Discrimination method and system for sorghum varieties based on decision fusion and near-infrared spectroscopy

Abstract

The application discloses a discrimination method and system for sorghum varieties based on decision fusion and near-infrared spectroscopy, and the method comprises the following steps: collecting and processing samples of sorghum grains; collecting near-infrared spectra of each sample to construct a calibration set; determining the tannin content and amylopectin content of each sample; constructing a first quantitative detection model between the near-infrared spectra of the samples and the tannin content, and constructing a second quantitative detection model between the near-infrared spectra of the samples and the amylopectin content; constructing a qualitative discrimination model between the near-infrared spectra of the samples and the category attribution; predicting the calibration set respectively to obtain three groups of prediction values, taking the three groups of prediction values as independent variables, and taking the category labels of the samples as dependent variables to construct a decision fusion model; for unknown sorghum samples, the tannin content and amylopectin content of the samples are predicted respectively, the qualitative discrimination model is used to predict the category attribution, and finally, the category prediction value is obtained; the method has the advantages of high precision and strong robustness.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of sorghum variety classification, and in particular to a method and system for classifying sorghum varieties based on decision fusion and near-infrared spectroscopy. BACKGROUND

[0002] Sorghum is one of the important food crops in China. Sorghum is rich in various nutrients and is mainly used as a raw material for brewing wine in China. Different wine enterprises use different sorghum varieties for brewing, and in the raw material procurement process, in order to ensure the quality of the processed wine, the enterprises give higher procurement prices for their special sorghum varieties than for other varieties. Traditional sorghum variety identification methods such as molecular identification and field identification are complex and time-consuming, and cannot meet the variety determination needs of wine enterprises in the procurement and storage process. Near-infrared spectroscopy technology, as a mature rapid detection technology, is very suitable for rapid identification of sorghum varieties. For example, Chinese patent publication CN103743705A discloses a rapid detection method for false sorghum and its similar species, determines the optimal principal component number as 9, after feature extraction, takes the 9 principal components as input variables of the neural network, establishes a three-layer BP neural network, and calculates the category to which each subcategory belongs. This patent application and other similar technical solutions that use spectroscopy to identify crop varieties are representative. They all use the category label as the output layer, use different algorithms to obtain the spectrum or some characteristic variables of the spectrum as the input layer, and then build a relationship model between the two.

[0003] However, using the category label as the output layer has certain limitations. The category label is usually a limited number of discrete values. Using it as the output for modeling can lead to insufficient model interpretability and the model is extremely dependent on the calibration sample. When the calibration sample is not representative, the model is not robust. Assuming that the calibration sample not only has differences in varieties, but also has differences other than varieties, if only the category label is used as the output value, then no matter what improved algorithm is used, it cannot guarantee a robust and reliable prediction result. To give a specific example, assume that Guojiaohong No. 1 from Luzhou and Qingqiao Yangsuo sorghum from Hainan are used as calibration samples for binary classification. If the category label (0 or 1) is used as the output value, the actual model built can be interpreted as either a variety discrimination model or a production area discrimination model, because the labels for these two types of models are the same. When the actual test sample and the calibration sample differ, for example, the actual test sample is Guojiaohong No. 1 from Hainan, the model may make a wrong prediction based on the production area rather than the variety of the object. Therefore, near-infrared qualitative discrimination analysis usually requires hundreds or thousands of samples to increase the representativeness of the calibration set as much as possible, but in actual application, collecting so many samples requires a high time cost and labor cost. SUMMARY

[0004] The technical problem to be solved by the present application is that the existing corn variety discrimination method has poor interpretability, low precision and poor robustness, and the existing spectrum detection technology has great dependence on sample representativeness.

[0005] The present application solves the above technical problems by the following technical means: a corn variety discrimination method based on decision fusion and near-infrared spectrum, comprising:

[0006] S1. Sample collection and processing of corn kernels;

[0007] S2. Collecting near-infrared spectrum of each sample and constructing a calibration set;

[0008] S3. Determining the tannin content and amylopectin content of each sample;

[0009] S4. Constructing a first quantitative detection model between the near-infrared spectrum of the sample and the tannin content, and constructing a second quantitative detection model between the near-infrared spectrum of the sample and the amylopectin content;

[0010] S5. Constructing a qualitative discrimination model between the near-infrared spectrum of the sample and the category attribution;

[0011] S6. Using the first quantitative detection model, the second quantitative detection model and the qualitative discrimination model to respectively predict the calibration set, obtaining three groups of prediction values, taking the three groups of prediction values as independent variables and the category label of the sample as dependent variable, and constructing a decision fusion model;

[0012] S7. For unknown corn samples, the spectrum is collected in the same way as in S2, the tannin content and amylopectin content are predicted by the first quantitative detection model and the second quantitative detection model in S4, the category attribution is predicted by the qualitative discrimination model in S5, and then the three groups of prediction values are further predicted by the decision fusion model in S6 to obtain the final category prediction value. When the final category prediction value is ≤0.5, the sample belongs to other varieties, and when the final category prediction value is >0.5, the sample belongs to the target variety.

[0013] Further, the S1 comprises:

[0014] Collecting more than 60 samples of complete mature corn kernels with known variety attribution to form a data set, the data set including a single variety sample or multiple variety samples or multiple variety mixed samples that the user is interested in, these samples being regarded as target variety samples, and the target variety samples being more than 30; the samples in the data set other than the target variety samples being other variety samples, and the other variety samples being more than 30; using a flour mill to process the corn kernels of each target variety sample and other variety sample into powder; assigning a category label 1 to the target variety samples and a category label 0 to the other variety samples.

[0015] Further, S2 includes:

[0016] Near-infrared spectra of each sample were acquired using a diffuse reflectance method on a near-infrared spectrometer. Each sample was acquired twice at different time periods. The near-infrared spectra of the first acquired sample were used as the calibration set for modeling, and the near-infrared spectra of the second acquired sample were used as the validation set to verify the predictive performance of the decision fusion model.

[0017] Further, S3 includes:

[0018] The tannin and amylopectin content of each sorghum sample was determined using a wet chemical method.

[0019] Furthermore, the method for determining the amylopectin content includes:

[0020] The preparation of standard solutions and absorbance detection were carried out according to the standard NY / T 55-1987 of the Ministry of Agriculture and Rural Affairs of the People's Republic of China. A spectrophotometer was used to read the absorbance of the prepared standard solutions at 640 nm and 532 nm. The absorbance read at 640 nm was used to construct a vector. The vector is formed by the absorbance read at 532 nm. ,vector sum vector Construct a matrix , with matrix amylopectin concentration is the independent variable. Using amylopectin as the dependent variable, a partial least squares regression equation for amylopectin concentration was constructed. The absorbance of the sample at 640 nm and 532 nm was read and substituted into the regression equation to obtain the amylopectin concentration. Using the formula Obtaining amylopectin content in, The quality of the sample to be tested, This represents the moisture percentage of the sample to be tested.

[0021] Furthermore, the process of constructing the regression equation for amylopectin concentration is as follows:

[0022] S321, Regarding the matrix Using formula Standardize, For standardization , Represents the mean. Representative standard deviation; for amylopectin concentration Standardization is performed using the following formula.

[0023] (1)

[0024] wherein, is the normalized ;

[0025] solving equation to get the first weight vector of , wherein is the largest eigenvalue of matrix , and the calculated is expressed as wherein represents the contribution weight of to the construction of the first score vector ; represents the contribution weight of to the construction of the first score vector ;

[0026] calculating the first score vector and the second score vector , wherein represents the first weight vector of ;

[0027] calculating the first loading vector of and the first loading vector of

[0028] calculating the first residual matrix of and the first residual matrix of ;

[0029] S322, taking the residual matrix obtained in the last iteration and as the new and , repeating all steps of S321 to obtain the second weight vector of 、 the third score vector of , the fourth score vector of , the second loading vector of , the second residual matrix of , and the second residual matrix Represented as ,in, represent For building Contribution weight, represent For building Contribution weight;

[0030] S323, to The regression formula is as follows:

[0031] (2)

[0032] in, for The residual, and All The regression coefficients, , ;

[0033] S324, will and use express:

[0034] (3)

[0035] (4)

[0036] According to formulas (1) and (2), Combining formulas (3) and (4), and after merging like terms, the final formula is as follows:

[0037]

[0038] in, It is the intercept. and yes The regression coefficient is given by the following formula:

[0039]

[0040]

[0041] 。

[0042] Furthermore, both the first and second quantitative detection models are modeled using partial least squares (PLS). The modeling parameters for the first quantitative detection model are as follows: the sample spectrum is first smoothed using the first derivative of the Savitsky-Golay method with 17 points, and then the near-infrared spectrum in the ranges of 928-1008nm, 1174-1328nm, and 1550-1617nm is extracted for modeling, and the PLS factor number during modeling is 9.

[0043] The modeling parameters of the second quantitative detection model are as follows: First, the sample spectrum is smoothed using the first derivative of the Savitsky-Golay method with 17 points. Then, the near-infrared spectra at 938nm, 960nm, 1034-1037nm, 1043-1052nm, 1062-1065nm, 1109-1118nm, 1124-1128nm, 1156-1161nm, 1214-1223nm, 1244nm, 1250nm, 1259nm, 1265nm, 1277nm, 1315-1325nm, 1378-1381nm, 1430nm, 1445nm, 1456nm, 1502nm, 1508nm, 1513nm, 1653nm, and 1686nm are used for modeling, and the PLS factor number during modeling is 10.

[0044] Furthermore, the qualitative discrimination model is modeled using partial least squares, and the modeling parameters of the qualitative discrimination model are:

[0045] Near-infrared spectra at 928nm, 996nm, 1087nm, 1223nm, 1322nm, 1345nm, 1669nm, and 1680nm were used for modeling, and the PLS factor was 8 during modeling.

[0046] Furthermore, the decision fusion model is modeled using discriminant partial least squares (PLS) method, and the modeling parameter of the decision fusion model is 3 PLS factors.

[0047] This invention also provides a sorghum variety discrimination system based on decision fusion and near-infrared spectroscopy, comprising:

[0048] The sample collection module is used for the collection and processing of sorghum grain samples.

[0049] The spectral acquisition module is used to acquire near-infrared spectra for each sample and construct a calibration set;

[0050] The chemical reference value acquisition module is used to determine the tannin and amylopectin content of each sample.

[0051] The quantitative model building module is used to build a first quantitative detection model between the near-infrared spectrum of a sample and the tannin content, and to build a second quantitative detection model between the near-infrared spectrum of a sample and the amylopectin content.

[0052] The qualitative discrimination model building module is used to build a qualitative discrimination model between the near-infrared spectrum of a sample and its class classification.

[0053] The decision fusion model construction module is used to predict the calibration set using the first quantitative detection model, the second quantitative detection model, and the qualitative discrimination model respectively, to obtain three sets of predicted values. These three sets of predicted values ​​are used as independent variables, and the class label of the sample is used as the dependent variable to construct the decision fusion model.

[0054] The result prediction module is used to acquire spectra of unknown sorghum samples in the same way as the spectral acquisition module. The first and second quantitative detection models in the quantitative model construction module are used to predict the tannin content and amylopectin content, respectively. The qualitative discriminant model in the qualitative discriminant model construction module is used to predict the category. Then, the decision fusion model in the decision fusion model construction module is used to further predict the above three sets of predicted values ​​to obtain the final category prediction value. When the final category prediction value is ≤0.5, the sample belongs to other varieties, and when the final category prediction value is >0.5, the sample belongs to the target variety.

[0055] Furthermore, the sample acquisition module is also used for:

[0056] A dataset is constructed by collecting more than 60 complete and mature sorghum grain samples with known cultivar affiliations. The dataset includes samples of a single variety, multiple varieties, or a mixture of multiple varieties of interest to the user. These samples are considered as target variety samples, with more than 30 target variety samples. Samples other than target variety samples in the dataset are considered as other variety samples, with more than 30 other variety samples. The sorghum grains of each target variety sample and other variety samples are processed into powder using a mill. The target variety samples are assigned a category label of 1, and the other variety samples are assigned a category label of 0.

[0057] Furthermore, the spectral acquisition module is also used for:

[0058] Near-infrared spectra of each sample were acquired using a diffuse reflectance method on a near-infrared spectrometer. Each sample was acquired twice at different time periods. The near-infrared spectra of the first acquired sample were used as the calibration set for modeling, and the near-infrared spectra of the second acquired sample were used as the validation set to verify the predictive performance of the decision fusion model.

[0059] Furthermore, the chemical reference value acquisition module is also used for:

[0060] The tannin and amylopectin content of each sorghum sample was determined using a wet chemical method.

[0061] Furthermore, the method for determining the amylopectin content includes:

[0062] The preparation of standard solutions and absorbance detection were carried out according to the standard NY / T 55-1987 of the Ministry of Agriculture and Rural Affairs of the People's Republic of China. A spectrophotometer was used to read the absorbance of the prepared standard solutions at 640 nm and 532 nm. The absorbance read at 640 nm was used to construct a vector. The vector is formed by the absorbance read at 532 nm. ,vector sum vector Construct a matrix , with matrix amylopectin concentration is the independent variable. Using amylopectin as the dependent variable, a partial least squares regression equation for amylopectin concentration was constructed. The absorbance of the sample at 640 nm and 532 nm was read and substituted into the regression equation to obtain the amylopectin concentration. Using the formula Obtaining amylopectin content in, The quality of the sample to be tested, This represents the moisture percentage of the sample to be tested.

[0063] Furthermore, the process of constructing the regression equation for amylopectin concentration is as follows:

[0064] S321, Regarding the matrix Using formula Standardize, For standardization , Represents the mean. Representative standard deviation; for amylopectin concentration Standardization is performed using the following formula.

[0065] (1)

[0066] in, For standardization ;

[0067] Solving formula get First weight vector The value of , where, For matrix The largest eigenvalue, after calculation Represented as in represent To construct the first score vector Contribution weight, represent To construct the first score vector Contribution weight;

[0068] Calculate the first score vector Second score vector ,in express The first weight vector;

[0069] calculate First load vector and First load vector

[0070] calculate The first residual matrix and The first residual matrix ;

[0071] S322, The residual matrix obtained from the previous iteration is... and As a new and Repeat all steps in S321 to obtain The second weight vector 、 The third score vector , The fourth score vector , The second load vector , The second load vector , The second residual matrix and The second residual matrix After calculation Represented as ,in, represent For building Contribution weight, represent For building Contribution weight;

[0072] S323, to The regression formula is as follows:

[0073] (2)

[0074] in, for The residual is the error term remaining after fitting the equation (2). and All The regression coefficients can be obtained using the least squares method, and the formula is as follows: , ;

[0075] S324, will and use express:

[0076] (3)

[0077] (4)

[0078] According to formulas (1) and (2), Combining formulas (3) and (4), and after merging like terms, the final formula is as follows:

[0079]

[0080] in, It is the intercept. and yes The regression coefficient is given by the following formula:

[0081]

[0082]

[0083] 。

[0084] Furthermore, both the first and second quantitative detection models are modeled using partial least squares (PLS). The modeling parameters for the first quantitative detection model are as follows: the sample spectrum is first smoothed using the first derivative of the Savitsky-Golay method with 17 points, and then the near-infrared spectrum in the ranges of 928-1008nm, 1174-1328nm, and 1550-1617nm is extracted for modeling, and the PLS factor number during modeling is 9.

[0085] The modeling parameters of the second quantitative detection model are as follows: First, the sample spectrum is smoothed using the first derivative of the Savitsky-Golay method with 17 points. Then, the near-infrared spectra at 938nm, 960nm, 1034-1037nm, 1043-1052nm, 1062-1065nm, 1109-1118nm, 1124-1128nm, 1156-1161nm, 1214-1223nm, 1244nm, 1250nm, 1259nm, 1265nm, 1277nm, 1315-1325nm, 1378-1381nm, 1430nm, 1445nm, 1456nm, 1502nm, 1508nm, 1513nm, 1653nm, and 1686nm are used for modeling, and the PLS factor number during modeling is 10.

[0086] Furthermore, the qualitative discrimination model is modeled using partial least squares, and the modeling parameters of the qualitative discrimination model are:

[0087] Near-infrared spectra at 928nm, 996nm, 1087nm, 1223nm, 1322nm, 1345nm, 1669nm, and 1680nm were used for modeling, and the PLS factor was 8 during modeling.

[0088] Furthermore, the decision fusion model is modeled using discriminant partial least squares (PLS) method, and the modeling parameter of the decision fusion model is 3 PLS factors.

[0089] The advantages of this invention are:

[0090] (1) This invention proposes a near-infrared discrimination method based on decision fusion. Based on the traditional sorghum variety classification model constructed using category labels, multiple near-infrared quantitative detection models for sorghum components are added. Then, the basic sorghum variety classification model (i.e., the qualitative discrimination model) and the newly added near-infrared quantitative detection models for sorghum components are used to predict the sample. The predicted classification values ​​and component values ​​are used as input to construct a new variety classification model, namely the decision-level fusion model, to re-determine the category of sorghum varieties. Its beneficial effect is that this method considers both the traditional category classification of sorghum varieties and the actual correlation between sorghum varieties and their components during calculation. By constructing the decision-level fusion model, the component content (predicted value) of the sorghum sample is used to correct the initial variety classification result. Since the final result is calculated jointly by the qualitative and quantitative discriminant models, this correction method can enhance the interpretability of the model, improve the discrimination effect, and increase the discrimination accuracy. Furthermore, only a few dozen samples are needed to model in near-infrared quantitative analysis because the output value of the quantitative model is not a discrete category label, but a continuous component value. The model can self-correct based on the differences of these component values, reducing the model's dependence on sample representativeness and improving prediction robustness in the case of a small number of samples.

[0091] (2) In this invention, absorbance at 640 nm and 532 nm is used as independent variables. X 1 and X 2. Amylopectin concentration is the dependent variable. Y A regression equation for amylopectin concentration was constructed using partial least squares (PLS factor number 2). The main reason for this was that during the colorimetric reaction, absorption peaks in the ranges of 600-680 nm and 520-555 nm were generated, which are the characteristic absorption peaks of the amylose-iodine complex and amylopectin-iodine complex, respectively. 640 nm and 532 nm are the maximum absorption wavelengths of the amylose-iodine complex and amylopectin-iodine complex, respectively, theoretically providing the best regression results; therefore, these two absorbance values ​​were chosen.

[0092] (3) The parameters such as band range, preprocessing and PLS factor number in S4 and S5 of this invention are obtained by combining multiple band selection algorithms and preprocessing algorithms, and the model is optimized and screened with the lowest cross-validation root mean square error as the model performance evaluation standard. Under the relevant parameter selection, the model has higher accuracy and better performance. Attached Figure Description

[0093] Figure 1 This is a flowchart of the sorghum variety discrimination method based on decision fusion and near-infrared spectroscopy disclosed in Embodiment 1 of the present invention;

[0094] Figure 2This is a scatter plot of the predicted values ​​of the discrimination model constructed in S5 and the decision fusion model constructed in S6 of the sorghum variety discrimination method based on decision fusion and near-infrared spectroscopy disclosed in Embodiment 1 of the present invention.

[0095] Figure 3 This is a scatter plot of the predicted values ​​of the discrimination model constructed in S5 and the decision fusion model constructed in S6 of the sorghum variety discrimination method based on decision fusion and near-infrared spectroscopy disclosed in Embodiment 1 of the present invention. Detailed Implementation

[0096] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0097] Example 1

[0098] like Figure 1 As shown, Embodiment 1 of the present invention provides a method for distinguishing sorghum varieties based on decision fusion and near-infrared spectroscopy, comprising the following steps:

[0099] S1. Sorghum Grain Sample Collection and Processing: In this embodiment, to verify the method proposed in this invention, a dataset was designed containing 81 sorghum grain samples from 26 sorghum varieties in different regions. Samples belonging to the varieties Guojiao Hong No. 1 and Qingke Yang (or a mixture thereof) were grouped into one category, considered as target variety samples, totaling 34 samples. The remaining sorghum samples by variety and origin were grouped into another category, considered as other variety samples, totaling 47 samples. Imperfect grains were removed from each sample, and the samples were ground using a cyclone mill. Target variety samples were assigned a category label of 1, and other variety samples were assigned a label of 0.

[0100] S2. Spectral Acquisition: Spectral acquisition was conducted using a portable near-infrared spectrometer. During acquisition, the sample was placed in a sample cup with a transparent quartz glass bottom, positioned above the instrument's detection window, and spectral acquisition was performed using diffuse reflectance. The spectral range was 928-1685 nm, with a resolution of 3 nm. Each sample was acquired twice at different time points. The spectrum of the sorghum powder acquired in the first acquisition was used as the calibration set for modeling, and the spectrum acquired in the second acquisition was used as the validation set to verify the model's predictive performance.

[0101] S3. Chemical reference values ​​were obtained: The tannin and amylopectin content of each sorghum sample were determined using a wet chemical method.

[0102] S31. The wet chemical method for determining tannin content refers to the Chinese National Standard Method GB / T15686-2008.

[0103] S32. The wet chemical method for determining amylopectin content is an improvement upon the standard "NY / T55-1987 Determination of Amylose in Rice, Corn and Millet Grains" issued by the Ministry of Agriculture and Rural Affairs of the People's Republic of China. The specific steps are as follows:

[0104] 1) Preparation of standard solutions: Weigh 0.05 g of dried potato amylose (pure product) and place it in a 10 ml centrifuge tube. Add 0.5 ml of anhydrous ethanol to moisten the sample, then add 4.5 ml of 1 mol / L sodium hydroxide solution. Disperse in a boiling water bath for 10 minutes, then dilute to 10 ml with water to obtain a 5 mg / ml amylose standard solution. Weigh 0.05 g of dried sorghum amylopectin (pure product) and place it in a 10 ml centrifuge tube. Add 0.5 ml of anhydrous ethanol to moisten the sample, then add 4.5 ml of 1 mol / L sodium hydroxide solution. Disperse in a boiling water bath for 10 minutes, then dilute to 10 ml with water to obtain a 5 mg / ml amylopectin standard solution.

[0105] 2) Standard sample detection: Take six 50ml volumetric flasks and add 0, 25, 50, 100, 150, and 200 μl of 5 mg / ml amylose standard solution, respectively. Then, add 500 μl, 475 μl, 450 μl, 400 μl, 350 μl, and 300 μl of 5 mg / ml amylopectin standard solution, respectively, for a total volume of 500 μl. Take another 100ml volumetric flask and add 5ml of 0.09mol / L sodium hydroxide solution as a blank. Then, add approximately 25 ml of water, 0.5 ml of 1 mol / L acetic acid, and 0.5 ml of iodine reagent (the iodine reagent is obtained by dissolving 2 g of iodine and 20 g of potassium iodide in distilled water and diluting to 100 ml, then taking 10 ml of the solution and diluting to 100 ml) to the volumetric flasks other than the blank flask, and then dilute to 50 ml with water. Add 50 ml of water, 1 ml of 1 mol / L acetic acid, and 1 ml of iodine reagent to the blank volumetric flask, and dilute to 100 ml with water. Develop the color for 10 minutes, using the blank as a control, and read the absorbance at 640 nm and 532 nm using a spectrophotometer.

[0106] It should be noted that during the colorimetric reaction, absorption peaks in the ranges of 600-680 nm and 520-555 nm are generated, which are the characteristic absorption peaks of the amylose-iodine complex and amylopectin-iodine complex generated in the reaction, respectively. Among them, 640 nm and 532 nm are the maximum absorption wavelengths of the amylose-iodine complex and amylopectin-iodine complex, respectively, theoretically providing the best regression effect; therefore, these two absorbance values ​​were chosen. This invention adjusts the standard solution preparation and the solution ratios in the absorbance detection process disclosed in the Ministry of Agriculture and Rural Affairs of the People's Republic of China standard "NY / T 55-1987 Determination of Amylose in Rice, Corn, and Millet Grains". Then, absorbance is read at two spectral points, and a mixed calibration curve is plotted. In contrast, "NY / T 55-1987 Determination of Amylose in Rice, Corn, and Millet Grains" reads absorbance at a single spectral point and plots the calibration curve. In practical applications, the standard solution preparation and absorbance detection method disclosed in the Ministry of Agriculture and Rural Affairs of the People's Republic of China standard "NY / T55-1987 Determination of Amylose in Rice, Corn, and Millet Grains" can be directly adopted without modifying the ratios. The specific selection should be based on actual needs. The following details the aspects of the amylopectin content determination process that have been significantly improved compared to the Chinese national standard.

[0107] 3) Plotting the mixed calibration curve: The vector is constructed using the absorbance of six 50ml volumetric flasks read at 640nm. The vector is formed by the absorbance of six 50 ml volumetric flasks read at 532 nm. ,vector sum vector Construct a matrix , with matrix amylopectin concentration is the independent variable. Using amylopectin concentration as the dependent variable, a regression equation was constructed using partial least squares (PLS factor number 2). The specific steps are as follows:

[0108] S321. Extract the first PLS component:

[0109] (1) Calculation First weight vector :

[0110] First of all Using formula Standardization is carried out, and the standardized version for ,in, Represents the mean. Representative standard deviation; for amylopectin concentration Standardization is performed using the following formula.

[0111] (1)

[0112] in, For standardization ;

[0113] Solving formula Obtain the first weight vector The value of , where, For matrix The largest eigenvalue, It needs to be normalized (module length is 1). After calculation... It can be represented as in represent To construct the first score vector Contribution weight, represent To construct the first score vector The contribution weight.

[0114] (2) Calculate the first score vector Second score vector ( (First PLS component score):

[0115] make 1. Calculation ， in express The first weight vector, which is related to The matching and calculation methods are the same, except that... The calculation process and The exchange will not be elaborated upon here.

[0116] (3) Calculation First load vector and First load vector :

[0117]

[0118]

[0119] (4) Calculate the residual matrix

[0120] Explain using the first extracted component and Then calculate the remaining unexplained portion (residual), using the following formula:

[0121]

[0122]

[0123] in, express The first residual matrix, express The first residual matrix.

[0124] S322. Extract the second PLS component, including the following steps:

[0125] The residual matrix obtained from the previous iteration and As a new and Repeat all steps in S321 to extract the second component pair. The calculation yields... The second weight vector 、 The third score vector , The fourth score vector , The second load vector , The second load vector , The second residual matrix and The second residual matrix After calculation Represented as ,in, represent For building Contribution weight, represent For building The contribution weight.

[0126] S323. Establish the regression equation:

[0127] right The regression formula is as follows:

[0128] (2)

[0129] for The residual is the error term remaining after fitting the equation (2). and All The regression coefficients can be obtained using the least squares method, and the formula is as follows: , .

[0130] S324. Convert the standardized equations back to the original dimensions:

[0131] Will and use express:

[0132] (3)

[0133] (4)

[0134] According to formulas (1) and (2), we can know 。

[0135] By further combining formulas (3) and (4) and merging like terms, we can finally obtain the following formula:

[0136]

[0137] in It is the intercept. and yes The regression coefficient is given by the following formula:

[0138]

[0139]

[0140] 。

[0141] 4) Sample determination:

[0142] First, the moisture content of the sample was determined using the drying and weighing method. Next, the equivalent of 0.05g of crude starch was weighed into a 10ml centrifuge tube, and 0.5ml of anhydrous ethanol was added to thoroughly moisten the sample. Then, 4.5ml of 1mol / L sodium hydroxide solution was added, the tube was capped, and the mixture was shaken for 10 seconds. Afterward, it was dispersed in a boiling water bath for 10 minutes, rapidly cooled, and then diluted to 10ml with water. Finally, 0.5ml of the sample was pipetteed into a 50ml volumetric flask, 25ml of water was added, followed by 0.5ml of 1mol / L acetic acid solution and 0.5ml of iodine reagent. The volume was then diluted to 50ml with water, and the sample was developed for 10 minutes. The absorbance was read at 640nm and 532nm using a spectrophotometer.

[0143] 5) Result Calculation: For the absorbance of each sample at 640 nm and 532 nm, the amylopectin concentration of the sample is predicted according to the curve in 3). (%, as a percentage of total sample weight), and then calculate the amylopectin content G0 (%, as a percentage of sample dry weight) using the following formula:

[0144] (2)

[0145] in The mass of the sample used to determine the amylopectin content, The moisture percentage of the sample used to calculate the amylopectin content.

[0146] The statistical results of the tannin and amylopectin content of sorghum determined by the above chemical methods are shown in Table 1. Both tannin and amylopectin contents have a relatively wide distribution, covering most types of sorghum used for brewing.

[0147] Table 1. Statistical results of tannin and amylopectin content in sorghum correction set.

[0148]

[0149] S4. Quantitative Model Construction: A quantitative detection model between the spectral density and tannin content of each sorghum sample was constructed using partial least squares method. Similarly, a quantitative detection model between the spectral density and amylopectin content of each sorghum sample was constructed using partial least squares method.

[0150] Both sets of models were selected from different combinations of preprocessing methods and band selection algorithms. After optimization, the optimal modeling parameters were determined, as shown in Table 2.

[0151] Table 2 Modeling parameters and performance of the sorghum composition model

[0152]

[0153] Table 2 shows the cross-validation determination coefficients (R²) for the two models. 2 cv The correlation coefficients (R²) of the two models were relatively high, at 0.854 and 0.882 respectively, and the root mean square error of cross-validation (RMSECV) were relatively low, at 0.151 and 1.080 respectively, indicating good modeling performance. When using the validation set samples for prediction, the prediction correlation coefficients (R²) of the two models were relatively high. p The values ​​were 0.889 and 0.867, respectively, and the root mean square error of prediction (RMSEP) were 0.187 and 1.689, respectively, indicating that the prediction performance was also good.

[0154] S5. Qualitative Discriminant Model Construction: Qualitative discriminant models were constructed for the spectral composition and category of each sorghum sample using the discriminant partial least squares (PLS) method. The modeling parameters were as follows: near-infrared spectra at 928 nm, 996 nm, 1087 nm, 1223 nm, 1322 nm, 1345 nm, 1669 nm, and 1680 nm were used for modeling, and the PLS factor was 8. When the spectral prediction value of a sample was ≤0.5, the sample belonged to the "Other Variety Samples" category; when the spectral prediction value was >0.5, the sample belonged to the "Target Variety Samples" category.

[0155] S6. Construction of Decision Fusion Model: Using the near-infrared spectra of the model prediction calibration set obtained in steps S4 and S5, three sets of predicted values ​​are obtained (predicted tannin content, predicted amylopectin content, and predicted category classification, respectively). These three sets of predicted values ​​are used as independent variables, and the category label of the sample is used as the dependent variable. A qualitative discriminant model is constructed between them using the discriminant partial least squares method. The PLS factor number of the model is 3. This model is the decision fusion model.

[0156] S7. Prediction: For unknown sorghum samples, the sorghum is processed into powder, and spectra are collected using the same method as in step S2. The quantitative model in step S4 is used to predict its tannin and amylopectin content, and the qualitative discrimination model in step S5 is used to predict its category. Then, the decision fusion model in step S6 is used to further predict the above three sets of predicted values ​​to obtain the final category prediction value. When the spectral prediction value of a sample is ≤0.5, the sample belongs to the "Other Variety Samples" category; when the spectral prediction value of a sample is >0.5, the sample belongs to the "Target Variety Samples" category.

[0157] To clearly illustrate the advantages of the proposed method compared to traditional near-infrared qualitative discrimination methods, the decision fusion model in S6 was used as the experimental group, and the qualitative model in S5 was used as the control group. The recognition performance of the two models on the calibration set and the validation sample set was statistically analyzed, as shown in Table 3.

[0158] Table 3 Comparison of modeling and predictive performance of sorghum variety discrimination models

[0159]

[0160] Table 3 shows that the sorghum variety discrimination model optimized by conventional chemometrics methods before decision fusion had 4 misclassifications for the calibration samples, with an accuracy of 0.951 and an F1 score of 0.941. After further calculation by the decision fusion model in step S6, the number of misclassifications for the calibration samples decreased to 2, and the accuracy and F1 score improved to 0.975 and 0.971, respectively. In prediction, the control model had 5 misclassifications, with an accuracy of 0.938 and an F1 score of 0.921. After further calculation by the decision fusion model in step S6, the number of misclassifications for the predicted samples decreased to 4, and the accuracy and F1 score improved to 0.951 and 0.943, respectively. This indicates that the method of the present invention has a certain improvement in discrimination performance compared to the general near-infrared method for sorghum variety discrimination.

[0161] Figure 2 This is a scatter plot of the predicted values ​​of the experimental and control group model prediction calibration set samples. Figure 3 This is a scatter plot of the predicted values ​​for the model's validation set samples in the experimental and control groups. Figure 2 and Figure 3 As can be seen, after optimization by the method of this invention, the prediction results of the model tend to change towards the correct classification direction. A considerable number of samples belonging to other varieties tend to have lower predicted values, while samples belonging to the target variety tend to have higher predicted values. This allows samples near the classification threshold to be classified more correctly.

[0162] Through the above technical solutions, the rapid determination method for sorghum variety categories provided by the present invention has the advantages of objective and accurate detection results and rapid detection process, which can effectively improve the quality detection efficiency related to sorghum variety categories in the sorghum production, processing and storage process.

[0163] Example 2

[0164] Based on Example 1, Example 2 of the present invention also provides a sorghum variety discrimination system based on decision fusion and near-infrared spectroscopy, including:

[0165] The sample collection module is used for the collection and processing of sorghum grain samples.

[0166] The spectral acquisition module is used to acquire near-infrared spectra for each sample and construct a calibration set;

[0167] The chemical reference value acquisition module is used to determine the tannin and amylopectin content of each sample.

[0168] The quantitative model building module is used to build a first quantitative detection model between the near-infrared spectrum of a sample and the tannin content, and to build a second quantitative detection model between the near-infrared spectrum of a sample and the amylopectin content.

[0169] The qualitative discrimination model building module is used to build a qualitative discrimination model between the near-infrared spectrum of a sample and its class classification.

[0170] The decision fusion model construction module is used to predict the calibration set using the first quantitative detection model, the second quantitative detection model, and the qualitative discrimination model respectively, to obtain three sets of predicted values. These three sets of predicted values ​​are used as independent variables, and the class label of the sample is used as the dependent variable to construct the decision fusion model.

[0171] The result prediction module is used to acquire spectra of unknown sorghum samples in the same way as the spectral acquisition module. The first and second quantitative detection models in the quantitative model construction module are used to predict the tannin content and amylopectin content, respectively. The qualitative discriminant model in the qualitative discriminant model construction module is used to predict the category. Then, the decision fusion model in the decision fusion model construction module is used to further predict the above three sets of predicted values ​​to obtain the final category prediction value. When the final category prediction value is ≤0.5, the sample belongs to other varieties, and when the final category prediction value is >0.5, the sample belongs to the target variety.

[0172] Specifically, the sample acquisition module is also used for:

[0173] A dataset is constructed by collecting more than 60 complete and mature sorghum grain samples with known cultivar affiliations. The dataset includes samples of a single variety, multiple varieties, or a mixture of multiple varieties of interest to the user. These samples are considered as target variety samples, with more than 30 target variety samples. Samples other than target variety samples in the dataset are considered as other variety samples, with more than 30 other variety samples. The sorghum grains of each target variety sample and other variety samples are processed into powder using a mill. The target variety samples are assigned a category label of 1, and the other variety samples are assigned a category label of 0.

[0174] Specifically, the spectral acquisition module is also used for:

[0175] Near-infrared spectra of each sample were acquired using a diffuse reflectance method on a near-infrared spectrometer. Each sample was acquired twice at different time periods. The near-infrared spectra of the first acquired sample were used as the calibration set for modeling, and the near-infrared spectra of the second acquired sample were used as the validation set to verify the predictive performance of the decision fusion model.

[0176] Specifically, the chemical reference value acquisition module is also used for:

[0177] The tannin and amylopectin content of each sorghum sample was determined using a wet chemical method.

[0178] More specifically, the method for determining the amylopectin content includes:

[0179] The preparation of standard solutions and absorbance detection were carried out according to the standard NY / T 55-1987 of the Ministry of Agriculture and Rural Affairs of the People's Republic of China. A spectrophotometer was used to read the absorbance of the prepared standard solutions at 640 nm and 532 nm. The absorbance read at 640 nm was used to construct a vector. The vector is formed by the absorbance read at 532 nm. ,vector sum vector Construct a matrix , with matrix amylopectin concentration is the independent variable. Using amylopectin as the dependent variable, a partial least squares regression equation for amylopectin concentration was constructed. The absorbance of the sample at 640 nm and 532 nm was read and substituted into the regression equation to obtain the amylopectin concentration. Using the formula Obtaining amylopectin content in, The quality of the sample to be tested, This represents the moisture percentage of the sample to be tested.

[0180] More specifically, the process of constructing the regression equation for amylopectin concentration is as follows:

[0181] S321, Regarding the matrix Using formula Standardize, For standardization , Represents the mean. Representative standard deviation; for amylopectin concentration Standardization is performed using the following formula.

[0182] (1)

[0183] in, For standardization ;

[0184] Solving formula get First weight vector The value of , where, For matrix The largest eigenvalue, after calculation Represented as in represent To construct the first score vector Contribution weight, represent To construct the first score vector Contribution weight;

[0185] Calculate the first score vector Second score vector ,in express The first weight vector;

[0186] calculate First load vector and First load vector

[0187] calculate The first residual matrix and The first residual matrix ;

[0188] S322, The residual matrix obtained from the previous iteration is... and As a new and Repeat all steps in S321 to obtain The second weight vector 、 The third score vector , The fourth score vector , The second load vector , The second load vector , The second residual matrix and The second residual matrix After calculation Represented as ,in, represent For building Contribution weight, represent For building Contribution weight;

[0189] S323, to The regression formula is as follows:

[0190] (2)

[0191] in, for The residual, and All The regression coefficients, , ;

[0192] S324, will and use express:

[0193] (3)

[0194] (4)

[0195] According to formulas (1) and (2), Combining formulas (3) and (4), and after merging like terms, the final formula is as follows:

[0196]

[0197] in, It is the intercept. and yes The regression coefficient is given by the following formula:

[0198]

[0199]

[0200] 。

[0201] Specifically, both the first and second quantitative detection models are modeled using partial least squares (PLS). The modeling parameters for the first quantitative detection model are as follows: the sample spectrum is first smoothed using the first derivative of the Savitsky-Golay method with 17 points, and then the near-infrared spectrum in the ranges of 928-1008nm, 1174-1328nm, and 1550-1617nm is extracted for modeling, and the PLS factor number during modeling is 9.

[0202] The modeling parameters of the second quantitative detection model are as follows: First, the sample spectrum is smoothed using the first derivative of the Savitsky-Golay method with 17 points. Then, the near-infrared spectra at 938nm, 960nm, 1034-1037nm, 1043-1052nm, 1062-1065nm, 1109-1118nm, 1124-1128nm, 1156-1161nm, 1214-1223nm, 1244nm, 1250nm, 1259nm, 1265nm, 1277nm, 1315-1325nm, 1378-1381nm, 1430nm, 1445nm, 1456nm, 1502nm, 1508nm, 1513nm, 1653nm, and 1686nm are used for modeling, and the PLS factor number during modeling is 10.

[0203] Specifically, the qualitative discrimination model is modeled using partial least squares, and the modeling parameters of the qualitative discrimination model are:

[0204] Near-infrared spectra at 928nm, 996nm, 1087nm, 1223nm, 1322nm, 1345nm, 1669nm, and 1680nm were used for modeling, and the PLS factor was 8 during modeling.

[0205] Specifically, the decision fusion model is modeled using discriminant partial least squares (PLS) and the modeling parameter of the decision fusion model is 3 PLS factors.

[0206] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for discriminating sorghum varieties based on decision fusion and near infrared spectroscopy, characterized by, The method comprises the following steps: S1. Sample collection and processing of sorghum grains Collect more than 60 samples of complete mature sorghum grains of known varieties to form a data set, which includes single variety samples or multiple variety samples or mixed samples of multiple varieties that users are interested in, which are regarded as target variety samples, and there are more than 30 target variety samples in the data set; samples other than the target variety samples in the data set are regarded as other variety samples, and there are more than 30 other variety samples; use a flour mill to process the sorghum grains of each target variety sample and other variety sample into powder; give the target variety samples a category label 1 and give the other variety samples a category label 0; S2. Collect near-infrared spectra of each sample to construct a calibration set; S3. Determine the tannin content and amylopectin content of each sample; S4. Construct a first quantitative detection model between the near-infrared spectra of the samples and the tannin content, and construct a second quantitative detection model between the near-infrared spectra of the samples and the amylopectin content; S5. Construct a qualitative discrimination model between the near-infrared spectra of the samples and the category attribution; S6. Use the first quantitative detection model, the second quantitative detection model and the qualitative discrimination model to respectively predict the calibration set to obtain three groups of prediction values, and use the three groups of prediction values as independent variables and the category labels of the samples as dependent variables to construct a decision fusion model; S7. For unknown sorghum samples, collect spectra in the same way as in S2, predict the tannin content and amylopectin content of each sample using the first quantitative detection model and the second quantitative detection model in S4, predict the category attribution of each sample using the qualitative discrimination model in S5, and then further predict the three groups of prediction values using the decision fusion model in S6 to obtain the final category prediction value. When the final category prediction value is ≤0.5, the sample belongs to other varieties, and when the final category prediction value is >0.5, the sample belongs to the target variety.

2. The method of discriminating sorghum varieties based on decision fusion and near infrared spectroscopy according to claim 1, characterized in that, The S2 comprises: Collect the near-infrared spectra of each sample on a near-infrared spectrometer in a diffuse reflection mode, and collect each sample twice in different time periods, use the first collected near-infrared spectra of the sample as a calibration set for modeling, and use the second repeatedly collected near-infrared spectra of the sample as a validation set for verifying the prediction performance of the decision fusion model.

3. The method of claim 1, wherein the method is characterized by: The S3 comprises: Determine the tannin content and amylopectin content of each sorghum sample by using a wet chemical method.

4. The method of claim 3, wherein the method is characterized by, The method for determining the amylopectin content comprises: The standard solution is prepared according to the standard NY / T 55-1987 of the Ministry of Agriculture and Rural Affairs of the People's Republic of China, and the absorbance is detected; the spectrophotometer is used to read the absorbance of the prepared standard solution at 640 nm and 532 nm, and the absorbance read at 640 nm constitutes a vector and the absorbance read at 532 nm constitutes a vector , the vector and the vector constitute a matrix , the matrix is the independent variable, the amylopectin concentration is the dependent variable, and the amylopectin concentration regression equation is constructed by using the partial least squares method; the absorbance of the sample to be measured at 640 nm and 532 nm is read, and the amylopectin concentration is obtained by substituting the amylopectin concentration regression equation, and the amylopectin content is obtained by using the formula wherein, the mass of the sample to be measured, is the moisture percentage of the sample to be measured.

5. The method of discriminating sorghum varieties based on decision fusion and near infrared spectroscopy according to claim 4, characterized in that, The construction process of the amylopectin concentration regression equation is as follows: S321、to the matrix using the formula standardized, for the standardized , representing the mean, representing the standard deviation; for amylopectin concentration using the following formula (1) wherein is the standardized ; solving formula obtained a first weight vector of the value, wherein is the maximum eigenvalue of the matrix , after the calculation is expressed as wherein represent the contribution weight of to build the first score vector represent the contribution weight of to build the first score vector computing a first score vector and a second score vector wherein denotes a first weight vector of computing the first load vector and the first load vector Computing the first residual matrix and the first residual matrix ; S322, the residual matrix obtained in the last iteration and as the new and , repeat all steps of S321 to obtain the second weight vector of 、 the third score vector of , the fourth score vector of , the second load vector of , the second load vector of , the second residual matrix of and the second residual matrix of , the calculated is expressed as , wherein represents the contribution weight of to the construction of , and represents the contribution weight of to the construction of ; S323、to Regression was performed, the formula of which is: （2） wherein is the residual, and are regression coefficients, , , ; S324、to and with denotes: （3） （4） According to formulas (1) and (2), , combined with formulas (3) and (4), after merging the same items, the final formula is obtained as follows: wherein, is the intercept, and is the regression coefficient, whose specific formula is: 。 6. The method of discriminating sorghum varieties based on decision fusion and near infrared spectroscopy according to claim 1, characterized in that, Both the first quantitative detection model and the second quantitative detection model are modeled by using a partial least squares method, and the modeling parameters of the first quantitative detection model are as follows: first, use a Savitsky-Golay method to perform 1st order derivative 17-point smoothing processing on the sample spectrum, then intercept the near-infrared spectra in the ranges of 928-1008 nm, 1174-1328 nm and 1550-1617 nm for modeling, and the number of PLS factors during modeling is 9. The modeling parameters of the second quantitative detection model are: first derivative 17-point smoothing processing of the sample spectrum by the Savitsky-Golay method, then near-infrared spectrum at 938 nm, 960 nm, 1034-1037 nm, 1043-1052 nm, 1062-1065 nm, 1109-1118 nm, 1124-1128 nm, 1156-1161 nm, 1214-1223 nm, 1244 nm, 1250 nm, 1259 nm, 1265 nm, 1277 nm, 1315-1325 nm, 1378-1381 nm, 1430 nm, 1445 nm, 1456 nm, 1502 nm, 1508 nm, 1513 nm, 1653 nm and 1686 nm is used for modeling, and the PLS factor number is 10 during modeling.

7. The method of discriminating sorghum varieties based on decision fusion and near infrared spectroscopy according to claim 1, characterized in that, The qualitative discrimination model is modeled by the partial least squares method, and the modeling parameters of the qualitative discrimination model are: The near-infrared spectrum at 928 nm, 996 nm, 1087 nm, 1223 nm, 1322 nm, 1345 nm, 1669 nm and 1680 nm is used for modeling, and the PLS factor number is 8 during modeling.

8. The method of discriminating sorghum varieties based on decision fusion and near infrared spectroscopy according to claim 1, characterized in that, The decision fusion model is modeled by the discriminant partial least squares method, and the modeling parameters of the decision fusion model are that the PLS factor number is 3.

9. A discrimination system for sorghum varieties based on decision fusion and near infrared spectroscopy, characterized by, It comprises: A sample collection module for collecting and processing sorghum grain samples; Collect more than 60 complete mature sorghum grain samples of known varieties to form a data set, which includes single variety samples or multiple variety samples or mixed samples of multiple varieties that the user is interested in. These samples are regarded as target variety samples, and there are more than 30 target variety samples; samples other than the target variety samples in the data set are regarded as other variety samples, and there are more than 30 other variety samples; use a flour mill to process the sorghum grains of each target variety sample and other variety sample into powder; give the target variety sample a class label 1 and the other variety sample a class label 0; A spectrum collection module for collecting near-infrared spectrum of each sample to build a calibration set; A chemical reference value acquisition module for determining the tannin content and amylopectin content of each sample; A quantitative model construction module for constructing a first quantitative detection model between the near-infrared spectrum and the tannin content of the sample, and a second quantitative detection model between the near-infrared spectrum and the amylopectin content of the sample; A qualitative discrimination model construction module for constructing a qualitative discrimination model between the near-infrared spectrum and the class attribution of the sample; A decision fusion model construction module for predicting the calibration set by using the first quantitative detection model, the second quantitative detection model and the qualitative discrimination model respectively to obtain three groups of predicted values, taking the three groups of predicted values as independent variables and the class label of the sample as dependent variable, and constructing a decision fusion model; The result prediction module is configured to collect the spectrum of the unknown sorghum sample in the same manner as in the spectrum collection module, to predict the tannin content and amylopectin content of the sample by using the first quantitative detection model and the second quantitative detection model in the quantitative model construction module, to predict the category of the sample by using the qualitative discrimination model in the qualitative discrimination model construction module, and to further predict the three groups of prediction values by using the decision fusion model in the decision fusion model construction module to obtain a final category prediction value. When the final category prediction value is less than or equal to 0.5, the sample belongs to other varieties, and when the final category prediction value is greater than 0.5, the sample belongs to the target variety.

Images