A method for detecting germination capacity of soybean seeds based on diffuse transmittance near infrared spectroscopy technology

Abstract

The present application relates to the technical field of agricultural detection, and particularly relates to a method for detecting germination capacity of soybean seeds based on diffuse transmittance near-infrared spectroscopy technology. The method comprises the following steps: performing diffuse transmittance near-infrared light scanning on seeds of soybean to be detected to obtain near-infrared spectroscopy results; introducing the near-infrared spectroscopy results into a germination capacity detection model to obtain the germination capacity of the soybean to be detected; the germination capacity detection model is constructed based on near-infrared transmittance spectroscopy and germination data of specific positions of multiple natural aging soybean seeds, and in the training process, a standard normal variable transformation preprocessing step is adopted for the spectroscopy data, a random forest model is used for training, a grid search algorithm is used for optimization, and 5-fold cross-validation and a test set are used for evaluation. The present application trains a germination capacity detection model through specific samples and processing methods, and the detection method based on the model can realize nondestructive, efficient and accurate detection of soybean seeds, and has important application value.

Description

Technical Field

[0001] This invention relates to the field of agricultural testing technology, and in particular to a method for detecting the germination ability of soybean seeds based on diffuse transmission near-infrared spectroscopy. Background Technology

[0002] Seed germination rate is a core indicator for evaluating seed quality and predicting yield. Current methods for obtaining or predicting germination rate mainly include: 1. Traditional biological detection: calculating germination rate through germination experiments under standard conditions, or using reagents such as TTC for vigor testing. 2. Near-infrared spectroscopy (NIR) prediction: classifying seeds into high / low quality based on the spectra of multiple seeds, or establishing models based on the reflectance spectra of single seeds to distinguish between artificially aged and non-artificially aged seeds. 3. Environmental parameter prediction: using machine learning to predict germination rate loss trends based on environmental temperature, humidity, and packaging conditions.

[0003] However, existing technologies have the following significant shortcomings in practical applications: 1. Destructive testing: After germination experiments and viability tests, seeds are consumed or destroyed and cannot be preserved or utilized. Therefore, the measurement results can only represent the statistical level of the sampled sample and cannot be directly equated with the true quality of the remaining seeds in the same batch. The validity of the results is greatly affected by the sampling amount and statistical error. 2. Limitations of spectral prediction models: Models based on multiple seeds are difficult to accurately quantify germination rates and are prone to overfitting; models based on single seeds often use reflectance scanning, which is structurally inaccurate. In addition, existing modeling data mostly comes from artificially aged seeds and cannot cover all the real factors of seed deterioration under natural aging conditions, resulting in insufficient prediction accuracy in practical applications. 3. Low accuracy of environmental prediction: Methods based on environmental parameters are greatly affected by seed type and can usually only predict the deterioration trend, making it difficult to achieve accurate quantitative prediction.

[0004] In summary, existing technologies either fail to achieve continuous tracking and detection of individual seeds due to their destructive nature, or suffer from insufficient prediction accuracy and universality due to incomplete detection information and flawed model construction methods. Therefore, there is an urgent need for a technical solution that can non-destructively, accurately, and continuously monitor and predict the germination rate of soybean seeds. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a method for detecting the germination ability of soybean seeds based on diffuse transmission near-infrared spectroscopy.

[0006] In a first aspect, the present invention provides a method for detecting the germination ability of soybeans, comprising: performing diffuse transmission near-infrared light scanning on the seeds of the soybean to be tested to obtain near-infrared spectral results; The near-infrared spectroscopy results are imported into the germination ability detection model to obtain the germination ability of the soybean to be tested; The germination ability detection model is constructed based on the near-infrared transmission spectrum and germination data of multiple naturally aged soybean seeds. During the training process, the spectral data is preprocessed using a standard normal variable transformation, trained using a random forest model, optimized using a grid search algorithm, and evaluated using 5-fold cross-validation.

[0007] The germination ability of soybean seeds is closely related to the content of their internal biochemical components (such as moisture, protein, oil, and active enzymes) and their microscopic and macroscopic physical structure. These intrinsic properties exhibit characteristic spectral responses in spectral data. This invention employs a targeted spectral acquisition method tailored to the characteristics of naturally aged soybean seeds, improving the correlation between the spectrum and germination. Combined with appropriate data preprocessing and feature reduction methods, redundant information in high-dimensional spectral data is eliminated. Finally, through a suitable cross-validation mechanism and hyperparameter optimization algorithm, a germination rate prediction model particularly suitable for soybean seeds is constructed. This model accurately maps the deep relationship between near-infrared spectral signals and seed germination, enabling non-destructive, efficient, and highly accurate detection of soybean seed germination ability.

[0008] Furthermore, the method for constructing the germination ability detection model includes: Multiple naturally aged soybean seeds were divided into a training set and a validation set; The training set near-infrared transmission spectral data were obtained by scanning soybean seeds in the training set with diffuse transmission near-infrared light. After performing standard normal variable transformation on the near-infrared transmission spectrum data of the training set, the data was trained based on the random forest algorithm. During the training process, the 5-fold cross-validation method was used for verification, and the grid search algorithm was used to optimize the parameters. The random forest algorithm uses sampling with replacement, randomly selecting the square root of the number of all features as a subset of candidate features during splitting, and the number of decision trees is 1500~1800.

[0009] The hyperparameters of the random forest algorithm used in this invention have been specifically optimized for diffuse transmission near-infrared spectral data of soybean seed germination, namely 'n_estimators': 1500~1800, bootstrap = True, max_features = 'sqrt', which can more effectively fit germination and spectral information and generate a prediction model with high accuracy.

[0010] Furthermore, the multi-grain naturally aged soybean seeds include seeds with germination rates of 90-100%, 70-90%, 30-70%, and 0-30%. Preferably, the multiple naturally aged soybean seeds are selected in the following manner: randomly select naturally aged seeds from different years and locations, and select seeds with germination rates of 95-100%, 70-90%, 30-60%, and 0-30%.

[0011] Furthermore, the number of naturally aged seeds with different germination rates is no less than 150.

[0012] The present invention preferably uses naturally aged seeds at a specific germination rate stage for model construction, and uses them as training samples, which can effectively improve the detection accuracy of the constructed model in practical application scenarios.

[0013] Furthermore, the parameters of the diffuse transmission near-infrared light scanning include: each spectrum is a composite value of 50 to 75 transmission scans, with a wavelength range of 11550 to 5800 cm⁻¹. -1 The resolution is 28~34 cm. -1 .

[0014] The preferred diffuse transmission near-infrared light scanning method of this invention can penetrate deeper into the soybean seed structure, precisely focusing on key germination tissues to capture the near-infrared spectra of components and physical structures such as water, proteins, and lipids during the germination process. Based on this high-quality data, combined with the aforementioned specific algorithm configuration, the mapping relationship between soybean seed germination-related factors and spectral signals can be more accurately reflected, thereby enabling the constructed prediction model to possess extremely high accuracy and robustness.

[0015] Furthermore, the diffuse transmission near-infrared scanning adopts a focused spot mode, with the focal point located on the epicotyl of the soybean seed, and the spot size is 1~2mm.

[0016] The preferred diffuse transmission near-infrared scanning method of this invention can better capture structural information related to the germination rate of soybean seeds. Furthermore, this invention also attempted to obtain near-infrared spectral data by scanning one side of the cotyledons of the soybean seed, but its prediction accuracy on the validation set decreased significantly. This makes it even more difficult to achieve high detection accuracy in practical applications.

[0017] Furthermore, the germination data of the multiple naturally aged soybean seeds represents the germination rate of soybean seeds from the same batch.

[0018] Secondly, the present invention provides a system for detecting the germination ability of soybeans, comprising: The near-infrared spectroscopy acquisition module is used to scan the seeds of soybeans to be tested with diffuse transmission near-infrared light to obtain near-infrared spectral results; The germination ability calculation module is used to import the near-infrared spectral results into the germination ability detection model to obtain the germination ability of the soybean to be tested.

[0019] For example, the germination ability of a single seed can be predicted using the method described above, and then the overall germination ability can be determined based on the germination ability results of all seeds.

[0020] Furthermore, it also includes: The germination ability detection model construction module is used to construct a germination ability detection model using the following methods: Multiple naturally aged soybean seeds were divided into a training set and a validation set; The training set near-infrared transmission spectral data were obtained by scanning soybean seeds in the training set with diffuse transmission near-infrared light. The germination status of each seed was investigated using the paper-based method of the national standard method, and the results were correlated with the spectrum. After performing standard normal variable transformation on the near-infrared transmission spectrum data of the training set, the data was trained based on the random forest algorithm. During the training process, the 5-fold cross-validation method was used for verification, and the grid search algorithm was used to optimize the parameters. Finally, the data was verified using the test set.

[0021] The present invention has the following beneficial effects: This invention provides a method for detecting the germination ability of soybean seeds based on diffuse transmission near-infrared spectroscopy. The germination ability detection model, constructed using specific samples and methods, achieves high accuracy while performing non-destructive testing of soybean seeds. This invention utilizes naturally aged seeds, and the germination ability detection model built based on full-band near-infrared spectral data enables efficient and accurate prediction of soybean seed germination ability, which has significant application value in the field of agricultural testing. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in this invention 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 invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the soybean seed germination ability detection provided in Embodiment 1 of the present invention.

[0024] Figure 2 This is a flowchart of the construction process of the germination ability detection model provided in Embodiment 1 of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of this invention, not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0026] Unless otherwise specified, the experimental methods involved in the following embodiments are conventional methods in the art. For example, you can refer to the experimental manual in the art or follow the conditions recommended in the manufacturer's instructions.

[0027] Unless otherwise specified, all experimental materials and reagents used in the following examples are commercially available.

[0028] Example 1: Construction of a Germination Capacity Detection Model 1. Dataset creation.

[0029] Soybean seeds from different years (e.g., 2021, 2022, 2023) and different locations (e.g., Shunyi, Hainan, Yangling, Chile) were randomly selected, and the germination rate was determined using the paper-based method (GB 3543 / 4-2025) in the national standard method. Batches with germination rates of 95%, 80%, 65%, and 20% were selected and mixed in equal proportions to form materials containing 1200 seeds.

[0030] The near-infrared spectra of the composite material were collected by performing near-infrared diffuse transmission scanning of the epicotyl region (using a focused spot mode with the focal point located on the epicotyl and a spot size of 1–2 mm). Each spectrum is a composite of 64 transmission scans, with a wavelength range of 11550–5800 cm⁻¹. -1 The resolution is 32 cm. -1 .

[0031] Then, the germination status was investigated in batches using the paper-based method of the national standard method, and the results were correlated with the spectrum.

[0032] 2. Model building.

[0033] During the training process, the present invention randomly splits the dataset of 1200 seeds into a training set containing 700 seeds and a test set containing 500 seeds.

[0034] After processing the transmission spectral data using standard normal variable transformation (SNV), the data was trained and validated using a random forest algorithm combined with 5-fold cross-validation. The optimal parameters were then calculated using a grid search algorithm ('n_estimators': 1700, bootstrap = True, max_features = 'sqrt').

[0035] This invention names this model based on random forest and related hyperparameters the Germinscope model. The model achieves an accuracy of 83.35% in predicting the germination rate of individual seeds in the test set. The overall predicted germination rate for the test set is 57.29%, while the actual germination rate is 55.29%.

[0036] When using the diffuse reflectance spectra of seeds from the same batch and the same modeling process, the accuracy of predicting the germination rate of individual seeds in the test set reached 75.25%. The overall predicted germination rate for the test set was 61.27%, while the actual germination rate was 55.29%.

[0037] Figure 1 This is a flowchart for testing the germination ability of soybean seeds. Figure 2 This is a flowchart of the construction process for the germination ability detection model.

[0038] Example 2 One sample containing 200 seeds was received. After powering on the near-infrared scanner, the diffuse transmission mode was selected, and each seed was placed with its hypocotyl facing down in the near-infrared scanning window. The near-infrared spectrum of each sample was obtained from 64 uniform scans (scanning method was the same as in Example 1). After exporting the near-infrared spectrum as a CSV file, the spectral data was analyzed using the Germinscope model (Example 1), and the results were exported as an Excel file. The predicted germination rate of the 200 seeds was 24. Therefore, the germination rate of this sample was 12%.

[0039] The actual germination rate of this batch of samples was 15.7% (based on the GB 3543 / 4-2025 method in Example 1). When the diffuse transmission mode was adjusted to diffuse reflection mode, the discrimination result was 22%.

[0040] Example 3 One sample containing 200 seeds was received. After powering on the near-infrared scanner, the diffuse transmission mode was selected, and the seeds were placed one by one with the hypocotyl facing down in the near-infrared scanning window. The near-infrared spectrum of each sample was obtained from 32 uniform scans. After exporting the near-infrared spectrum as a CSV file, the spectral data was analyzed using the Germinscope model (Example 1), and the results were exported as an Excel file. The prediction result for the 200 seeds was that 0 would not germinate. Therefore, the germination rate of this sample was 100%.

[0041] The actual germination rate of this batch of samples was 96% (based on the GB 3543 / 4-2025 method in Example 1). When the diffuse transmission mode was adjusted to diffuse reflection mode, the discrimination result was 85%.

[0042] Example 4 One sample containing 500 seeds, known to be of poor quality, was received. After turning on the near-infrared scanner and selecting diffuse transmission mode, the seeds were placed one by one with the hypocotyl facing down in the near-infrared scanning window. The near-infrared spectrum of each sample was obtained from 64 uniform scans. After exporting the near-infrared spectrum as a CSV file, the spectral data was analyzed using the Germinscope model (Example 1), and the results were exported as an Excel file. The results predicted a germination rate of 53% for this sample. After selecting all seeds predicted not to germinate, the actual germination rate of this batch of samples upon planting was 88%, making it a high-quality seed material.

[0043] The actual germination rate of this batch of samples was 55% (based on the GB 3543 / 4-2025 method in Example 1). When the diffuse transmission mode was adjusted to diffuse reflection mode, the discrimination result was 40%.

[0044] Modeling was performed using different data preprocessing methods for Example 1. 1. This comparative example uses the dataset collected and established in Example 1. During the training process, the present invention randomly splits the dataset of 1200 seeds into a training set containing 700 seeds and a test set containing 500 seeds.

[0045] 2. After processing the diffuse transmittance spectral data using the second derivative and standard normal variable transformation (SNV), training and validation are performed based on the random forest algorithm combined with 5-fold cross-validation. Then, the optimal parameters are calculated using a grid search algorithm. The remaining steps are the same as in Example 1.

[0046] The model constructed using this comparative example achieved an accuracy of 77.31% in predicting the germination rate of individual seeds in the test set. The predicted germination rate for this batch of seeds was 61.08%, while the actual measured germination rate was 55.29%. This result shows that the prediction accuracy significantly decreased after employing different data preprocessing methods.

[0047] Comparative Example 2 uses different algorithms for modeling 1. This comparative example uses the dataset collected and established in Example 1. During the training process, the present invention randomly splits the dataset of 1200 seeds into a training set containing 700 seeds and a test set containing 500 seeds.

[0048] 2. After processing the diffuse transmittance spectral data using Standard Normal Variable Transform (SNV), training and validation are performed based on a linear discriminant algorithm combined with 5-fold cross-validation. The optimal parameters are then calculated using a grid search algorithm. The remaining steps are the same as in Example 1.

[0049] The model constructed using this comparative example achieved an accuracy of 71.25% in predicting the germination rate of individual seeds in the test set. The predicted germination rate for this batch of seeds was 62.5%, while the actual measured germination rate was 55.29%. This result shows that the prediction accuracy significantly decreased after employing different algorithms.

[0050] Comparative Example 3 was modeled using different hyperparameters. 1. This comparative example uses the dataset collected and established in Example 1. During the training process, the present invention randomly splits the dataset of 1200 seeds into a training set containing 700 seeds and a test set containing 500 seeds.

[0051] 2. The Random Forest algorithm is based on ('n_estimators': 2000, bootstrap = True, max_features = 'log2'). The remaining steps are the same as in Example 1.

[0052] The model constructed using this comparative example achieved an accuracy of 78.01% in predicting the germination rate of individual seeds in the test set. The predicted germination rate for this batch of seeds was 51.89%, while the actual measured germination rate was 55.29%. This result shows that the prediction accuracy significantly decreased after using different hyperparameters.

[0053] Comparative Example 4 1. This comparative study uses different datasets for model construction. Specifically, the experimental materials used are as follows: Two hundred soybean seeds were placed in an environment of 40 degrees Celsius and 100% humidity for 72 hours for aging treatment.

[0054] The same near-infrared spectroscopy acquisition method and germination rate detection method as in Example 1 were used to construct the dataset.

[0055] 2. After processing the transmission spectrum data using standard normal variable transformation (SNV), 170 samples were trained and validated based on the random forest algorithm combined with 5-fold cross-validation. The optimal parameters were then calculated using a grid search algorithm ('n_estimators': 50, bootstrap = True, max_features = 'sqrt').

[0056] Using this model to predict the remaining 30 test grains, the single-grain accuracy was 83.33%.

[0057] Using this model, the germination rate of 400 unaged seeds in the same batch was predicted. The actual germination rate of this batch of seeds was 100%, the predicted germination rate was 65.5%, and the single-seed accuracy rate was 65.5%.

[0058] Using the Germinscope model constructed in Example 1 to predict the germination rate of the same batch of samples, the predicted germination rate was 97.25%, and the single-seed accuracy was 97.25%. This shows that while modeling with artificially aged seeds can achieve a high level of accuracy on the validation set, the accuracy drops significantly in real-world applications. The method provided by this invention maintains high accuracy and is more suitable for real-world seed germination rate detection scenarios.

[0059] Comparative Example 5 This comparative example uses diffuse transmission infrared spectroscopy detection at different scanning positions, specifically including: 1. Using near-infrared diffuse transmission mode, 300 soybean seeds were scanned one by one from one side of the cotyledons, and the germination rate was determined using the same national standard method (GB 3543 / 4-2025) as in Example 1. The data were correlated and merged to construct a dataset.

[0060] 2. After processing the transmission spectrum data using standard normal variable transformation (SNV), 240 samples were trained and validated using the random forest algorithm combined with 5-fold cross-validation. The optimal parameters were then calculated using a grid search algorithm ('n_estimators': 1000, bootstrap = True, max_features = 'sqrt').

[0061] Using this model to predict the remaining 60 test seeds, the single-seed accuracy was 70.0%, the predicted germination rate was 75%, and the actual germination rate of the test seed set was 60%.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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; and these 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 detecting the germination ability of soybeans, characterized in that, include: Near-infrared spectra were obtained by diffuse transmission near-infrared light scanning of soybean seeds. The near-infrared spectroscopy results are imported into the germination ability detection model to obtain the germination ability of the soybean to be tested; The germination ability detection model is constructed based on the near-infrared transmission spectrum and germination data of multiple naturally aged soybean seeds. During the training process, the spectral data is preprocessed using a standard normal variable transformation, trained using a random forest model, optimized using a grid search algorithm, and evaluated using 5-fold cross-validation.

2. The method according to claim 1, characterized in that, The method for constructing the germination ability detection model includes: Multiple naturally aged soybean seeds were divided into a training set and a validation set; The training set near-infrared transmission spectral data were obtained by scanning soybean seeds in the training set with diffuse transmission near-infrared light. After performing standard normal variable transformation on the near-infrared transmission spectrum data of the training set, the data was trained based on the random forest algorithm. During the training process, the 5-fold cross-validation method was used for verification, and the grid search algorithm was used to optimize the parameters. The random forest algorithm uses sampling with replacement, randomly selecting the square root of the number of all features as a subset of candidate features during splitting, and the number of decision trees is 1500~1800.

3. The method according to claim 1 or 2, characterized in that, The multi-grain naturally aged soybean seeds include: seeds with germination rates of 90-100%, 70-90%, 30-70%, and 0-30%; Preferably, the multiple naturally aged soybean seeds are selected in the following manner: randomly select naturally aged seeds from different years and locations, and select seeds with germination rates of 95-100%, 70-90%, 30-60%, and 0-30%.

4. The method according to claim 3, characterized in that, The number of naturally aged seeds with different germination rates was no less than 150.

5. The method according to any one of claims 1-4, characterized in that, The parameters of the diffuse near-infrared light scan include: each spectrum is the combined value of 50 to 75 transmission scans, with a wavelength range of 11550 to 5800 cm⁻¹. -1 The resolution is 28~34 cm. -1 .

6. The method according to any one of claims 1-5, characterized in that, The diffuse near-infrared scanning uses a focused spot mode, with the focal point located on the epicotyl of the soybean seed, and the spot size is 1~2mm.

7. The method according to any one of claims 1-6, characterized in that, The germination data of the multiple naturally aged soybean seeds represents the germination rate of soybean seeds from the same batch.

8. A system for detecting the germination ability of soybeans, characterized in that, include: The near-infrared spectroscopy acquisition module is used to perform diffuse transmission near-infrared light scanning on multiple seeds of soybean to obtain near-infrared spectral results. The germination ability calculation module is used to import the near-infrared spectral results into the germination ability detection model to obtain the germination ability of the soybean to be tested.

9. The system according to claim 8, characterized in that, Also includes: The germination ability detection model construction module is used to construct a germination ability detection model using the following methods: Multiple naturally aged soybean seeds were divided into a training set and a validation set; The training set near-infrared transmission spectral data were obtained by scanning soybean seeds in the training set with diffuse transmission near-infrared light. The germination status of each seed was investigated using the paper-based method of the national standard method, and the results were correlated with the spectrum. After performing standard normal variable transformation on the near-infrared transmission spectrum data of the training set, the data was trained based on the random forest algorithm. During the training process, the 5-fold cross-validation method was used for verification, and the grid search algorithm was used to optimize the parameters. Finally, the data was verified using the test set.

Images