A near-infrared spectroscopy prediction method for soybean seed vitality based on machine learning and deep learning algorithms

Abstract

The present invention relates to the field of agricultural planting technology, and in particular to a soybean seed vitality near infrared spectrum prediction method based on machine learning and deep learning algorithms, comprising the following steps: S1, bean selection; S2, preparing soybean seed samples with different vitality levels by artificial aging method; S3, obtaining near infrared spectra of soybean seed samples; S4, obtaining the true value of seed vitality index of soybean seed samples; S5, abnormal spectrum removal; S6, spectrum preprocessing; S7, characteristic band extraction; S8, model construction and evaluation of the model; S9, obtaining a prediction model for soybean seed vitality. The present invention can not only predict the germination rate of soybean seeds with light and dark seed coats, but also predict the germination index, normal seedling rate, germination potential and bud length, and can more comprehensively predict the vitality of soybean seeds.

Description

Technical Field

[0001] The present invention relates to the field of agricultural planting technology, and in particular to a soybean seed vitality near-infrared spectroscopy prediction method based on machine learning and deep learning algorithms. Background Art

[0002] Mechanized sowing is a strong guarantee for the large-scale promotion of the soybean-corn strip composite planting model, but mechanized sowing puts higher requirements on the germination and emergence speed, uniformity, and stress resistance of soybean seeds.

[0003] Seed vigor is defined as: "a general term for the potential ability of seeds to quickly and uniformly germinate and grow into normal seedlings under a wide range of field conditions". Seed vigor will directly affect the emergence rate of seeds in the field and their ability to resist adverse environments. It is an important indicator of seed quality and an important factor affecting the high and stable yield of crops. Predicting the vitality of soybean seeds with high accuracy before sowing can test seed quality, reduce the occurrence of soybean seed breakage, seedling shortage and uneven emergence, achieve more ideal field planting effects, and effectively reduce the amount of field sowing, saving sowing costs and sowing costs such as manpower and material resources. It is an indispensable part of agricultural production. Therefore, the development of soybean seed vigor prediction technology has important practical significance.

[0004] The determination methods of seed vigor can be divided into direct determination method and indirect determination method. The traditional evaluation of seed vigor is carried out by direct determination using the sand culture method, which requires the calculation of seed vigor germination index, normal seedling rate, germination potential, germination rate and seedling length after germination experiments. This method has the advantages of intuitive display and accurate measurement, but it also has disadvantages such as cumbersome operation, complex and time-consuming prediction process, and easy damage to seeds. It cannot meet the needs of modern agriculture for large-scale, rapid and non-destructive seed prediction and future development trends.

[0005] Near infrared spectroscopy (NIRS) is a modern technology that uses near infrared light to efficiently analyze substances. At present, existing technologies use near infrared technology to determine the seed viability of plant seeds and predict seed vitality, but the soybean seed near infrared spectroscopy prediction models constructed by most research institutes are based on a single variety and yellow seed coat soybeans. The modeling methods used are mostly linear models such as partial least squares. Such models have problems such as single detection traits, low prediction accuracy, and difficulty in applying to the detection of dark seed coat soybean seed vitality. They have poor versatility and universality and have obvious limitations. Summary of the invention

[0006] Features and advantages of the invention are set forth in part in the following description, or may be obvious from the description, or may be learned through practice of the invention.

[0007] In order to overcome the problems of the prior art, the present invention provides a soybean seed vitality near infrared spectrum prediction method based on machine learning and deep learning algorithms, comprising the following steps:

[0008] S1. Bean selection: select N portions of soybeans that are free of insects, mildew, damage, and impurities, where N≥1, and select most of them as a training set and the rest as a test set; the soybeans include two groups: light-colored seed coat soybeans and dark-colored seed coat soybeans, wherein the light-colored seed coat soybeans include yellow-colored seed coat soybeans and green-colored seed coat soybeans, and the dark-colored seed coat soybeans include brown-colored seed coat soybeans and black-colored seed coat soybeans;

[0009] S2. Prepare soybean seed samples with different levels of vitality using artificial aging method;

[0010] S3, obtaining a near infrared spectrum of the soybean seed sample: placing the soybean seed sample prepared in S2 into a near infrared spectrometer to obtain a near infrared spectrum of the soybean seed;

[0011] S4. Obtaining the true value of the seed vigor index of the soybean seed sample: conducting a germination experiment using a sand culture method, and calculating the true value of the seed vigor index after the germination is completed; the seed vigor index includes germination index, normal seedling rate, germination potential, germination rate and seedling sprout length;

[0012] S5, abnormal spectrum removal: removing abnormal spectrum data in the near infrared spectrum data collected in step S3;

[0013] S6, spectrum preprocessing: the spectrum processed in step S5 was preprocessed as follows: first-order derivative, second-order derivative, convolution smoothing, and baseline correction;

[0014] S7, extraction of characteristic bands: performing Pearson correlation analysis on the spectrum preprocessed in step S6 to extract effective characteristic bands;

[0015] S8, constructing a model and evaluating the model: using the spectral data obtained after extraction in step S7 as an input value, and the true value of the seed vigor index obtained in step S4 as an output value, carrying out the model construction with XGBoost and BP neural network algorithms, and using the training set data to respectively construct light-colored and dark-colored seed coat soybean seed vigor prediction models; evaluating the model by comparing the difference between the true value and the predicted value of the seed vigor index, and at the same time comparing and analyzing the determination coefficient and root mean square error of the training set and the test set to evaluate the model;

[0016] S9, obtaining a prediction model for soybean seed vitality: obtaining a prediction model for the vitality of light-colored and dark-colored seed coat soybean seeds according to the evaluation result of step S8.

[0017] Preferably, the prediction model includes:

[0018] The prediction model for GI, GR, and SL of light-colored seed coat soybeans was a first-order derivative preprocessing combined with the XGBoost model;

[0019] The prediction model for predicting GE and NSR of light-colored seed coat soybeans was the second-order derivative preprocessing and XGBoost model;

[0020] The prediction model for predicting GI and NSR of dark-coated soybeans is a baseline correction preprocessing combined with a BP neural network model;

[0021] The prediction model for predicting GE of dark-coated soybeans is a baseline correction preprocessing combined with an XGBoost model;

[0022] The prediction model for predicting GR of dark-coated soybeans is a convolution smoothing preprocessing combined with an XGBoost model;

[0023] The prediction model for predicting SL of dark-seed soybean is the first-order derivative combined with the XGBoost model.

[0024] Preferably, the proportion of soybeans in the training set is between 2 / 3 and 3 / 4.

[0025] Preferably, the number of varieties of light-colored seed coat soybeans is ≥25; the number of dark-colored seed coat soybean varieties is ≥15.

[0026] Preferably, the number of grades in the soybean seed samples with different grades of vitality levels prepared by the artificial aging method is ≥4.

[0027] The beneficial effects of the present invention: the prediction method provided by the present invention is used to predict the seed vigor of soybeans with different seed coat colors, which can not only predict the germination rate of soybean seeds, but also predict the germination index, normal seedling rate, germination potential and sprout length. Compared with the seed vigor prediction model previously constructed only based on the germination rate, it can more comprehensively and accurately evaluate the vitality of soybean seeds. Compared with the seed vigor prediction model constructed only based on a single soybean variety by other technologies, the prediction model constructed by the present invention has higher universality and versatility in production applications. In addition, it can also predict the seed vigor of dark seed coat soybeans and has a wider range of applications. The present application has good stability and accuracy on the training set and the test set, and has a high generalization ability. It can be used to predict the germination rate, germination potential, germination index, normal seedling rate and seedling length of soybeans with a variety of seed coat colors, and can more comprehensively and accurately evaluate the vitality of soybean seeds. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The present invention will be described in detail below with reference to the accompanying drawings and in combination with examples, and the advantages and implementation methods of the present invention will become more obvious. The contents shown in the accompanying drawings are only used to explain the present invention and do not constitute any limitation to the present invention in any sense. In the accompanying drawings:

[0029] Figure 1 It is a flowchart of a soybean seed vitality near-infrared spectroscopy prediction method based on machine learning and deep learning algorithms in a specific embodiment of the present invention;

[0030] Figure 2 It is a principal component analysis diagram of spectral data of light-colored seed coat soybean samples in a soybean seed vitality near-infrared spectroscopy prediction method based on machine learning and deep learning algorithms in a specific embodiment of the present invention;

[0031] Figure 3 It is a principal component analysis diagram of spectral data of dark seed coat soybean samples in a soybean seed vitality near-infrared spectroscopy prediction method based on machine learning and deep learning algorithms in a specific embodiment of the present invention;

[0032] Figure 4 The original spectrum and pre-processed spectrum of a light-colored seed coat soybean sample of a soybean seed vitality near-infrared spectroscopy prediction method based on machine learning and deep learning algorithms in a specific embodiment of the present invention: wherein, A: original spectrum (NO); B: convolution smoothing (SG) pre-processed spectrum; C: standard normal variable (SNV) pre-processed spectrum; D: multivariate scattering correction (MSC) pre-processed spectrum; E: baseline correction (LBC) pre-processed spectrum; F: maximum normalization (MN) pre-processed spectrum; G: first-order derivative (1D) pre-processed spectrum; H: second-order derivative (2D) pre-processed spectrum;

[0033] Figure 5 The present invention provides an original spectrum and preprocessed spectrum of a dark seed coat soybean sample of a soybean seed vitality near-infrared spectroscopy prediction method based on machine learning and deep learning algorithms in a specific embodiment of the present invention: wherein, A: original spectrum (NO); B: convolution smoothing (SG) preprocessed spectrum; C: standard normal variable (SNV) preprocessed spectrum; D: multivariate scattering correction (MSC) preprocessed spectrum; E: baseline correction (LBC) preprocessed spectrum; F: maximum normalization (MN) preprocessed spectrum; G: first-order derivative (1D) preprocessed spectrum; H: second-order derivative (2D) preprocessed spectrum. DETAILED DESCRIPTION

[0034] The following embodiments of the present invention are described in further detail in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0035] like Figure 1 As shown, the present invention provides a soybean seed vitality near infrared spectrum prediction method based on machine learning and deep learning algorithms, which specifically includes the following steps:

[0036] S1. Bean selection: select N portions of soybeans that are free of insect infestation, mildew, damage, and impurities, N≥1, and select most of them as the training set, and the remaining part as the test set; the soybeans include two groups of light-colored seed coat soybeans and dark-colored seed coat soybeans, wherein light-colored seed coat soybeans include yellow seed coat soybeans and green seed coat soybeans, and dark-colored seed coat soybeans include brown seed coat soybeans and black seed coat soybeans; the proportion of soybeans selected in the training set should be between 2 / 3 and 3 / 4; the number of light-colored soybean varieties should be ≥25; the number of dark-colored soybean varieties should be ≥15. In this embodiment, 43 varieties of soybeans with different origins, colors, and appearances were sampled, light-colored soybeans were 27 commercially approved varieties, and dark-colored soybeans were 16 local varieties collected in Yunnan, and the specific soybean varieties are shown in Table 1:

[0037] Table 1 Soybean material information

[0038]

[0039] S2. Prepare soybean seed samples with different vigor levels by artificial aging method: Take soybean varieties of Table 1, prepare seeds with different vigor levels by artificial aging treatment of different lengths, and obtain soybean samples with different seed vigor levels of at least 4 levels. In this embodiment, the aging method is hot water treatment, and 5 soybean seed samples with different vigor levels are obtained. The treatment conditions are soaking for 0 min, 4 min, 6 min, 8 min and 12 min at 58±1℃, respectively, and drying to the original weight in a 37℃ oven after treatment. Some varieties cannot tolerate hot water treatment for too long. After removing incomplete seeds such as broken and cracked seeds in hot water treatment, 108 light-colored soybean samples with different vigor levels and 67 dark-colored soybean samples with different vigor levels are obtained and preserved for testing.

[0040] S3. Obtaining a near-infrared spectrum of a soybean seed sample: placing the soybean seed sample prepared in S2 into a near-infrared spectrometer to obtain a near-infrared spectrum of the soybean seed: spreading the seed sample evenly over the entire sample tray, flattening the surface to ensure that the sample tray is completely covered, placing it in a near-infrared spectrometer to obtain a near-infrared spectrum of the soybean seed sample;

[0041] S4. Obtain the true values ​​of the five seed vitality indicators: All soybean seeds after the test in step S3 were subjected to a germination experiment using the sand culture method in accordance with the national standard GB / T3543.4-1995 "Germination Experiment of Crop Seed Inspection Procedures", and cultured for 8 days under 12h light and 12h dark conditions at 25°C, recording germination data every 24h, and measuring the length of seedlings on the 8th day. When the cotyledons are ≥1cm above the sand surface, it is judged to be germinated, and when the cotyledons are intact and the primary leaves are fully expanded, it is judged to be a normal seedling. After germination, the true values ​​of the five indicators of seed vitality germination index (English abbreviation: GI), normal seedling rate (English abbreviation: NSR), germination potential (English abbreviation: GE), germination rate (English abbreviation: GR) and seedling length (English abbreviation: SL) were calculated respectively; the calculation formulas for the relevant indicators are shown in Table 2:

[0042] Table 2 Calculation formulas for seed vitality related indicators

[0043]

[0044] Note: Gt represents the number of newly germinated seeds on day t; Dt represents the number of days corresponding to Gt.

[0045] S5. Abnormal sample removal: PCA screening was performed on the near infrared spectral data of 324 light-colored grain samples collected in S3, such as Figure 2 As shown in Figure 1, the cumulative contribution rate of PC1 and PC2 reached 85+14=99%, and the 21 samples outside the confidence ellipse 1 were eliminated; the near-infrared spectral data of 201 dark grain samples collected by S3 were screened and eliminated by PCA, as shown in Figure 1. Figure 3 As shown, the cumulative contribution rate of PC1 and PC2 reaches 84+15=99%, and the 12 samples located outside the confidence ellipse 1 are eliminated.

[0046] S6. Spectral preprocessing: After removing abnormal spectra, the original near-infrared spectra of light-colored soybean seed coat and dark-colored soybean seed coat in S5 are as follows: Figure 4 A and Figure 5 As shown in A; the spectra after the abnormal spectrum was removed were preprocessed as follows: first-order derivative (1D), second-order derivative (2D), convolution smoothing (SG), standard normal variable (SNV), multivariate scatter correction (MSC), maximum normalization (MN) and baseline correction (LBC). The near-infrared spectra after preprocessing are shown in Figure 4 and 5 As shown;

[0047] S7. Extraction of characteristic bands: Perform Pearson correlation analysis on the spectra preprocessed in S6 to extract effective characteristic bands. The characteristic bands of the spectrum of light-colored seed coat soybean samples are shown in Table 3:

[0048] Table 3 Spectral characteristic bands of light-colored seed coat soybean samples

[0049]

[0050] The maximum normalization (MN) and standard normal variable (SNV) preprocessing methods did not extract characteristic bands with high correlation; the spectral characteristic bands of dark seed coat soybean samples are shown in Table 4:

[0051] Table 4 Spectral characteristic bands of dark-coated soybean samples

[0052]

[0053] S8. Sample set division: The true values ​​of the five seed vitality-related indicators obtained in step S4 are randomly divided into training set and test set in a ratio of 7:3. Among them, there are 212 light-colored soybean samples in the training set and 91 in the test set. The descriptive statistics are shown in Table 5:

[0054] Table 5 Descriptive statistics of light-colored soybean sample set

[0055]

[0056] There are 132 dark soybean samples in the training set and 57 in the test set. The descriptive statistics are shown in Table 6:

[0057] Table 6 Descriptive statistics of dark soybean sample set

[0058]

[0059] The data range of the five seed vitality indicators in the training set and the test set is relatively wide, which can better cover the range of soybean seed vitality and is highly representative, and can be used for model construction.

[0060] S9. Construct a model and evaluate the model: Use the spectral data extracted in step S7 as the input value, and the true values ​​of the five seed vitality-related indicators obtained in step S4 as the output value, respectively use multiple linear regression (MLR), partial least squares regression (PLSR), support vector machine regression (SVR), XGBoost and BP neural network algorithm (BPNN) to construct the model, and use the training set data to construct light-colored and dark-colored seed coat soybean seed vitality prediction models respectively; then, use the constructed model to predict the test set data.

[0061] The model was evaluated by comparing the differences between the true and predicted values ​​of the five seed vigor-related indicators. The R 2 (coefficient of determination) and RMSE (root mean square error) were used to evaluate the model, where R 2The value close to 1 indicates that the model has good accuracy, and the value close to 0 indicates that the model has low error and good stability. The model evaluation results of different pretreatment and modeling algorithm combinations for predicting five seed vitality indicators of light-colored seed coat soybeans are shown in Table 7:

[0062] Table 7 Evaluation results of different combinations of pretreatment and modeling algorithms for predicting seed vigor of light-colored seed coat soybeans

[0063]

[0064]

[0065] Note: NO: original spectrum; SG: convolution smoothing; SNV: standard normal variate; LBC: baseline correction; 1D: first-order derivative; 2D: second-order derivative; MLR: multivariate linear regression; PLSR: partial least squares regression; SVR: support vector machine regression; BPNN: BP neural network algorithm.

[0066] The model evaluation results of different pretreatment and modeling algorithm combinations for predicting five seed vigor indices of dark seed coat soybean are shown in Table 8:

[0067] Table 8 Evaluation results of different combinations of pretreatment and modeling algorithms for predicting seed vigor of dark-coated soybeans

[0068]

[0069]

[0070]

[0071] Note: NO: original spectrum; SG: convolution smoothing; SNV: standard normal variate; MSC: multivariate scatter correction; LBC: baseline correction; MN: maximum normalization; 1D: first-order derivative; 2D: second-order derivative; MLR: multivariate linear regression; PLSR: partial least squares regression; SVR: support vector machine regression; BPNN: BP neural network algorithm.

[0072] S10. Obtaining a prediction model for soybean seed vigor: According to the evaluation results of the model in step S9, the optimal combination of seed vigor prediction models for light-colored seed coat soybeans and dark-colored seed coat soybeans is obtained, as shown in Table 9:

[0073] Table 9 Optimal combinations of seed vigor prediction models for light-colored seed coat soybeans and dark-colored seed coat soybeans

[0074]

[0075] The optimal model for predicting GI, GR, and SL of light-colored seed coat soybeans was the first-order derivative preprocessing combined with the XGBoost model. 2 The RMSE of the training set is 0.5, 5.93 and 1.14 respectively, and the R 2 The results were 0.86, 0.81 and 0.85 respectively, and the RMSE of the test set were 0.65, 10.10 and 1.88 respectively. The optimal model for predicting GE and NSR of light-colored seed coat soybean was the second-order derivative pretreatment and XGBoost model. The R 2 The RMSE of the training set is 6.79 and 3.17 respectively, and the R 2 They reach 0.91 and 0.85 respectively, and the test set RMSE is 7.82 and 9.99 respectively.

[0076] The optimal model for predicting the GI of dark-coated soybeans is the baseline correction preprocessing combined with the BP neural network model. 2 and RMSE were 0.96 and 0.04 respectively, and the R 2 and RMSE were 0.93 and 0.48 respectively; the optimal prediction model for predicting GE of dark-coated soybean was the baseline correction preprocessing combined with the XGBoost model, and the R 2 and RMSE were 0.97 and 4.49 respectively, and the R 2 and RMSE were 0.90 and 8.45 respectively; the optimal model for predicting GR of dark-coated soybean was convolution smoothing preprocessing combined with XGBoost model, and the R 2 and RMSE were 0.94 and 6.67 respectively, and the R 2 and RMSE were 0.82 and 8.95 respectively; the optimal prediction model for predicting NSR of dark-seed soybean was baseline correction preprocessing combined with BP neural network model, and the R 2 and RMSE were 0.96 and 3.30 respectively, and the R 2 and RMSE were 0.79 and 8.84 respectively; the optimal prediction model for predicting SL of dark-coated soybean was the first-order derivative combined with the XGBoost model, and the R 2 and RMSE were 0.98 and 0.71 respectively, and the R 2 and RMSE are 0.95 and 0.97 respectively.

[0077] The above preprocessing methods and model methods performed best on the training set and test set, with high accuracy and stability. A near-infrared prediction model of soybean seed vitality with strong versatility and high generalization ability has been obtained, which can predict soybean germination rate, germination potential, germination index, normal seedling rate and seedling length. This model can predict soybean seed vitality more comprehensively.

[0078] The preferred embodiments of the present invention are described above with reference to the accompanying drawings. A person skilled in the art may implement the present invention in a variety of variations without departing from the scope and essence of the present invention. For example, a feature shown or described as part of one embodiment may be used in another embodiment to obtain another embodiment. The above are only preferred feasible embodiments of the present invention, and do not limit the scope of rights of the present invention. Any equivalent changes made using the contents of the present specification and the accompanying drawings are included in the scope of rights of the present invention.

Claims

1. A soybean seed vitality near infrared spectroscopy prediction method based on machine learning and deep learning algorithms, characterized in that: The following steps are involved: S1. Bean selection: select N portions of soybeans that are free of insects, mildew, damage, and impurities, where N≥1, and select most of them as a training set and the rest as a test set; the soybeans include two groups: light-colored seed coat soybeans and dark-colored seed coat soybeans, wherein the light-colored seed coat soybeans include yellow-colored seed coat soybeans and green-colored seed coat soybeans, and the dark-colored seed coat soybeans include brown-colored seed coat soybeans and black-colored seed coat soybeans; S2. Prepare soybean seed samples with different levels of vitality using artificial aging method; S3, obtaining a near infrared spectrum of the soybean seed sample: placing the soybean seed sample prepared in S2 into a near infrared spectrometer to obtain a near infrared spectrum of the soybean seed; S4. Obtaining the true value of the seed vigor index of the soybean seed sample: conducting a germination experiment using a sand culture method, and calculating the true value of the seed vigor index after the germination is completed; the seed vigor index includes germination index, normal seedling rate, germination potential, germination rate and seedling sprout length; S5, abnormal spectrum removal: removing abnormal spectrum data in the near infrared spectrum data collected in step S3; S6, spectrum preprocessing: the spectrum processed in step S5 was preprocessed as follows: first-order derivative, second-order derivative, convolution smoothing, and baseline correction; S7, extraction of characteristic bands: performing Pearson correlation analysis on the spectrum preprocessed in step S6 to extract effective characteristic bands; S8, constructing a model and evaluating the model: using the spectral data obtained after extraction in step S7 as an input value, and the true value of the seed vigor index obtained in step S4 as an output value, carrying out the model construction with XGBoost and BP neural network algorithms, and using the training set data to respectively construct light-colored and dark-colored seed coat soybean seed vigor prediction models; evaluating the model by comparing the difference between the true value and the predicted value of the seed vigor index, and at the same time comparing and analyzing the determination coefficient and root mean square error of the training set and the test set to evaluate the model; S9, obtaining a prediction model for soybean seed vitality: obtaining a prediction model for the vitality of light-colored and dark-colored seed coat soybean seeds according to the evaluation results of step S8, wherein the prediction model includes: The prediction model for GI, GR, and SL of light-colored seed coat soybeans was a first-order derivative preprocessing combined with the XGBoost model; The prediction model for predicting GE and NSR of light-colored seed coat soybeans was the second-order derivative preprocessing and XGBoost model; The prediction model for predicting GI and NSR of dark-coated soybeans is a baseline correction preprocessing combined with a BP neural network model; The prediction model for predicting GE of dark-coated soybeans is a baseline correction preprocessing combined with an XGBoost model; The prediction model for predicting GR of dark-coated soybeans is a convolution smoothing preprocessing combined with an XGBoost model; The prediction model for predicting SL of dark-seed soybean is the first-order derivative combined with the XGBoost model.

2. The soybean seed vigor near infrared spectroscopy prediction method based on machine learning and deep learning algorithm according to claim 1, characterized in that: The proportion of soybeans in the training set is between 2 / 3 and 3 / 4.

3. The soybean seed vigor near infrared spectroscopy prediction method based on machine learning and deep learning algorithm according to claim 1, characterized in that: The number of varieties of soybeans with light-colored seed coats shall be ≥25; the number of varieties of soybeans with dark-colored seed coats shall be ≥15.

4. The soybean seed vigor near infrared spectroscopy prediction method based on machine learning and deep learning algorithm according to claim 1, characterized in that: The number of grades in the soybean seed samples with different grades of vitality levels prepared by the artificial aging method is ≥4.

Images