Soybean activity detection method fusing UVE feature selection and sparse Gaussian process regression

Abstract

The invention discloses a soybean activity detection method fusing UVE feature selection and sparse Gaussian process regression, and relates to the technical field of agricultural product quality detection. The method comprises the following steps: performing grouping gradient aging treatment on fresh soybean seeds, and obtaining three types of vigor data including germination potential, germination rate and germination index through a paper germination test; then, collecting a near infrared spectrum of a sample, and carrying out Savitzky-Golay and SNV pretreatment on the near infrared spectrum of the sample; determining a variable stability threshold value of each vitality index based on an improved PLSR model, and screening a data set through UVE; dividing a data set, training an SGPR model through a Matern32 kernel function, a K-means clustering optimization induction point and a maximized edge likelihood function, and carrying out verification set optimization and test set verification; after a sample to be detected is subjected to spectrum processing and feature screening, three types of vitality index predicted values are output through the exclusive SGPR model. According to the invention, nondestructive and high-precision detection of soybean activity is realized, and the problems of high time consumption and strong destructiveness of the existing method are solved.

Description

Technical Field

[0001] This invention relates to the field of agricultural product quality testing technology, specifically to a soybean vigor detection method that integrates UVE feature selection and sparse Gaussian process regression. Background Technology

[0002] Soybeans, as an important dual-purpose crop for both grain and oil, rely heavily on seed vigor, which directly determines germination rate, uniform growth, and yield. Seed vigor testing is a core component of seed production, storage, and field application. Traditional methods include germination tests, low-temperature stress methods, and conductivity measurements. Germination tests require 7-14 days to collect germination potential and germination rate data, resulting in a long testing cycle and cumbersome operation. While low-temperature stress and conductivity measurements can shorten the testing time, they still damage the seed structure, failing to achieve non-destructive testing and thus hindering the practical needs of large-scale rapid seed screening. To address these limitations, near-infrared spectroscopy, with its non-destructive, rapid, and efficient advantages, has become a research hotspot in the field of agricultural product quality testing.

[0003] Chinese patent (publication number CN119138146B) discloses a near-infrared spectral prediction method for soybean seed vigor based on machine learning and deep learning algorithms. This method constructs a model using XGBoost and BP neural network algorithms to predict multiple indicators such as germination rate, germination potential, and germination index of soybeans with light and dark seed coats, thus improving the model's universality. However, the BP neural network used in this method is a black box model, resulting in poor interpretability of the results. It also requires a large number of samples for training and is not adaptable to small sample scenarios. The feature band extraction relies on Pearson correlation analysis, which is not very specific and fails to effectively remove redundant information in high-dimensional spectra, thus limiting the modeling efficiency.

[0004] Furthermore, while existing research can handle nonlinear relationships well and improve the prediction accuracy of maize seed germination rate through hybrid kernel function optimization, the computational complexity is mostly O(n³) or higher. This results in extremely low modeling efficiency in high-dimensional near-infrared spectral data scenarios, making it difficult to meet the actual needs of rapid soybean seed detection. At the same time, some technologies only focus on predicting a single vigor index for a single crop, ignoring the need for collaborative evaluation of multiple varieties and multiple indicators, and have not been specifically optimized for small-sample, high-dimensional soybean spectral data, resulting in insufficient generalization ability.

[0005] In summary, existing seed vigor detection technologies still suffer from problems such as time-consuming and destructive testing methods, making them unsuitable for large-scale rapid testing needs. Furthermore, feature selection methods lack specificity, redundant information in high-dimensional spectral data is not effectively removed, and some technologies focus only on a single indicator or a single crop, making it difficult to comprehensively and accurately reflect the vigor level of soybean seeds.

[0006] Therefore, developing a soybean seed vigor detection method that is non-destructive, rapid, highly accurate, efficient in modeling, and capable of evaluating multiple indicators has become a pressing technical challenge in this field. Summary of the Invention

[0007] Based on the aforementioned technical problems, this application discloses a soybean vigor detection method that integrates UVE feature selection and sparse Gaussian process regression, specifically including: Fresh soybean seeds were grouped and aged within a preset period to obtain a grouped vigor gradient sample. Paper germination test was performed on all samples, and the germination potential (GE), germination rate (GP), and germination index (GI) were calculated to obtain soybean vigor data. Each sample was scanned using a near-infrared spectrometer, and the near-infrared spectral data of soybeans were obtained after preprocessing. Based on germination potential (GE), germination rate (GP), and germination index (GI), the stability thresholds of the variables corresponding to each index are determined by an improved partial least squares regression (PLSR) model. The improved PLSR model is optimized by leave-one-out cross-validation.

[0008] Based on variable stability thresholds and soybean near-infrared spectral data, non-informative variables with wavelength values ​​below the corresponding thresholds were removed, resulting in a dataset highly correlated with each vitality index. The dataset is divided into a training set, a validation set, and a test set. A sparse Gaussian process regression (SGPR) model is trained using the training set and the model parameters are optimized using the validation set. Based on the sparse Gaussian process regression (SGPR) model, the soybean vigor detection results of the soybean seeds to be tested are output, including the germination potential, germination rate, and predicted germination index of the soybean seeds.

[0009] Preferably, the soybean vitality data is obtained by: setting a preset aging period of... In the germination test Germination potential (GE) of each group of samples was counted on day 2. The germination rate (GP) of each group of samples was statistically analyzed daily. The germination index (GI) is calculated based on the number of germinated seeds and the number of germination days. The formula is as follows:

[0010] in, For soybean seed samples in the germination test The number of seeds that germinate in a day This refers to the number of days since germination. The number of days for the germination test. The statistical period for the germination experiment; By summarizing the values ​​of germination potential, germination rate, and germination index, complete soybean vitality data can be obtained.

[0011] Preferably, the soybean near-infrared spectral data is obtained by performing a single spectral scan on each soybean seed sample, completing the first round of scanning of all samples according to preset parameters, repeating the scanning operation, and averaging the multiple scan data of each sample to obtain the original near-infrared spectral data of each sample. Savitzky-Golay convolution smoothing combined with standard normal variable transformation was used to process the raw near-infrared spectral data to eliminate interference factors such as baseline drift, random noise, and scattering effects, correct spectral data deviation, and obtain smooth and accurate soybean near-infrared spectral data.

[0012] Preferably, determining the variable stability threshold for each indicator specifically involves: performing data matching and partitioning on the measured data and near-infrared spectral data of the three vitality indicators to obtain three independent modeling datasets; constructing partial least squares regression (PLSR) models based on the three datasets; optimizing each PLSR model using leave-one-out cross-validation; introducing random noise variables into the three optimized PLSR models based on leave-one-out cross-validation and calculating their stability values; and obtaining the variable stability threshold based on the maximum stability value.

[0013] Preferably, the construction of the partial least squares regression (PLSR) model is specifically as follows: the preprocessed near-infrared spectral data is used as the independent variable and matched with the measured data of three vigor indicators, namely germination potential (GE), germination rate (GP), and germination index (GI), to obtain three independent modeling datasets: GE-spectrum, GP-spectrum, and GI-spectrum. Based on three datasets, partial least squares regression (PLSR) models were constructed to extract potential correlation factors between spectral data and various vitality indicators. A linear regression relationship between spectral characteristics and soybean vitality indicators was established, as shown in the formula:

[0014] in, This is a predicted value for the vitality index. For the intercept term, For the first The regression coefficients of the wavelength variable, For the preprocessed first Spectral values ​​at each wavelength point, This represents the total number of spectral wavelength variables. Fit the residuals to the model; The optimal number of latent variables for each PLSR model is determined based on minimizing the root mean square error of cross-validation (RMSECV), using the following formula:

[0015] in, To model the total number of samples, For the first Measured values ​​of the vitality index for each sample. For the first Cross-validation predicted values ​​of the vitality index for each sample.

[0016] Preferably, the variable stability threshold based on the maximum stability value is obtained by: optimizing each PLSR model by leave-one-out cross-validation, using one sample in each dataset as the validation set and the remaining samples as the training set, repeatedly training and validating the model until all samples are traversed, and determining the optimal number of latent variables for each PLSR model. Based on the three PLSR models optimized using leave-one-out cross-validation, random noise variables were introduced into each model, and their stability values ​​were calculated. The maximum absolute value of the stability value of the noise component was extracted and set as the variable stability threshold for the corresponding vitality index. The variable stability thresholds for the three indices GE, GP, and GI were obtained using the following formulas:

[0017]

[0018] in, For the first The stability threshold of a vitality indicator variable. The number of iterations for leave-one-out cross-validation for the PLSR model. For the first In the cross-validation, the projected importance value of the random noise variable. The original number of wavelength variables. For the first The variance explained by each latent variable to the dependent variable For the i-th cross-validation, the i-th Weighting coefficients of noise variables under each latent variable. For the first The magnitude of the weight coefficients of all variables under one latent variable. This represents the optimal number of latent variables in the PLSR model.

[0019] Preferably, the dataset is specifically composed of: based on the stability thresholds of variables such as germination potential (GE), germination rate (GP), and germination index (GI), combined with preprocessed soybean near-infrared spectral data, wavelengths with no or low information are filtered out through targeted wavelength variable screening. For each vigor indicator, its specific variable stability threshold is used as the judgment criterion to screen the corresponding wavelength point-stability value dataset, removing wavelength variables with stability values ​​lower than the corresponding threshold and retaining wavelength variables with stability values ​​higher than the corresponding threshold. Highly correlated feature wavelengths are extracted for each vigor indicator to form the corresponding dataset.

[0020] Preferably, the sparse Gaussian process regression (SGPR) model is specifically configured as follows: the dataset is divided into three groups of datasets according to GE, GP, and GI, and the dataset is further divided to obtain a training set, a validation set, and a test set; The SGPR model was trained using the training set data, and the Gaussian process posterior distribution was constructed using the fully independent training condition FITC approximation. The hyperparameter set was optimized by maximizing the marginal likelihood function based on the validation set data. The generalization ability of the model was verified using the test set data. Data collection and feature processing were performed on the soybean seeds to be tested to obtain a dataset of soybean seeds to be tested. The dataset was then input into the GE / GP / GI dedicated SGPR model that had been trained and optimized, and the predicted values ​​of germination potential, germination rate and germination index of the test samples were output.

[0021] Preferably, the partitioning of the dataset is specifically achieved by simultaneously minimizing the Euclidean distance between the spectral and vitality indices, as shown in the formula:

[0022] in, For the sample With sample The overall distance , Samples , The characteristic wavelength spectral value vector, , Samples , The vitality index value, This is for Euclidean distance calculation. To balance the distance weighting of spectral and vitality indicators; The training set, validation set, and test set are divided according to a preset ratio, and samples that are far apart and scattered are assigned to different subsets.

[0023] Preferably, the training of the SGPR model specifically involves: For each dataset, an SGPR model was initialized separately. The model was set with characteristic wavelength spectral values ​​as input variables and vitality index values ​​as output variables. An induction point set was introduced to reduce the computational complexity of the model. K-means clustering combined with the elbow rule was used to cluster the characteristic wavelength data of each training set. The relative decrease rate of the sum of squares within each cluster was calculated to determine the optimal number of clusters for the GE / GP / GI model. The sample closest to the cluster center in each cluster was selected as the model induction point. The Matern32 anisotropic kernel function was selected as the kernel function of the SGPR model, and the hyperparameter set was initialized using the following formula:

[0024]

[0025] in, For kernel function, The standard deviation of the signal. The kernel function length scale, These are the weighting coefficients for each feature wavelength dimension. The dimension of the characteristic wavelength. For anisotropic Euclidean distance, Weights for each feature wavelength dimension, , The first , The soybean sample in the first Near-infrared spectral response values ​​at each characteristic wavelength position; The SGPR model is trained using the training set data; the hyperparameter set is optimized by maximizing the marginal likelihood function using the validation set data as a benchmark, and the parameters are iteratively updated until the RMSE prediction error of the validation set reaches the minimum value, thus determining the optimal hyperparameters for each model; the generalization ability of the model is verified using the test set data, and the sparse Gaussian process regression SGPR model is obtained after the accuracy reaches the target.

[0026] Compared with the prior art, the technical solution of this application has the following technical effects: This invention collects sample spectral data using near-infrared spectroscopy technology without damaging the seed structure, and the subsequent modeling and prediction process is efficient and convenient. Combined with the sparsity optimization of the SGPR model, the modeling and detection time is significantly shortened, making it suitable for the actual production needs of large-scale rapid screening of soybean seeds. Compared with the traditional germination test of 7-14 days, the detection efficiency is improved by more than 80%.

[0027] This invention replaces the existing linear PLSR model and the traditional GPR model with a sparse Gaussian process regression (SGPR) model. It can accurately capture the complex nonlinear relationship between near-infrared spectral data and soybean vitality indicators. Furthermore, by optimizing the selection of induced points, the computational complexity is reduced from O(n³) to O(nm²). Since m is much smaller than n, the computational complexity is greatly reduced, solving the problem of low efficiency in traditional GPR modeling. At the same time, compared with black-box models such as BP neural networks, this application uses the PLSR model to determine the variable stability threshold and UVE to select feature wavelengths. The model logic is clear, highly interpretable, and has excellent adaptability to small sample scenarios, achieving high-precision prediction without the need for training with a large number of samples.

[0028] This invention constructs dedicated PLSR models to determine variable stability thresholds using three core vigor indicators: germination potential (GE), germination rate (GP), and germination index (GI). By using UVE to target and eliminate non-informative variables, redundant information in high-dimensional spectral data is effectively eliminated, improving the model's prediction accuracy. Compared to existing technologies that only focus on a single vigor indicator, this application simultaneously outputs predicted values ​​for three key vigor indicators, which can comprehensively and accurately reflect the vigor level of soybean seeds, overcoming the shortcomings of the one-sided evaluation in existing technologies.

[0029] This invention obtains samples with eight vigor gradients through a 0-7 day gradient aging process. The modeling samples cover the entire range of soybean seeds from high to low vigor, ensuring that the model can adapt to soybean seeds with different aging levels. Compared with existing technologies that only target a single crop and a single vigor level, this application specifically optimizes the kernel function and feature selection process for soybean seeds, resulting in stronger generalization ability. It can be directly applied to vigor detection in multiple stages of soybean seed production and storage.

[0030] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0031] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0033] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows: Figure 1 The overall flowchart of the soybean vigor detection method that integrates UVE feature selection and sparse Gaussian process regression; Figure 2 The overall architecture diagram of the soybean vigor detection method that integrates UVE feature selection and sparse Gaussian process regression; Figure 3 This is a diagram of the architecture of the improved partial least squares regression (PLSR) model in this application. Figure 4 This is a diagram of the architecture of the sparse Gaussian process regression (SGPR) model in this application; Figure 5 This is a wavelength point distribution diagram after feature wavelength screening in the embodiments of this application; Figure 6 This is a comparison chart of grouped prediction results for different aging days in the embodiments of this application; Figure 7 This is a comparison chart of overall performance index data in the embodiments of this application. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0035] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0036] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0037] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0038] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0039] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0040] Example 1 mainly describes a soybean vigor detection method that integrates UVE feature selection and sparse Gaussian process regression. Figure 1 , Figure 2 As shown, it specifically includes: Fresh soybean seeds were divided into 8 groups and aged for 0-7 days to obtain 8 vigor gradient samples. Paper germination test was performed on all samples, and the germination potential (GE), germination rate (GP), and germination index (GI) were calculated to obtain soybean vigor data. Each sample was scanned using a near-infrared spectrometer, and the near-infrared spectral data of soybeans were obtained after preprocessing. Based on germination potential (GE), germination rate (GP), and germination index (GI), the stability thresholds of the variables corresponding to each index are determined by an improved partial least squares regression (PLSR) model. The improved PLSR model is optimized by leave-one-out cross-validation.

[0041] Based on variable stability thresholds and soybean near-infrared spectral data, non-informative variables with wavelength values ​​below the corresponding thresholds were removed, resulting in a dataset highly correlated with each vitality index. The dataset is divided into a training set, a validation set, and a test set. A sparse Gaussian process regression (SGPR) model is trained using the training set and the model parameters are optimized using the validation set. Based on the sparse Gaussian process regression (SGPR) model, the soybean vigor detection results of the soybean seeds to be tested are output, including the germination potential, germination rate, and predicted germination index of the soybean seeds.

[0042] Furthermore, the soybean vigor data were obtained as follows: the germination potential (GE) of each group of samples was calculated on the 4th day of the germination experiment, and the germination rate (GP) of each group of samples was calculated on the 7th day. The germination index (GI) is calculated based on the number of germinated seeds and the number of germination days. The formula is as follows:

[0043] in, For soybean seed samples in the germination test The number of seeds that germinate in a day This refers to the number of days since germination. The number of days for the germination test. The statistical period for the germination experiment; By summarizing the values ​​of germination potential, germination rate, and germination index, complete soybean vitality data can be obtained.

[0044] Furthermore, the soybean near-infrared spectral data are obtained as follows: a single spectral scan is performed on each soybean seed sample. After completing the first round of scanning of all samples according to preset parameters, the scanning operation is repeated. The average value of the multiple scan data of each sample is taken to obtain the original near-infrared spectral data of each sample. Savitzky-Golay convolution smoothing combined with standard normal variable transformation was used to process the raw near-infrared spectral data to eliminate interference factors such as baseline drift, random noise, and scattering effects, correct spectral data deviation, and obtain smooth and accurate soybean near-infrared spectral data.

[0045] Furthermore, the scanning parameters of the near-infrared spectrometer are: spectral range Spectral resolution Each sample was scanned at least 16 times and the average value was taken to obtain the raw spectral data. The preprocessing was to process the raw spectral data by Savitzky-Golay convolution smoothing combined with standard normal variable transformation. The window size of the convolution smoothing was 21 and the polynomial degree was 2.

[0046] Furthermore, the stability thresholds for each indicator were determined as follows: the measured data and near-infrared spectral data of the three vitality indicators were matched and divided to obtain three independent modeling datasets; based on the three datasets, partial least squares regression (PLSR) models were constructed respectively; leave-one-out cross-validation was used to optimize each PLSR model; based on the three PLSR models optimized by leave-one-out cross-validation, random noise variables were introduced and their stability values ​​were calculated, and the stability thresholds for the variables were obtained based on the maximum stability value.

[0047] Furthermore, partial least squares regression (PLSR) models were constructed. Specifically, the preprocessed near-infrared spectral data were used as independent variables and matched with the measured data of three vigor indicators: germination potential (GE), germination rate (GP), and germination index (GI), to obtain three independent modeling datasets: GE-spectrum, GP-spectrum, and GI-spectrum. Based on three datasets, partial least squares regression (PLSR) models were constructed to extract potential correlation factors between spectral data and various vitality indicators. A linear regression relationship between spectral characteristics and soybean vitality indicators was established, as shown in the formula:

[0048] in, This is a predicted value for the vitality index. For the intercept term, For the first The regression coefficients of the wavelength variable, For the preprocessed first Spectral values ​​at each wavelength point, This represents the total number of spectral wavelength variables. Fit the residuals to the model; The optimal number of latent variables for each PLSR model is determined based on minimizing the root mean square error of cross-validation (RMSECV), using the following formula:

[0049] in, To model the total number of samples, For the first Measured values ​​of the vitality index for each sample. For the first Cross-validation predicted values ​​of the vitality index for each sample.

[0050] Furthermore, the variable stability threshold is obtained based on the maximum stability value. Specifically, the model is optimized by leave-one-out cross-validation for each PLSR model. One sample in each dataset is used as the validation set, and the remaining samples are used as the training set. The model is repeatedly trained and validated until all samples are traversed to determine the optimal number of latent variables for each PLSR model. Based on the three PLSR models optimized using leave-one-out cross-validation, random noise variables were introduced into each model, and their stability values ​​were calculated. The maximum absolute value of the stability value of the noise component was extracted and set as the variable stability threshold for the corresponding vitality index. The variable stability thresholds for the three indices GE, GP, and GI were obtained using the following formulas:

[0051]

[0052] in, For the first The stability threshold of a vitality indicator variable. The number of iterations for leave-one-out cross-validation for the PLSR model. For the first In the cross-validation, the projected importance value of the random noise variable. The original number of wavelength variables. For the first The variance explained by each latent variable to the dependent variable For the i-th cross-validation, the i-th Weighting coefficients of noise variables under each latent variable. For the first The magnitude of the weight coefficients of all variables under one latent variable. This represents the optimal number of latent variables in the PLSR model.

[0053] Furthermore, the optimal number of latent variables for leave-one-out cross-validation of the PLSR model is 19 for GE, 18 for GP, and 14 for GI; the variable stability thresholds are 48.54 for GE, 45.31 for GP, and 76.4385 for GI (all of which are the maximum absolute values ​​of the noise part of the corresponding model).

[0054] Furthermore, the dataset is specifically designed as follows: based on the stability thresholds of variables such as germination potential (GE), germination rate (GP), and germination index (GI), combined with preprocessed soybean near-infrared spectral data, non-informative and low-information wavelengths are removed through targeted wavelength variable screening. For each vigor indicator, its specific variable stability threshold is used as the criterion to screen the corresponding "wavelength point-stability value" dataset, removing wavelength variables with stability values ​​below the corresponding threshold (these variables are non-informative variables unrelated to the vigor indicator), and retaining wavelength variables with stability values ​​above the corresponding threshold (these variables are effective variables highly correlated with the vigor indicator). Highly correlated feature wavelengths are extracted for each vigor indicator to form the corresponding dataset.

[0055] Furthermore, targeted wavelength variable screening is used to eliminate uninformative and low-information wavelengths. Specifically, for each vitality indicator, the corresponding "wavelength point-stability value" dataset is screened based on its specific variable stability threshold. Wavelength variables with stability values ​​below the corresponding threshold (these variables are uninformative variables unrelated to the vitality indicator) are eliminated, while wavelength variables with stability values ​​above the corresponding threshold (these variables are effective variables highly correlated with the vitality indicator) are retained.

[0056] Furthermore, the specific process of generating and organizing the datasets is as follows: the effective wavelength variables obtained after screening the sub-indicators are summarized, organized by wavelength point number, and germination potential GE dataset, germination rate GP dataset, and germination index GI dataset are generated respectively. It is also verified that all feature wavelengths cover the characteristic absorption peaks of soybean near-infrared spectrum to ensure the effectiveness and representativeness of feature wavelengths. The spectral data corresponding to each dataset is extracted to form "feature wavelength-spectral value" modeling data that matches each vitality index, providing highly relevant and low-redundancy input data for the subsequent training and optimization of the SGPR model.

[0057] Furthermore, the Sparse Gaussian Process Regression (SGPR) model specifically involves dividing the dataset into three groups: GE, GP, and GI, and then further dividing the dataset to obtain the training set, validation set, and test set. The SGPR model was trained using the training set data, and the Gaussian process posterior distribution was constructed using the fully independent training condition FITC approximation. The hyperparameter set was optimized by maximizing the marginal likelihood function based on the validation set data. The generalization ability of the model was verified using the test set data. Data collection and feature processing were performed on the soybean seeds to be tested to obtain a dataset of soybean seeds to be tested. The dataset was then input into the GE / GP / GI dedicated SGPR model that had been trained and optimized, and the predicted values ​​of germination potential, germination rate and germination index of the test samples were output.

[0058] Furthermore, the partitioning of the dataset specifically involves simultaneously minimizing the Euclidean distance between the spectral and vitality indices, as shown in the formula:

[0059] in, For the sample With sample The overall distance , Samples , The characteristic wavelength spectral value vector, , Samples , The vitality index value, This is for Euclidean distance calculation. To balance the distance weighting of spectral and vitality indicators; The training set, validation set, and test set are divided according to a preset ratio, and samples that are far apart and scattered are assigned to different subsets.

[0060] Furthermore, the above specifically refers to: For each dataset, an SGPR model was initialized separately. The model was set with characteristic wavelength spectral values ​​as input variables and vitality index values ​​as output variables. An induction point set was introduced to reduce the computational complexity of the model. K-means clustering combined with the elbow rule was used to cluster the characteristic wavelength data of each training set. The relative decrease rate of the sum of squares within each cluster was calculated to determine the optimal number of clusters for the GE / GP / GI model. The sample closest to the cluster center in each cluster was selected as the model induction point. The Matern32 anisotropic kernel function was selected as the kernel function of the SGPR model, and the hyperparameter set was initialized using the following formula:

[0061]

[0062] in, For kernel function, The standard deviation of the signal. The kernel function length scale, These are the weighting coefficients for each feature wavelength dimension. The dimension of the characteristic wavelength. For anisotropic Euclidean distance, Weights for each feature wavelength dimension, , The first , The soybean sample in the first Near-infrared spectral response values ​​at each characteristic wavelength position; The SGPR model is trained using the training set data; the hyperparameter set is optimized by maximizing the marginal likelihood function using the validation set data as a benchmark, and the parameters are iteratively updated until the RMSE prediction error of the validation set reaches the minimum value, thus determining the optimal hyperparameters for each model; the generalization ability of the model is verified using the test set data, and the sparse Gaussian process regression SGPR model is obtained after the accuracy reaches the target.

[0063] Furthermore, the hyperparameter optimization criterion adopts maximizing the marginal likelihood function, as shown in the formula:

[0064] in, Let be the marginal likelihood function. These are the measured values ​​of soybean vitality index. The kernel function matrix between the induced points, It is a diagonal matrix. The square of the noise standard deviation. This is the set of induction points selected using K-means clustering and the elbow rule.

[0065] This embodiment details a soybean vigor detection method integrating UVE feature selection and sparse Gaussian process regression. Fresh soybean seeds are grouped and subjected to gradient aging. Germination potential, germination rate, and germination index are obtained through paper germination experiments. Near-infrared spectra of the samples are then collected and preprocessed using Savitzky-Golay and SNV. The variable stability thresholds for each vigor index are determined based on an improved PLSR model, and the dataset is screened using UVE. The dataset is then divided, and an SGPR model (using Matern32 kernel function, K-means clustering to select induction points, and maximizing the marginal likelihood function to optimize hyperparameters) is trained and optimized on a validation set and validated on a test set. After spectral processing and feature selection of the test samples, the predicted values ​​of the three vigor indices are output using a dedicated SGPR model.

[0066] Example 2, based on Example 1, details the process of using this method with fresh soybean seeds as the detection object, from sample preparation, data acquisition, data processing, model analysis to viability detection, as follows: Sample preparation: The calibration set consisted of 8 groups of soybean seeds aged 0-7 days in advance, with 50 seeds in each group, for a total of 400 seeds (germination test completed, GE, GP, GI measured values ​​obtained for model training and optimization); the test samples consisted of 2 groups of commercially available soybean seeds, with 50 seeds in each group, numbered as Test Group 1 and Test Group 2 (germination test not performed, as unknown samples for testing).

[0067] Data Collection: Soybean seeds of test group 1 and test group 2 were weighed using an electronic balance, and the mass of each group was recorded. The mass of test group 1 was 128.6g and that of test group 2 was 130.2g. Damaged, insect-infested, and deformed seeds were removed from both groups to ensure that 50 whole and uniformly sized soybeans were retained in each group. The screened test samples were placed in a constant temperature drying oven at 30℃ for 2 hours until the seed moisture content stabilized at 10%±0.5%. After drying, the samples were cooled to room temperature. The test samples were then packaged in sealed bags, labeled as test group 1-20250324 and test group 2-20250324, and placed in a desiccator for later use. Turn on the near-infrared spectrometer, warm it up for 30 minutes, and set the core acquisition parameters to be consistent with the calibration set modeling parameters, spectral range. Spectral resolution The scan was performed 16 times per sample in diffuse reflectance mode. Fifty soybeans from test group 1 were evenly spread in the sample cell of the spectrometer, ensuring complete coverage. Scanning was then initiated, and after 16 scans, a single set of raw spectral curves was generated. The average value of the 16 scans was automatically calculated as the raw spectral data for test group 1. The sample cell was wiped clean with a soft cloth, and spectral acquisition was repeated to complete the raw spectral acquisition for test group 2, obtaining the raw spectral data for test group 2.

[0068] Data processing: The original spectral files of the two sets of samples to be tested were read, and a spectral matrix was constructed. The random noise of the spectrum was eliminated by SG convolution with a window size of 21 and a polynomial degree of 2, and a smoothed spectral matrix was obtained. The smoothed spectrum still retains 1557 wavelength points, and the fluctuation of absorbance values ​​is reduced by 32.6% compared with the original spectrum. The smoothed spectral matrix is ​​transformed by standard normal variables to eliminate baseline drift interference caused by sample particle size and packing density, resulting in a preprocessed spectral matrix. The saved calibration set model parameter file is called, and the UVE stability thresholds of the three vitality indices are read: GE threshold 48.54, GP threshold 45.31, and GI threshold 76.4385; the noise variable matrix of the preprocessed spectral matrix calibration set model is combined to calculate the stability value at each wavelength point and generate a stability value vector. like Figure 5 The wavelength distribution map shown is a result of the characteristic wavelength screening. For the GE index, wavelengths with an absolute value of stability ≤ 53.96 were removed, and 86 GE-specific characteristic wavelengths were screened out (concentrated in the protein / amide bond characteristic absorption region of 6000~5500cm⁻¹). For the GP index: wavelengths with an absolute stability value ≤ 69.37 were removed, and 79 GP-specific characteristic wavelengths were selected (concentrated in the 5500~4500cm⁻¹ soybean oil / fatty acid characteristic absorption region). For the GI index: wavelengths with an absolute stability value ≤ 91.14 were removed, and 62 GI-specific characteristic wavelengths were selected (concentrated in the 4500~4000cm⁻¹ soybean starch / carbohydrate characteristic absorption region). Based on the screening results, the absorbance values ​​at corresponding wavelengths were extracted, and the GE characteristic spectral matrices of the samples to be tested were constructed respectively. (2×86) Characteristic spectral matrix (2×79) GI characteristic spectral matrix (2×62).

[0069] Model Analysis: Load the pre-trained and optimized GE / GP / GI-specific SGPR models. The number of induction points is 35 for the GE model, 32 for the GP model, and 28 for the GI model. The maximum values ​​of the marginal likelihood functions of the models are -128.5 (GE), -132.7 (GP), and -119.3 (GI), which are consistent with those when training was completed. The three characteristic spectral matrices of the sample to be tested were input into the corresponding SGPR model. The model was forward propagated to calculate the predicted value of the vigor index. The measured value was recorded simultaneously. The sample was collected 8 times a day, and the predicted data of soybean seed vigor index are shown in Table 1 below: Table 1. Predicted data on soybean seed vigor indicators

[0070] According to Table 1 and Figure 6 As shown in the comparison chart of prediction results for different aging days, the relative error of the model prediction values ​​of this method is controlled within 3% for all samples, and around 1% overall. The predicted values ​​output by the model are in high agreement with the measured values. Furthermore, the data covers the full viability range from 99.00% (fresh) to 29.00% (severely aged), verifying that this method can achieve extremely high accuracy in the detection of each specific sample.

[0071] Experimental data from all stages were collected and compared with the existing GPR model analysis method described in the background section. The overall performance indicators of the soybean seed vigor detection method are shown in Table 2 below: Table 2 Overall Performance Indicators of Soybean Seed Vigor Detection Method

[0072] According to Table 2 and Figure 7 The comparison chart of overall performance indicators shows that the traditional GPR model directly inputs full-band spectral data, introducing a large amount of redundant noise unrelated to seed vigor; while the present invention, through UVE feature selection, eliminates non-informative variables and retains only the core feature wavelengths, achieving a comprehensive RMSE (1.03) and Both methods significantly outperform the GPR model analysis method; furthermore, this invention reduces the complexity to O(n) through the induced point strategy. Since m is much smaller than n, the computational complexity is greatly reduced compared to the O(n) of the GPR model analysis method. It has a significant advantage in terms of time consumption for a single sample detection and achieves high-precision prediction with full range coverage.

[0073] This embodiment describes in detail the process of using this method with fresh soybean seeds as the detection object, from sample preparation, data acquisition, data processing, model analysis to viability detection. By combining UVE feature selection and SGPR, it surpasses the existing GPR model analysis method in terms of accuracy, efficiency, and robustness. Moreover, all indicators are far superior to the preset industrial thresholds, demonstrating strong industrial application value.

[0074] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A soybean vigor detection method integrating UVE feature selection and sparse Gaussian process regression, characterized in that, include: Fresh soybean seeds were grouped and aged within a preset period to obtain a grouped vigor gradient sample. Paper germination test was performed on all samples, and the germination potential (GE), germination rate (GP), and germination index (GI) were calculated to obtain soybean vigor data. Soybean near-infrared spectral data were obtained by scanning each sample using a near-infrared spectrometer and then preprocessing it. Based on germination potential (GE), germination rate (GP), and germination index (GI), the stability threshold of the variables corresponding to each index is determined by an improved partial least squares regression (PLSR) model. The improved PLSR model is optimized by leave-one-out cross-validation. Based on variable stability thresholds and soybean near-infrared spectral data, non-informative variables with wavelength values ​​below the corresponding thresholds were removed, resulting in a dataset highly correlated with each vitality index. The dataset is divided into a training set, a validation set, and a test set. A sparse Gaussian process regression (SGPR) model is trained using the training set and the model parameters are optimized using the validation set. Based on the sparse Gaussian process regression (SGPR) model, the soybean vigor detection results of the soybean seeds to be tested are output, including the germination potential, germination rate, and predicted germination index of the soybean seeds.

2. The soybean vigor detection method integrating UVE feature selection and sparse Gaussian process regression according to claim 1, characterized in that, The soybean vitality data obtained specifically includes: a preset aging period of... In the germination test Germination potential (GE) of each group of samples was counted on day 2. The germination rate (GP) of each group of samples was statistically analyzed daily. The germination index (GI) is calculated based on the number of germinated seeds and the number of germination days. The formula is as follows: ； in, For soybean seed samples in the germination test The number of seeds that germinate in a day This refers to the number of days since germination. The number of days for the germination test. The statistical period for the germination experiment; By summarizing the values ​​of germination potential, germination rate, and germination index, complete soybean vitality data can be obtained.

3. The soybean vigor detection method integrating UVE feature selection and sparse Gaussian process regression according to claim 1, characterized in that, The soybean near-infrared spectral data are specifically obtained by performing a single spectral scan on each soybean seed sample, completing the first round of scanning of all samples according to preset parameters, repeating the scanning operation, and averaging the multiple scan data of each sample to obtain the original near-infrared spectral data of each sample. Savitzky-Golay convolution smoothing combined with standard normal variable transformation was used to process the raw near-infrared spectral data to eliminate interference factors such as baseline drift, random noise, and scattering effects, correct spectral data deviation, and obtain smooth and accurate soybean near-infrared spectral data.

4. The soybean vigor detection method integrating UVE feature selection and sparse Gaussian process regression according to claim 1, characterized in that, The determination of the variable stability thresholds for each indicator is specifically as follows: the measured data and near-infrared spectral data of the three vitality indicators are matched and divided to obtain three independent modeling datasets; based on the three datasets, partial least squares regression (PLSR) models are constructed respectively; leave-one-out cross-validation is used to optimize each PLSR model; based on the three PLSR models optimized by leave-one-out cross-validation, random noise variables are introduced and their stability values ​​are calculated, and the variable stability thresholds are obtained based on the maximum stability value.

5. The soybean vigor detection method based on the fusion of UVE feature selection and sparse Gaussian process regression according to claim 4, characterized in that, The partial least squares regression (PLSR) model is constructed by using preprocessed near-infrared spectral data as independent variables and matching them with measured data of three vigor indicators: germination potential (GE), germination rate (GP), and germination index (GI), to obtain three independent modeling datasets: GE-spectrum, GP-spectrum, and GI-spectrum. Based on three datasets, partial least squares regression (PLSR) models were constructed to extract potential correlation factors between spectral data and various vitality indicators. A linear regression relationship between spectral characteristics and soybean vitality indicators was established, as shown in the formula: ； in, This is a predicted value for the vitality index. For the intercept term, For the first The regression coefficients of the wavelength variable, For the preprocessed first Spectral values ​​at each wavelength point, This represents the total number of spectral wavelength variables. Fit the residuals to the model; The optimal number of latent variables for each PLSR model is determined based on minimizing the root mean square error of cross-validation (RMSECV), using the following formula: ； in, To model the total number of samples, For the first Measured values ​​of the vitality index for each sample. For the first Cross-validation predicted values ​​of the vitality index for each sample.

6. The soybean vigor detection method based on the fusion of UVE feature selection and sparse Gaussian process regression according to claim 4, characterized in that, The variable stability threshold based on the maximum stability value is obtained by: optimizing each PLSR model by leave-one-out cross-validation, using one sample in each dataset as the validation set and the remaining samples as the training set, repeatedly training and validating the model until all samples are traversed, and determining the optimal number of latent variables for each PLSR model. Based on the three PLSR models optimized using leave-one-out cross-validation, random noise variables were introduced into each model, and their stability values ​​were calculated. The maximum absolute value of the stability value of the noise component was extracted and set as the variable stability threshold for the corresponding vitality index. The variable stability thresholds for the three indices GE, GP, and GI were obtained using the following formulas: ； ； in, For the first The stability threshold of a vitality indicator variable. The number of iterations for leave-one-out cross-validation for the PLSR model. For the first In the cross-validation, the projected importance value of the random noise variable. The original number of wavelength variables. For the first The variance explained by each latent variable to the dependent variable For the i-th cross-validation, the i-th Weighting coefficients of noise variables under each latent variable. For the first The magnitude of the weight coefficients of all variables under one latent variable. This represents the optimal number of latent variables in the PLSR model.

7. The soybean vigor detection method integrating UVE feature selection and sparse Gaussian process regression according to claim 1, characterized in that, The dataset specifically comprises: based on the stability thresholds of variables such as germination potential (GE), germination rate (GP), and germination index (GI), combined with preprocessed soybean near-infrared spectral data, wavelengths with no or low information are filtered out through targeted wavelength variable screening. For each vigor indicator, its specific variable stability threshold is used as the criterion to screen the corresponding wavelength point-stability value dataset, removing wavelength variables with stability values ​​lower than the corresponding threshold and retaining wavelength variables with stability values ​​higher than the corresponding threshold. Highly correlated feature wavelengths are extracted for each vigor indicator to form the corresponding dataset.

8. The soybean vigor detection method based on the fusion of UVE feature selection and sparse Gaussian process regression according to claim 1, characterized in that, The Sparse Gaussian Process Regression (SGPR) model specifically involves dividing the dataset into three groups: GE, GP, and GI, and further dividing the dataset to obtain a training set, a validation set, and a test set. The SGPR model was trained using the training set data, and the Gaussian process posterior distribution was constructed using the fully independent training condition FITC approximation. The hyperparameter set was optimized by maximizing the marginal likelihood function based on the validation set data. The generalization ability of the model was verified using the test set data. Data collection and feature processing were performed on the soybean seeds to be tested to obtain a dataset of soybean seeds to be tested. The dataset was then input into the GE / GP / GI dedicated SGPR model that had been trained and optimized, and the predicted values ​​of germination potential, germination rate and germination index of the test samples were output.

9. The soybean vigor detection method based on the fusion of UVE feature selection and sparse Gaussian process regression according to claim 8, characterized in that, The specific method for partitioning the dataset is to simultaneously minimize the Euclidean distance between the spectral and vitality indices, as shown in the formula: ； in, For the sample With sample The overall distance , Samples , The characteristic wavelength spectral value vector, , Samples , The vitality index value, This is for Euclidean distance calculation. To balance the distance weighting of spectral and vitality indicators; The training set, validation set, and test set are divided according to a preset ratio, and samples that are far apart and scattered are assigned to different subsets.

10. The soybean vigor detection method based on the fusion of UVE feature selection and sparse Gaussian process regression according to claim 8, characterized in that, The training of the SGPR model is specifically as follows: For each dataset, an SGPR model was initialized separately. The model was set with characteristic wavelength spectral values ​​as input variables and vitality index values ​​as output variables. An induction point set was introduced to reduce the computational complexity of the model. K-means clustering combined with the elbow rule was used to cluster the characteristic wavelength data of each training set. The relative decrease rate of the sum of squares within each cluster was calculated to determine the optimal number of clusters for the GE / GP / GI model. The sample closest to the cluster center in each cluster was selected as the model induction point. The Matern32 anisotropic kernel function was selected as the kernel function of the SGPR model, and the hyperparameter set was initialized using the following formula: ； ； in, For kernel function, The standard deviation of the signal. The kernel function length scale, These are the weighting coefficients for each feature wavelength dimension. The dimension of the characteristic wavelength. For anisotropic Euclidean distance, Weights for each feature wavelength dimension, , The first , The soybean sample in the first Near-infrared spectral response values ​​at each characteristic wavelength position; The SGPR model is trained using the training set data; the hyperparameter set is optimized by maximizing the marginal likelihood function using the validation set data as a benchmark, and the parameters are iteratively updated until the RMSE prediction error of the validation set reaches the minimum value, thus determining the optimal hyperparameters for each model; the generalization ability of the model is verified using the test set data, and the sparse Gaussian process regression SGPR model is obtained after the accuracy reaches the target.

Images