation process of nut oils through Raman technology combined with PLSR and RF-PLSR model. LWT 2021, 146, 111290. [Google Scholar] [CrossRef]
  31. Zhu, J.; Fan, X.; Han, L.; Zhang, C.; Wang, J.; Pan, L.; Tu, K.; Peng, J.; Zhang, M. Quantitative analysis of caprolactam in sauce-based food using infrared spectroscopy combined with data fusion strategies. J. Food Compos. Anal. 2021, 104, 104130. [Google Scholar] [CrossRef]
  32. Tian, S.; Liu, W.; Xu, H. Improving the prediction performance of soluble solids content (SSC) in kiwifruit by means of near-infrared spectroscopy using slope/bias correction and calibration updating. Food Res. Int. 2023, 170, 112988. [Google Scholar] [CrossRef]
  33. Saha, D.; Senthilkumar, T.; Singh, C.B.; Manickavasagan, A. Quantitative detection of metanil yellow adulteration in chickpea flour using line-scan near-infrared hyperspectral imaging with partial least square regression and one-dimensional convolutional neural network. J. Food Compos. Anal. 2023, 120, 105290. [Google Scholar] [CrossRef]
  34. Torres, I.; Sánchez, M.; Vega-Castellote, M.; Pérez-Marín, D. Fraud detection in batches of sweet almonds by portable near-infrared spectral devices. Foods 2021, 10, 1221. [Google Scholar] [CrossRef] [PubMed]
  35. Zhao, T.; Chen, M.; Jiang, X.; Shen, F.; He, X.; Fang, Y.; Liu, Q.; Hu, Q. Integration of spectra and image features of Vis/NIR hyperspectral imaging for prediction of deoxynivalenol contamination in whole wheat flour. Infrared Phys. Technol. 2020, 109, 103426. [Google Scholar] [CrossRef]
  36. Azadnia, R.; Rajabipour, A.; Jamshidi, B.; Omid, M. New approach for rapid estimation of leaf nitrogen, phosphorus, and potassium contents in apple-trees using Vis/NIR spectroscopy based on wavelength selection coupled with machine learning. Comput. Electron. Agric. 2023, 207, 107746. [Google Scholar] [CrossRef]
  37. An, H.; Zhai, C.; Zhang, F.; Ma, Q.; Sun, J.; Tang, Y.; Wang, W. Quantitative analysis of Chinese steamed bread staling using NIR, MIR, and Raman spectral data fusion. Food Chem. 2023, 405, 134821. [Google Scholar] [CrossRef]
  38. Xu, P.; Fu, L.; Xu, K.; Sun, W.; Tan, Q.; Zhang, Y.; Zha, X.; Yang, R. Investigation into maize seed disease identification based on deep learning and multi-source spectral information fusion techniques. J. Food Compos. Anal. 2023, 119, 105254. [Google Scholar] [CrossRef]
  39. Khamsopha, D.; Woranitta, S.; Teerachaichayut, S. Utilizing near infrared hyperspectral imaging for quantitatively predicting adulteration in tapioca starch. Food Control 2021, 123, 107781. [Google Scholar] [CrossRef]
  40. Pudełko, A.; Chodak, M.; Roemer, J.; Uhl, T. Application of FT-NIR spectroscopy and NIR hyperspectral imaging to predict nitrogen and organic carbon contents in mine soils. Measurement 2020, 164, 108117. [Google Scholar] [CrossRef]
  41. Cai, Z.; Huang, Z.; He, M.; Li, C.; Qi, H.; Peng, J.; Zhou, F.; Zhang, C. Identification of geographical origins of Radix Paeoniae Alba using hyperspectral imaging with deep learning-based fusion approaches. Food Chem. 2023, 422, 136169. [Google Scholar] [CrossRef] [PubMed]
  42. Castro, R.C.; Ribeiro, D.S.M.; Santos, J.L.M.; Páscoa, R.N.M.J. Authentication/discrimination, identification and quantification of cinnamon adulterants using NIR spectroscopy and different chemometric tools: A tutorial to deal with counterfeit samples. Food Control 2023, 147, 109619. [Google Scholar] [CrossRef]
  43. Assis, C.; Gama, E.M.; Nascentes, C.C.; Oliveira, L.S.; Anzanello, M.J.; Sena, M.M. A data fusion model merging information from near infrared spectroscopy and X-ray fluorescence. Searching for atomic-molecular correlations to predict and characterize the composition of coffee blends. Food Chem. 2020, 325, 126953. [Google Scholar] [CrossRef] [PubMed]
  44. Arndt, M.; Rurik, M.; Drees, A.; Bigdowski, K.; Kohlbacher, O.; Fischer, M. Comparison of different sample preparation techniques for NIR screening and their influence on the geographical origin determination of almonds (Prunus dulcis MILL.). Food Control 2020, 115, 107302. [Google Scholar] [CrossRef]
Foods 12 02904 g001 550
Figure 1. Schematic diagram of hyperspectral data and the image feature extraction process.
Figure 1. Schematic diagram of hyperspectral data and the image feature extraction process.
Foods 12 02904 g001
Foods 12 02904 g002 550
Figure 2. Schematic diagram of data fusion strategy.
Figure 2. Schematic diagram of data fusion strategy.
Foods 12 02904 g002
Foods 12 02904 g003 550
Figure 3. Samples of AR. (A) pure AC powder; (B) adulterated AC powder (50% adulteration content); (C) AJ powder; (D) pure AL powder; (E) adulterated AL powder (50% adulteration content). Scale bar = 1 cm.
Figure 3. Samples of AR. (A) pure AC powder; (B) adulterated AC powder (50% adulteration content); (C) AJ powder; (D) pure AL powder; (E) adulterated AL powder (50% adulteration content). Scale bar = 1 cm.
Foods 12 02904 g003
Foods 12 02904 g004 550
Figure 4. HSI ((A): AC; (C): AL) and NIRS ((B): AC; (D): AL) spectrograms of pure and adulterated AC and AL (adulteration concentrations: 0, 10%, 20%, 30%, 40%, 50%, 70%, 90%). Note: AC from Anhui, Gansu, Guangxi, Hebei, Heilongjiang, Jilin, Liaoning, Inner Mongolia, Shaanxi, Sichuan, Yunnan, and Zhejiang; AL from Anhui, Guangxi, Henan, Hubei, Jiangsu, Jiangxi, Shaanxi, and Zhejiang.
Figure 4. HSI ((A): AC; (C): AL) and NIRS ((B): AC; (D): AL) spectrograms of pure and adulterated AC and AL (adulteration concentrations: 0, 10%, 20%, 30%, 40%, 50%, 70%, 90%). Note: AC from Anhui, Gansu, Guangxi, Hebei, Heilongjiang, Jilin, Liaoning, Inner Mongolia, Shaanxi, Sichuan, Yunnan, and Zhejiang; AL from Anhui, Guangxi, Henan, Hubei, Jiangsu, Jiangxi, Shaanxi, and Zhejiang.
Foods 12 02904 g004
Foods 12 02904 g005 550
Figure 5. Plot of the RMSE of Y with principal components. Note: There was the SGS + PLSR model for adulterated AC based on NIRS data.
Figure 5. Plot of the RMSE of Y with principal components. Note: There was the SGS + PLSR model for adulterated AC based on NIRS data.
Foods 12 02904 g005
Foods 12 02904 g006 550
Figure 6. Feature variables selected by the MLF strategy. Note: (A,B) denote the NIRS (1 Der + CARS) and HSI datasets (2 Der + CARS) used to predict the adulteration level of adulterated AC powder, respectively. (C,D) denote the NIRS (SGS + SPA) and HSI datasets (2 Der + SPA) used to predict the adulteration level of adulterated AL powder, respectively.
Figure 6. Feature variables selected by the MLF strategy. Note: (A,B) denote the NIRS (1 Der + CARS) and HSI datasets (2 Der + CARS) used to predict the adulteration level of adulterated AC powder, respectively. (C,D) denote the NIRS (SGS + SPA) and HSI datasets (2 Der + SPA) used to predict the adulteration level of adulterated AL powder, respectively.
Foods 12 02904 g006
Table
Table 1. PLSR Modeling results under different methods of processing.
Table 1. PLSR Modeling results under different methods of processing.
ObjectDataMethodPrincipal
Components
Number		Variables
Number		R2T (%)RMSET (%)R2P (%)RMSEP (%)
Adulterated ACNIRSRaw data21155795.328.6395.008.89
SGS20155797.216.7996.826.83
SNV21155797.726.2095.887.51
MSC20155797.765.8296.907.73
1 Der15155799.771.9497.616.60
2 Der9155799.980.6087.4515.53
1 Der + CARS107699.043.9197.346.88
1 Der + SPA47493.969.9386.4914.25
1 Der + GA72495.878.4590.339.90
1 Der + CARS + CF158599.063.8798.505.14
1 Der + CARS + TF3212498.994.0998.075.31
1 Der + CARS + C-TF4213398.964.1497.775.81
HSIRaw data2251281.9216.3979.9917.24
SGS2151281.6816.3981.4217.15
SNV2151283.7115.3077.1719.58
MSC2051283.6915.7081.2716.47
1 Der1251286.9014.3184.4414.62
2 Der951288.2512.9788.3515.39
2 Der + CARS52985.6614.3285.0615.78
2 Der + SPA107689.9912.6585.5915.16
2 Der + GA31084.5914.9380.8618.00
2 Der + SPA + CF198593.469.5890.4814.59
2 Der + SPA + TF1512488.1313.3981.3017.62
2 Der + SPA + C-TF1113390.3911.8386.1316.38
LLFRaw data24206996.007.8995.558.80
SGS27206997.366.5096.587.52
SNV21206998.534.9097.147.36
MSC29206998.704.6095.198.86
1 Der27206999.921.1597.716.73
2 Der 15206999.722.1491.6311.58
1 Der + CARS102296.547.6194.878.12
1 Der + SPA2411799.125.5696.377.91
1 Der + GA157598.025.6394.0310.73
1 Der + SPA + CF2612698.524.9396.168.07
1 Der + SPA +TF3516598.265.5695.078.85
1 Der + SPA + C-TF3417498.315.2097.316.95
MLFCARS1010599.153.6198.176.55
SPA915096.447.6186.2415.81
GA103496.587.5095.408.30
CARS + CF1611499.093.7598.535.28
CARS + TF3915399.223.5698.035.58
CARS + C-TF4116299.851.2598.615.06
Adulterated ALNIRSRaw data21155797.815.9392.4511.73
SGS30155799.662.2497.607.08
SNV29155799.871.4396.378.47
MSC27155799.921.1092.7611.78
1 Der11155798.674.7482.6314.20
2 Der 10155799.901.2668.1621.09
SGS + CARS91491.6911.1986.6315.42
SGS + SPA109598.744.3996.0610.35
SGS + GA76397.026.7693.1912.40
SGS + SPA + CF1110499.333.2395.0310.05
SGS + SPA + TF1014396.197.5388.5715.05
SGS + SPA + C-TF1315299.652.3292.1612.14
HSIRaw data2151289.6612.2588.7715.61
SGS2351291.7911.4379.0217.12
SNV2151294.349.4387.5414.25
MSC2251291.1111.8384.7014.98
1 Der951293.4310.1187.1514.23
2 Der 951299.123.6991.6611.58
2 Der + CARS63593.1010.3987.0914.48
2 Der + SPA89498.015.6088.7313.97
2 Der + GA41686.6713.5285.0317.71
2 Der + SPA + CF1010399.662.3883.0017.21
2 Der + SPA + TF914299.662.2686.5216.78
2 Der + SPA + C-TF915186.3214.5476.6117.66
LLFRaw data36206998.165.4791.9512.29
SGS40206999.811.7597.196.37
SNV33206999.921.1095.5510.20
MSC44206999.911.1995.8010.56
1 Der11206999.492.9092.6810.78
2 Der8206999.143.7288.8613.27
SGS + CARS134493.2710.3186.6014.98
SGS + SPA3011099.053.8692.8212.16
SGS + GA1511698.504.8392.2713.10
SGS + SPA + CF2411999.303.4290.2412.05
SGS + SPA + TF1915899.512.7990.9417.42
SGS + SPA + C-TF2016798.694.4786.2516.38
MLFCARS154995.808.1194.3410.07
SPA2318999.622.3796.2211.00
GA207997.955.5495.1311.17
SPA + CF2319899.622.3798.535.28
SPA + TF2823799.762.0297.875.85
SPA + C-TF3424699.921.1699.002.16
Note: AC = Atractylodes chinensis (DC.) Koidz.; AL = Atractylodes lancea (Thunb.) DC.; NIRS = near-infrared spectroscopy; HIS = hyperspectral imaging; SGS = Savitzky-Golay smoothing; SNV = standard normalized variate; MSC = multiplicative scatter correction; 1 Der = first derivative; 2 Der = second derivative; CARS = competitive adaptive reweighted sampling; SPA = successive projection algorithm; GA = genetic algorithm; CF = color features; TF = texture features; C-TF = color-texture features; R2T = correlation coefficient of training sets; RMSET = root mean square error of training sets; R2P = correlation coefficient of prediction sets; RMSEP = root mean square error of prediction sets.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Jiang, Z.; Lv, A.; Zhong, L.; Yang, J.; Xu, X.; Li, Y.; Liu, Y.; Fan, Q.; Shao, Q.; Zhang, A. Rapid Prediction of Adulteration Content in Atractylodis rhizoma Based on Data and Image Features Fusions from Near-Infrared Spectroscopy and Hyperspectral Imaging Techniques. Foods 2023, 12, 2904. https://doi.org/10.3390/foods12152904

AMA Style

Jiang Z, Lv A, Zhong L, Yang J, Xu X, Li Y, Liu Y, Fan Q, Shao Q, Zhang A. Rapid Prediction of Adulteration Content in Atractylodis rhizoma Based on Data and Image Features Fusions from Near-Infrared Spectroscopy and Hyperspectral Imaging Techniques. Foods. 2023; 12(15):2904. https://doi.org/10.3390/foods12152904

Chicago/Turabian Style

Jiang, Zhiwei, Aimin Lv, Lingjiao Zhong, Jingjing Yang, Xiaowei Xu, Yuchan Li, Yuchen Liu, Qiuju Fan, Qingsong Shao, and Ailian Zhang. 2023. "Rapid Prediction of Adulteration Content in Atractylodis rhizoma Based on Data and Image Features Fusions from Near-Infrared Spectroscopy and Hyperspectral Imaging Techniques" Foods 12, no. 15: 2904. https://doi.org/10.3390/foods12152904

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop