Wolfberry production place tracing method based on deep learning algorithm combining HSI and NMR technologies

Abstract

The invention discloses a wolfberry origin tracing method based on a deep learning algorithm combining HSI and NMR technologies, and relates to a traditional Chinese medicine origin tracing method. The method comprises the following steps: 1) pretreating to-be-detected wolfberry samples from different producing areas, and collecting hyperspectral image data and nuclear magnetic resonance hydrogen spectrum data; 2) preprocessing the collected hyperspectral image data, and constructing a wolfberry origin traceability model based on the preprocessed hyperspectral image data in combination with a deep learning algorithm; (3) preprocessing the acquired nuclear magnetic resonance hydrogen spectrum data, and respectively screening spectrum peak wavebands and difference components of the Chinese wolfberry fruits from different producing areas by using multivariate statistical analysis; and 4) screening the hyperspectral characteristic wavelength based on a SHapley Addive explanation algorithm, and analyzing the correlation between the spectral characteristic wavelength and the difference component. The established Chinese wolfberry origin traceability model has the advantages of being simple in data processing, low in cost, capable of achieving high-throughput Chinese wolfberry origin identification and the like.

Description

Technical Field

[0001] The present invention relates to a method for tracing the origin of traditional Chinese medicine, and in particular to a method for tracing the origin of wolfberry using a deep learning algorithm combining HSI and NMR technologies. Background Art

[0002] As a precious tonic Chinese medicine rich in nutrients and with significant medical and health benefits, wolfberry has attracted widespread attention in the field of functional food and nutrition and health care. The plant is a perennial deciduous shrub. There are about 80 species of Lycium barbarum known worldwide, and my country has 7 species and 2 variants. Wolfberry was first recorded in Shennong's Herbal Classic, and it was listed as a top-grade product, saying: "Long-term consumption strengthens muscles and bones, lightens the body and prevents aging, and is resistant to cold and heat." With the in-depth exploration of wolfberry components by modern pharmacology, especially the key component of wolfberry, Lycium barbarum polysaccharide, it has become a research focus because of its antioxidant, anti-inflammatory, immune system regulation and anti-cancer effects. Wolfberry has a history of cultivation and application in my country for more than two thousand years, and its main production areas include Ningxia, Hebei, Xinjiang and Inner Mongolia. Among them, Ningxia wolfberry (located in China's Loess Plateau) is the only medicinal wolfberry in China and is famous worldwide for its excellent quality. The nutritional and medicinal value of wolfberry and its market pricing are closely related to its origin and variety. Accurately identifying the origin of wolfberry is particularly important for ensuring product quality, enhancing consumer trust and expanding international markets.

[0003] Traditional methods for tracing the origin of wolfberry mainly rely on physical, chemical and sensory analysis, but these methods are often time-consuming and rely on the subjective judgment of experienced experts. In recent years, a variety of new analytical methods have been introduced, such as near-infrared spectroscopy, hyperspectral imaging (HSI), electronic nose, nuclear magnetic resonance (NMR), etc., to more accurately and efficiently identify the origin of wolfberry. Near-infrared spectroscopy and hyperspectral imaging technology can provide rich spectral and spatial information about the sample, and are mainly used to detect the chemical and physical properties of wolfberry; nuclear magnetic resonance has shown excellent performance in identifying and quantifying sample components. By integrating these technologies, the geographical origin of wolfberry can be analyzed more comprehensively, overcoming the limitations of a single method. Although the combination of multiple technologies can provide more comprehensive information, the data obtained by hyperspectral imaging and nuclear magnetic resonance technology pose a challenge to traditional machine learning due to its large data volume and high dimensionality. Summary of the invention

[0004] The purpose of the present invention is to solve the above problems existing in the prior art and realize accurate, rapid, non-destructive and high-throughput traceability analysis of the origin of wolfberry. The present invention provides a wolfberry origin traceability method combining a deep learning algorithm of hyperspectral imaging (HSI) and nuclear magnetic resonance (NMR); to cope with difficulties such as large amount of training data and high dimension. It not only reduces the amount of data collection and improves the calculation efficiency of the model, but also ensures the interpretability of the deep learning model used, and can identify the origin of wolfberry more accurately.

[0005] The present invention comprises the following steps:

[0006] 1) After pre-treatment of the wolfberry samples from different origins to be tested, hyperspectral image data and nuclear magnetic resonance hydrogen spectrum data are collected;

[0007] 2) Preprocess the collected hyperspectral image data, and build a wolfberry origin traceability model based on the preprocessed hyperspectral image data combined with a deep learning algorithm;

[0008] 3) Preprocessing the collected H NMR spectrum data, and using multivariate statistical analysis to screen the spectral peak bands and differential components of wolfberries from different origins;

[0009] 4) Screening of hyperspectral characteristic wavelengths based on SHapley Additive exPlanations algorithm and correlation analysis between spectral characteristic wavelengths and difference components.

[0010] In step 1), the hyperspectral image data is collected by:

[0011] (1) Placing the wolfberry samples from different origins to be tested on a stage;

[0012] (2) Adjusting the acquisition resolution, spectral band number, exposure time, gain, stage height, and motor speed of the hyperspectral system, and then collecting visible and near-infrared hyperspectral image data of wolfberry; the collected hyperspectral image data has a resolution of 882×960, and the band is a visible and near-infrared hyperspectral image with a total of 128 bands within a wavelength range of 400 to 1000 nm;

[0013] In step 1), the step of collecting the H NMR spectrum data is:

[0014] (1) The whole wolfberry was rapidly frozen using liquid nitrogen, then ground into fine powder using a mortar and pestle and sieved;

[0015] (2) dissolving wolfberry powder in a mixed solvent of methanol, chloroform and water; the ratio of methanol, chloroform and water is 2:2:3;

[0016] (3) After vortex mixing, the mixed solution was cooled on ice for 15 min and then centrifuged for 10 min; the vortex mixing time was 60 s, the centrifugal force was 10,000 g, and the centrifugal temperature was 4°C;

[0017] (4) The supernatant was transferred to a centrifuge tube and nitrogen blown for 30 min using a nitrogen blower, followed by freeze drying for 12 h to remove methanol and water;

[0018] (5) The dried material was redissolved in 600 μL of deuterated phosphate buffer containing 0.05% TSP (sodium 3-(trimethylsilyl)propionate-2,2,3,3-d4); the pH of the deuterated phosphate buffer was 5.7, and TSP was a deuterated aqueous solution, which was used as a calibration substance in subsequent data processing;

[0019] (6) After vortex mixing for 5 min, 550 μL of the supernatant was transferred to a 5 mm NMR tube to prepare for H NMR spectrum data collection;

[0020] (7) The prepared samples were placed on a nuclear magnetic resonance spectrometer equipped with an ultra-low temperature probe to collect nuclear magnetic resonance hydrogen spectrum data, and combined with multivariate statistical analysis to screen the different components from different origins. Data collection was performed on an 850MHz nuclear magnetic resonance spectrometer using a CPTCI probe at 850.32MHz using a ZGPR pulse sequence, with the following parameters set: temperature 298K, spectrum width 14KHz, sampling points 32K, relaxation time 4.0s, acquisition time 1.9s, and scan times 64 times.

[0021] In step 2), the collected hyperspectral image data is preprocessed, and the specific steps of constructing a wolfberry origin tracing model based on the preprocessed hyperspectral image data combined with a deep learning algorithm may be:

[0022] (1) The collected hyperspectral image data is corrected by a standard diffuse reflectance reference plate; the correction formula used is: Where R is the corrected image, I raw is the original image, I dark is the black reference image, I ref To collect images of a standard diffuse reflectance reference plate (HSIA-CT-400x400, Dualix, China);

[0023] (2) Identify and crop the region of interest of the corrected visible-near infrared hyperspectral image;

[0024] (3) The cropped image data of the region of interest was converted and saved, and a wolfberry origin traceability model based on deep learning was constructed; the deep learning algorithm used was an improved ResNet-34 model, and the ResNet-34 model was used to train the hyperspectral data. The trained model was used to calculate the SHAP value of each wavelength through the SHAP algorithm, and the spectral wavelengths that contributed most to the correct prediction of the origin by the model were screened;

[0025] (4) Calculate the average spectrum of each wolfberry and build a wolfberry origin traceability model based on traditional machine learning; the traditional machine learning models used include logistic regression (LR) and support vector machine (SVM) models.

[0026] In step 3), the preprocessing of the collected H NMR spectrum data may be specifically as follows: using MestReNova software to calibrate the collected H NMR spectrum data, marking the TSP position as 0 ppm to ensure the comparability of chemical shifts between different samples; adjusting the phase of the spectrum to ensure that the spectrum peak is located at the correct position; adjusting the baseline to eliminate the influence of baseline drift on the integration accuracy; then removing the interfering peaks, and integrating and normalizing the spectrum by total area.

[0027] In step 3), the specific steps of using multivariate statistical analysis to screen the spectral peak bands and differential components of wolfberries from different origins can be: importing the integrated nuclear magnetic resonance hydrogen spectrum data into SIMCA software, and performing multivariate statistical analysis in the SIMCA software, such as PCA (principal component analysis) and PLS-DA (partial least squares discriminant analysis); setting the parameters of the four-dimensional volcano map, including fold change, p-value, absolute correlation coefficient (│r│) and variable importance projection (VIP); and screening out the geographical markers of wolfberries from different origins through the four-dimensional volcano map.

[0028] In step 4), the SHapley Additive exPlanations algorithm is used to screen the hyperspectral characteristic wavelengths and the correlation analysis between the spectral characteristic wavelengths and the difference components, the correlation between the single-band image data and the difference components is explored, and the Pearson correlation coefficient is used to analyze the correlation between the two. Finally, the image data of the band with the highest correlation with the difference components of each origin is selected to optimize the wolfberry origin traceability model.

[0029] Compared with the prior art, the technical effects created by the present invention are embodied in:

[0030] The present invention uses hyperspectral imaging and nuclear magnetic resonance technology to collect data of wolfberries from different origins. Hyperspectral imaging technology has the advantages of being non-destructive, fast, and high-throughput, but its explanatory power for specific substances is relatively weak; while nuclear magnetic resonance technology can accurately identify and identify the different components of wolfberries, but the nuclear magnetic resonance data collection process has high technical requirements, high costs, and cannot achieve high throughput. Therefore, the hyperspectral image data is associated with the wolfberry components, and the image data with characteristic wavelengths is screened out and combined with the deep learning algorithm to construct a wolfberry origin traceability model, which has the advantages of simple data processing, low cost, and the ability to achieve high-throughput wolfberry origin identification. The method based on the association analysis of spectral characteristic wavelengths and differential components of wolfberry origin can not only simplify the data dimension, but also effectively improve the performance of the model. Combining the correlation between multispectral information and geographical markers, a new idea is provided for wolfberry origin traceability, and the accuracy and reliability of the wolfberry origin traceability model are improved. The present invention not only improves the classification accuracy, but also significantly reduces the calculation time. The application of the present invention is not limited to wolfberry, but can also be extended to the origin traceability and quality detection of other agricultural products, and has a wide range of application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 The figure shows the average reflectance spectra of wolfberry samples from five different origins, where Side A represents the average spectrum of the front side of the wolfberry; Side B represents the average spectrum of the back side of the wolfberry. The abbreviations TX in the figure represent wolfberries from Tongxin County, Ningxia; ZN1 represents wolfberries from Zhongning No. 1 variety in Ningxia; ZN2 represents wolfberries from Zhongning No. 2 variety in Ningxia; NMH represents wolfberries from Nomhong, Qinghai; DLH represents wolfberries from Delingha, Qinghai, the same below.

[0032] Figure 2 Confusion matrices of RseNet-34 and improved RseNet-34 models for traceability analysis of wolfberries from different origins based on RGB images and full-wavelength hyperspectral image data, respectively.

[0033] Figure 3 The SHAP algorithm was used to filter the characteristic wavelengths of hyperspectral image data and construct the accuracy results of wolfberry traceability analysis from different origins.

[0034] Figure 4 This is the full relaxation spectrum of H NMR of wolfberry from different origins.

[0035] Figure 5 The score plot (left) and four-dimensional volcano plot (right) of the OPLS-DA model of the nuclear magnetic resonance data of wolfberries from different origins.

[0036] Figure 6 This is a correlation analysis diagram of the characteristic spectrum of wolfberry and its differential components. DETAILED DESCRIPTION

[0037] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention clearer and more understandable, they are further explained below in conjunction with specific embodiments, but the present invention is not limited to the description.

[0038] The present invention introduces a deep learning algorithm to cope with difficulties such as large amount of training data and high dimension. At the same time, the SHapleyAdditive exPlanations (SHAP) algorithm is combined to perform feature screening on the training data. The present invention not only reduces the amount of data collection and improves the model calculation efficiency, but also ensures the interpretability of the deep learning model used. By integrating hyperspectral imaging and nuclear magnetic resonance technology with deep learning to establish a wolfberry origin traceability method, the origin of wolfberry can be more accurately identified, which provides new ideas for ensuring the authenticity of the product origin, enhancing consumer trust, and improving the quality evaluation system of wolfberry.

[0039] Example:

[0040] Experimental materials: Wolfberry samples were collected from five different regions in China, including Tongxin County, Ningxia, Zhongning No. 1, Zhongning No. 2, Zhongning County, Ningxia, Nomhong County, Qinghai, and Delingha County, Qinghai. Wolfberry samples from different origins were picked manually and dried naturally for several days. 105 wolfberry samples were collected from each origin, and a total of 525 samples were collected. After the wolfberries were dried, they were packed into self-sealing bags and immediately stored at -80°C to preserve their biochemical and physical properties, ensuring the accuracy of the collected hyperspectral images and nuclear magnetic resonance hydrogen spectrum data.

[0041] Hyperspectral image data acquisition: The present invention uses a set of hyperspectral imaging systems (GaiaField-V10E) developed by Jiangsu Shuangli Hepu Technology Co., Ltd. to collect hyperspectral image data of wolfberries from different origins. First, multiple wolfberry samples to be tested are placed on the stage, and the distance between the stage and the lens is 35cm; then, in the SpecView software in the computer, the exposure time of the camera is set to 2.4ms, the image resolution of the collected image is set to 882×960, the number of spectral bands is set to 128, the gain is set to 2, and the Bin mode is used. The starting point of the motor is set to 0.8cm, the forward speed of the motor is set to 0.1135cm / s, and the wavelengths of the three channels of the RGB image are 750nm, 531.7nm and 461.8nm respectively; finally, the dark box cabinet door is closed to collect the hyperspectral image data of the wolfberry sample.

[0042] H NMR spectrum data collection: After the hyperspectral image data of all wolfberry samples were collected, 30 wolfberry samples were taken from each origin (a total of 150 samples from five origins) to collect their one-dimensional H NMR spectrum data. Soak a single wolfberry in a mortar filled with liquid nitrogen for grinding and sieving. Then, take 100 mg of wolfberry powder and place it in a 5 mL centrifuge tube, and add 0.4 mL of methanol, 0.4 mL of chloroform and 0.6 mL of ultrapure water to make the extractant. Vortex the dissolved solution for 60 seconds to mix it. The mixed solution was placed at 4 ° C and 10000g for 10 minutes, and then the supernatant was taken and transferred to a new centrifuge tube. The centrifuge tube was placed in a nitrogen blower for 30 minutes and then placed in a -80 ° C refrigerator for 12 hours. Then the centrifuge tube was moved to a freeze dryer for lyophilization for 12 hours to remove methanol and water. 600 μL of phosphate buffer prepared with heavy water at pH = 5.7 was added to all the freeze-dried wolfberry powders for re-dissolution. The dissolved solution was vortexed for 60 seconds to mix, and then centrifuged at 4°C and 10000g for 10 minutes. 550 μL of the supernatant was taken into a 5 mm NMR tube (ST500, NORELL, Inc., Morganton, North Carolina, USA) for sampling. The one-dimensional nuclear magnetic resonance hydrogen spectrum data of all wolfberry samples were collected in an 850 MHz Bruker AVANCE III NMR instrument equipped with an ultra-low temperature probe. The CPTCI probe was used at 850.32 MHz and the ZGPR pulse sequence was used for detection. The acquisition parameters were set as follows: temperature 298 K, spectral width 14 KHz, sampling points 32k, relaxation delay 4.0 s, acquisition time 1.9 s, and scan times 64 times. In addition, for accurate quantification, the sample with the best spectral peak was selected from each origin to increase the ZGPR sampling with a relaxation delay of 30 s.

[0043] Data preprocessing: After the hyperspectral image data is collected, it needs to be preprocessed by black and white plate correction, region of interest extraction, average spectrum calculation, etc. Black and white plate correction uses: Formula, where R is the corrected image, I raw is the original image, I dark is the black reference image, I refTo collect the image of the standard diffuse reflectance reference plate; the region of interest uses the image recognition method to identify a single wolfberry as the region of interest and crop it. The cropped hyperspectral image of the single wolfberry is used as the input data for the subsequent deep learning algorithm; the average spectrum calculation is to calculate the average value of each band image of the whole wolfberry as the subsequent input data based on traditional machine learning. After the one-dimensional nuclear magnetic resonance hydrogen spectrum data is collected, the MestReNova software is used to calibrate the collected data (mark the TSP position as 0ppm), adjust the phase and baseline, remove the interference peaks, and then integrate the spectrum and normalize the total area. The integration interval is δ0.5-10, and the integration interval is set to 0.002ppm. Since the number of protons corresponding to the TSP peak is 9, the spectrum peak at TSP is normalized to 9 to facilitate subsequent quantitative analysis. The quantitative formula is Among them C x is the mass concentration of a component x; A' x A' is the integrated area of ​​metabolite x in the non-full relaxation spectrum; TSP A is the integrated area of ​​the calibration substance TSP in the non-full relaxation spectrum; x0 (∞) is the integrated area of ​​a component x0 in the full relaxation spectrum; A x0 (t) is the integrated area of ​​a component x0 in the non-full relaxation spectrum; A TSP0 (t) is the integrated area of ​​the calibration substance TSP in the non-full relaxation spectrum; A TSP0 (∞) is the integrated area of ​​the calibration substance TSP in the full relaxation spectrum; N TSP is the proton number of the calibration substance TSP 9; N x is the number of protons corresponding to component x; MW x is the molar mass of component X; MW TSP is the molar mass of the calibration substance TSP; C TSP is the molar concentration (M) of the standard substance TSP.

[0044] Establishment of wolfberry origin traceability model: This embodiment establishes wolfberry origin traceability models based on hyperspectral and nuclear magnetic resonance data, respectively. For the average spectral reflectance data in the hyperspectral, traditional machine learning algorithms logistic regression (LR) and support vector machine (SVM) are used to establish wolfberry origin traceability models, respectively. For hyperspectral image data, the improved ResNst34 model is used to establish the wolfberry origin traceability model, and the improved ResNet34 model network structure parameters are shown in Table 1. For the integrated nuclear magnetic resonance hydrogen spectrum data, it is imported into SIMCA 14.1 software for multivariate statistical analysis, and finally a four-dimensional volcano map is established to screen the geographical markers of wolfberries from different origins, wherein the parameters of the four-dimensional volcano map include fold change, p value, absolute correlation coefficient (│r│) and variable importance projection (variable importance projection, VIP).

[0045] Table 1 Improved ResNet34 model network structure parameters

[0046]

[0047] Note: conv3d / 2d means 3D / 2D convolution layer. BN3d / 2d means batch normalization.

[0048] Experimental results: The average spectral reflectance curves of wolfberries from different origins were calculated as follows: Figure 1 As shown, from Figure 1 It can be seen that the trend of the average spectral reflectance curve of wolfberries from different origins is consistent, but the values ​​are different, which is caused by the different concentrations of the ingredients contained in wolfberries from different origins. The figure also shows that the spectral differences of wolfberries from different sides (Side A and Side B) from the same origin are small, indicating that there will be no significant error when using hyperspectral imaging technology to collect data from only one side.

[0049] Although there are differences in the spectra of wolfberries from different origins, these differences are difficult to identify with the naked eye. Therefore, traditional machine learning LR (linear) and SVM (non-linear) algorithms are used to build traceability models for wolfberries from different origins. The accuracy of the models is shown in Table 2. It can be seen from Table 2 that the classification accuracy of wolfberry origin traceability based on Side A or Side B data does not exceed 90%. However, when all Side ApB data are used to build the model, the classification accuracy is significantly improved, with the LR model reaching 95.24% and the SVM model classification accuracy reaching 99.43%.

[0050] Table 2 Accuracy of wolfberry origin traceability model established based on LR and SVM algorithms

[0051]

[0052] Note: "Side ApB" means using all Side A data to train the model and using all Side B data as the test set.

[0053] Considering that the classification accuracy of the LR and SVM models is less than 90%, this embodiment introduces a deep learning algorithm to further analyze high-dimensional spectral image data rich in spatial and spectral information. Figure 2 As shown in the figure, a more accurate classification of the origin of wolfberry is achieved, with an overall classification accuracy of 95.63%, an increase of 6.26% compared to traditional machine learning methods.

[0054] In general, the improved ResNet-34 model has a high recognition accuracy of 95.63% for 5 different origins of wolfberry. However, using 3D convolution to process high-dimensional spectral image data not only increases the amount of calculation but also increases the complexity of the model. The "black box" nature of the deep learning model leads to the lack of interpretability of the model. Therefore, this embodiment uses the SHAP algorithm to screen the bands in the model that contribute more to the origin tracing. The wolfberry origin tracing classification results constructed based on the screened characteristic bands are shown in the figure. Figure 3 As shown in the figure, it can be seen that using only 35% of the spectral band data can achieve the same classification accuracy as the full-band spectral data, while the model reduces the calculation time by 29.9%. By combining the method based on hyperspectral data with deep learning, it is possible to achieve fast, non-destructive, and high-throughput wolfberry origin traceability analysis.

[0055] One-dimensional nuclear magnetic resonance proton spectra of wolfberries from different origins Figure 4 As shown. It can be seen from the figure that most of the peaks exist between δ1-6 and δ6.5-9.5. In the high field region of δ0.8-3.5 (mainly containing amino acid and organic acid peaks), the content of wolfberry components from different origins is not much different, but the content of alanine (δ1.48) and unknown-1 (δ1.51) in Ningxia TX, ZN1 and ZN2 wolfberries is significantly higher than that in Qinghai NMH and DLH. In the middle and low field region of δ3.5-5.5, the peaks are mainly carbohydrates, while in the low field region of δ6-10, the peaks are mainly aromatic compounds. By comparing chemical shifts and coupling constants, and querying the FooDB and HMDB databases, 62 components were identified as shown in Table 3.

[0056] Table 3 Substances identified in H NMR spectra of wolfberries from different origins and their absolute quantitative results

[0057]

[0058]

[0059]

[0060]

[0061] Figure 5 The OPLS-DA score map and four-dimensional volcano map are established based on the nuclear magnetic resonance hydrogen spectrum data and the content of different components of wolfberry. The model parameters include R 2 X, R 2 Y and Q 2 . Figure 5 Figure a compares wolfberries from different provinces in China. There is a clear distinction between the Ningxia (NX) and Qinghai (QH) samples along the vertical axis. The R 2 Y=0.871 and Q 2 =0.833, indicating its effectiveness in distinguishing wolfberry producing areas. The geographical markers of wolfberry between Ningxia and Qinghai were screened out by constructing a four-dimensional volcano map based on multivariate and univariate analysis parameters. Combined with Table 3, it can be concluded that the contents of glutamic acid, malic acid and proline in wolfberry from Qinghai are significantly higher than those in Ningxia samples, while the content of dihydroxyacetone in Ningxia wolfberry is 2.7 times that of Qinghai. Figure 5 Figures b and c in the figure show the analysis results of wolfberry in various counties in China. From the OPLS-DA score chart on the left, we can see that there is a significant difference between the samples of Tongxin County (TX) and Zhongning County (ZN) in Ningxia, and between Nuomuhong County (NMH) and Delingha County (DLH) in Qinghai. 2 X, R 2 Y and Q 2 The model parameters show that it has a significant geographical differentiation effect, demonstrating a high degree of predictability and reliability. Observing the volcano map on the right, it can be seen that the concentrations of various nutrients such as 1-methyluric acid, choline, and glutamate in Zhongning wolfberry are higher than those in Tongxin wolfberry, which also shows the nutritional value and quality superiority of Zhongning wolfberry. Zhongning County is famous for its high-quality wolfberries cultivated under its unique geographical and climatic conditions. However, even different wolfberry varieties from the same origin show differences in composition. Figure 5 The OPLS-DA analysis results in Figure d clearly distinguish between Zhongning No. 1 and Zhongning No. 2 varieties. Figure 5The volcano plot on the right of Figure d further identified several significantly different components, such as glucosamine 6-phosphate, gulolactone, and isoleucine, which were found in higher concentrations in Zhongning No. 1 than in Zhongning No. 2. These components have shown significant health benefits in clinical practice, such as relieving knee pain and improving functional mobility. Lysine, as an essential amino acid, not only supports proper growth and reduces anxiety, but also shows potential in cancer treatment. These findings not only emphasize the importance of in-depth component analysis of wolfberry using nuclear magnetic resonance technology, but also provide a scientific basis for the identification of the geographical origin and quality assessment of wolfberry.

[0062] The wolfberry origin tracing model based on the characteristic bands screened by hyperspectral image data combined with deep learning and based on nuclear magnetic resonance data combined with multivariate statistical analysis can well distinguish wolfberries from different origins. However, the disadvantages of large amount of hyperspectral data, high dimensionality, difficulty in preparing samples by nuclear magnetic resonance and high cost limit its application in rapid, non-destructive and high-throughput wolfberry origin tracing. This example further screens more streamlined spectral bands by analyzing the correlation between the characteristic wavelengths of the spectrum and the differential components of the wolfberry origin. The results are as follows: Figure 6 As shown. Through HSI-NMR analysis, most of the substances in the wolfberry samples have a significant correlation with the red band (633-768nm), which is consistent with the red appearance of the wolfberry fruit. Using three specific bands (633.6nm, 665.6nm and 980nm) that are highly correlated with geographical indicators as the input of the improved ResNet-34 model, the constructed wolfberry origin traceability classification results are shown in Table 4. Among them, the wolfberries in Ningxia and Qinghai regions achieved a distinction accuracy of 95%, the classification accuracy of wolfberries in Tongxin County and Zhongning County, Ningxia reached 96.35%, and the distinction accuracy of Zhongning No. 1 and Zhongning No. 2, which are planted in close locations, also reached 93.75%, and the calculation time of the model was reduced by 46.26%. This shows that the wolfberry origin traceability model constructed by the screened spectral bands based on the association analysis of the wolfberry hyperspectral characteristic bands and the differential components of different origins in this embodiment can well distinguish wolfberries from different origins. This invention can provide a new method for the traceability analysis of wolfberries from different origins in China, and provide technical support for the subsequent development of portable multi-spectral detection equipment.

[0063] Table 4 Accuracy of wolfberry origin traceability model based on improved ResNet34 model

[0064]

[0065] Finally, the above specific embodiments are only representative examples of the present invention. Obviously, the technical solution of the present invention is not limited to the above embodiments, but can also be applied in many aspects. All variations that can be directly derived or associated with the content disclosed by ordinary technicians in this field should be considered as the protection scope of the present invention.

Claims

1. A method for tracing the origin of wolfberry using a deep learning algorithm combining HSI and NMR technology, characterized in that The following steps are involved: 1) After pre-treatment of the wolfberry samples from different origins to be tested, hyperspectral image data and nuclear magnetic resonance hydrogen spectrum data are collected; 2) Preprocess the collected hyperspectral image data, and build a wolfberry origin traceability model based on the preprocessed hyperspectral image data combined with a deep learning algorithm; 3) Preprocessing the collected H NMR spectrum data, and using multivariate statistical analysis to screen the spectral peak bands and differential components of wolfberries from different origins; 4) Screening of hyperspectral characteristic wavelengths based on SHapley Additive exPlanations algorithm and correlation analysis between spectral characteristic wavelengths and difference components.

2. A method for tracing the origin of wolfberry using a deep learning algorithm combining HSI and NMR technology as claimed in claim 1, characterized in that In step 1), the hyperspectral image data is collected by: (1) Placing the wolfberry samples from different origins to be tested on a stage; (2) The acquisition resolution, number of spectral bands, exposure time, gain, stage height and motor speed of the hyperspectral system were adjusted, and then the visible and near-infrared hyperspectral image data of wolfberry were collected; the resolution of the collected hyperspectral image data was 882×960, and the band was a visible and near-infrared hyperspectral image with a total of 128 bands within the wavelength range of 400 to 1000 nm.

3. A method for tracing the origin of wolfberry using a deep learning algorithm combining HSI and NMR technology as claimed in claim 1, characterized in that In step 1), the step of collecting the H NMR spectrum data is: (1) The whole wolfberry was rapidly frozen using liquid nitrogen, then ground into fine powder using a mortar and pestle and sieved; (2) dissolving wolfberry powder in a mixed solvent of methanol, chloroform and water; the ratio of methanol, chloroform and water is 2:2:3; (3) After vortex mixing, the mixed solution was cooled on ice for 15 min and then centrifuged for 10 min; the vortex mixing time was 60 s, the centrifugal force was 10,000 g, and the centrifugal temperature was 4°C; (4) The supernatant was transferred to a centrifuge tube and nitrogen blown for 30 min using a nitrogen blower, followed by freeze drying for 12 h to remove methanol and water; (5) The dried material was redissolved in 600 μL of deuterated phosphate buffer containing 0.05% TSP (sodium 3-(trimethylsilyl)propionate-2,2,3,3-d4); the pH of the deuterated phosphate buffer was 5.7, and TSP was a deuterated aqueous solution, which was used as a calibration substance in subsequent data processing; (6) After vortex mixing for 5 min, 550 μL of the supernatant was transferred to a 5 mm NMR tube to prepare for H NMR spectrum data collection; (7) The prepared samples were placed on a nuclear magnetic resonance spectrometer equipped with an ultra-low temperature probe to collect nuclear magnetic resonance hydrogen spectrum data, and combined with multivariate statistical analysis to screen the different components from different origins; data collection was carried out on an 850 MHz nuclear magnetic resonance spectrometer using a CPTCI probe at 850.32 MHz using a ZGPR pulse sequence, and the parameters were set as follows: temperature 298 K, spectral width 14 KHz, sampling points 32 K, relaxation time 4.0 s, acquisition time 1.9 s, and scan times 64 times.

4. A method for tracing the origin of wolfberry using a deep learning algorithm combining HSI and NMR technology as claimed in claim 1, characterized in that In step 2), the specific steps of preprocessing the collected hyperspectral image data and constructing a wolfberry origin tracing model based on the preprocessed hyperspectral image data combined with a deep learning algorithm are as follows: (1) The collected hyperspectral image data is corrected by a standard diffuse reflectance reference plate; the correction formula used is: Where R is the corrected image, I raw is the original image, I dark is the black reference image, I ref To collect images of standard diffuse reflection reference plates; (2) Identify and crop the region of interest of the corrected visible-near infrared hyperspectral image; (3) The cropped image data of the region of interest was converted and saved, and a wolfberry origin traceability model based on deep learning was constructed; the deep learning algorithm used was an improved ResNet-34 model, and the ResNet-34 model was used to train the hyperspectral data. The trained model was used to calculate the SHAP value of each wavelength through the SHAP algorithm, and the spectral wavelengths that contributed most to the correct prediction of the origin by the model were screened; (4) The average spectrum of each wolfberry was calculated, and a wolfberry origin traceability model was established based on traditional machine learning. The traditional machine learning models used included LR and SVM models.

5. A method for tracing the origin of wolfberry using a deep learning algorithm combining HSI and NMR technology as claimed in claim 1, characterized in that In step 3), the preprocessing of the collected H NMR spectrum data is specifically as follows: using MestReNova software to calibrate the collected H NMR spectrum data, marking the TSP position as 0 ppm to ensure that the chemical shifts between different samples are comparable; adjusting the phase of the spectrum to ensure that the spectrum peak is located at the correct position; adjusting the baseline to eliminate the influence of baseline drift on the integration accuracy; Then the interfering peaks were removed, and the spectra were integrated and normalized for total area.

6. A method for tracing the origin of wolfberry using a deep learning algorithm combining HSI and NMR technology as claimed in claim 1, characterized in that In step 3), the specific steps of using multivariate statistical analysis to screen the spectral peak bands and differential components of wolfberries from different origins are: importing the integrated nuclear magnetic resonance hydrogen spectrum data into SIMCA software, and performing multivariate statistical analysis in the SIMCA software; setting the parameters of the four-dimensional volcano map, including the difference multiple, p value, absolute correlation coefficient │r│ and variable importance projection; The geographical markers of wolfberries from different origins were screened out through four-dimensional volcano maps.

7. A method for tracing the origin of wolfberry using a deep learning algorithm combining HSI and NMR technology as claimed in claim 1, characterized in that In step 4), the SHapley Additive exPlanations algorithm is used to screen the hyperspectral characteristic wavelengths and the correlation analysis between the spectral characteristic wavelengths and the difference components, the correlation between the single-band image data and the difference components is explored, and the Pearson correlation coefficient is used to analyze the correlation between the two. Finally, the image data of the band with the highest correlation with the difference components of each origin is selected to optimize the wolfberry origin traceability model.

Images