A deep learning-based method and system for predicting crop seed quality

Abstract

This invention discloses a deep learning-based method and system for predicting crop seed quality. It uses infrared spectral data of multiple crop seeds and corresponding quality indicators as a training set to train a crop seed quality prediction model based on a convolutional neural network, resulting in a trained crop seed quality prediction model. The quality indicators are used to evaluate the quality of crop seeds. The crop seed quality prediction model is used to predict the quality of crop seeds based on their infrared spectral data. Taking advantage of the large wavelength range and numerous sampling points in the near-infrared spectral characteristics of crop seeds, this invention designs a one-dimensional convolutional neural network structure and introduces Inception and LSTM modules, enabling the network to extract more diverse feature information and achieve more accurate predictions.

Description

Technical Field

[0001] This invention belongs to the field of non-destructive testing technology for crop seeds, specifically relating to a method and system for predicting crop seed quality based on deep learning. Background Technology

[0002] Detecting the quality of relevant components in crop seeds is crucial for assisting experts in seed selection and breeding, guiding farmers in production, and increasing crop yields. Traditional agricultural product quality testing techniques, such as morphological methods, protein electrophoresis, and DNA molecular markers, require expensive precision instruments and involve destroying the structure of crop seeds to obtain quality data. This approach is not only inefficient but also relies heavily on expert experience for data evaluation, making it subjective and unsuitable for the current needs of large-scale and industrialized crop seed quality testing in my country. Therefore, developing non-destructive rapid testing technologies for crop seed quality has become an important research topic. With the development of near-infrared spectroscopy, its applications in non-destructive testing are increasing. However, in the field of non-destructive testing of crop seeds, most near-infrared spectroscopy techniques rely on traditional algorithms for modeling, resulting in low accuracy and an inability to support high-precision crop seed quality testing. Summary of the Invention

[0003] This invention employs near-infrared spectroscopy for non-destructive testing of crop seeds and applies deep learning methods to non-destructive testing, proposing a deep learning-based crop seed quality prediction method and system. This method is based on 1D-CNN-LSTM and near-infrared spectroscopy for non-destructive testing of crop seed quality, and the 1D-CNN-LSTM network structure is used to accurately predict the quality of crop seeds.

[0004] One of the objectives of this invention is a deep learning-based method for predicting crop seed quality, comprising:

[0005] Construct a crop seed quality prediction model based on convolutional neural networks;

[0006] The model is trained using infrared spectral data of multiple crop seeds and their corresponding quality indicators as a training set to obtain a trained crop seed quality prediction model. The quality indicators are used to evaluate the quality of crop seeds. The crop seed quality prediction model is used to predict the quality of rapeseed based on the infrared spectral data of crop seeds.

[0007] Furthermore, the convolutional neural network includes an Inception module, which is located in one or more convolutional layers of the convolutional neural network, and is used to convert each convolutional kernel in the same convolutional layer into multiple convolutional kernels of different sizes. The size of the converted convolutional kernels is smaller than the size of the original convolutional kernels, so as to extract more spectral information features.

[0008] Spectral data can be viewed as a very long one-dimensional vector. When processing such vectors, convolutional neural networks have difficulty establishing global connections, such as the beginning and end information of the spectrum. LSTM networks can extract global features better. Therefore, the convolutional neural network further includes an LSTM (Long Short-Term Memory) module, which is located between the fully connected layer and the stretching layer in the convolutional neural network. It is used to extract spectral information features to enhance the connection between spectral sampling points.

[0009] Furthermore, the method also includes data removal of the infrared spectral data of the multiple crop seeds to eliminate environmental factors that cause some sample spectra to deviate from the population sample set.

[0010] Furthermore, the method also includes preprocessing the infrared spectral data of the multiple crop seeds to remove noise signals and baseline drift signals present in the original spectral images.

[0011] Furthermore, the preprocessing methods include Savitsky-Golay smoothing and / or multivariate scattering correction and / or standard normal transformation and / or mean centering and / or normalization.

[0012] Furthermore, the output layer of the convolutional neural network uses the softmax activation function; the convolutional layer, stretching layer, pooling layer, and fully connected layer use the ReLU (Rectified LinearUnit) activation function; and the softmax activation function is a normalized exponential function.

[0013] A second objective of this invention is a crop seed quality prediction system based on deep learning, comprising a convolutional neural network construction module, an infrared spectral data acquisition module, a quality index acquisition module, and a model training module.

[0014] The convolutional neural network building module is used to build a crop seed quality prediction model based on convolutional neural networks;

[0015] The infrared spectral data acquisition module is used to acquire infrared spectral data of multiple crop seeds;

[0016] The quality index acquisition module is used to obtain quality indexes for multiple crop seeds.

[0017] The model training module is used to train a crop seed quality prediction model; the crop seed quality prediction model is used to predict the quality of crop seeds based on their infrared spectral data.

[0018] Furthermore, the convolutional neural network includes an Inception module and / or an LSTM module. The Inception module is located in one or more convolutional layers of the convolutional neural network and is used to convert each convolutional kernel in the same convolutional layer into multiple convolutional kernels of different sizes. The size of the converted convolutional kernels is smaller than the size of the original convolutional kernels, so as to extract more spectral information features. The LSTM module is located between the fully connected layer and the stretching layer in the convolutional neural network and is used to extract spectral information features to enhance the connection between spectral sampling points.

[0019] Beneficial effects:

[0020] This invention takes crop seeds as the research object. Based on the characteristics of the large wavelength range and many sampling points of near-infrared spectral features, a one-dimensional convolutional neural network structure is designed, and an Inception module and an LSTM module are introduced to enable the network to extract more diverse feature information and make more accurate predictions.

[0021] Near-infrared spectral information was collected using a spectral acquisition instrument. The protein and oil content of rapeseed seeds were determined using a platform, and outlier samples were removed to construct a standard dataset. The proposed 1D-CNN-LSTM network model and the conventional PLS algorithm model were trained separately, and the results were compared. In the protein evaluation index, R² improved by 0.05, and RMSEP decreased by 0.21; in the oil content evaluation index, R² improved by 0.07, and RMSEP decreased by 0.81, showing significant improvements in recognition accuracy.

[0022] In the non-destructive testing of crop seeds, this invention introduces deep learning for the first time to predict their quality. By comparing with traditional methods, it can be concluded that the deep learning method is still applicable to the testing of crop seeds and has a significant improvement in accuracy. Attached Figure Description

[0023] Figure 1 This is the structure of the 1D-CNN-LSTM network described in this invention;

[0024] Figure 2 This refers to the LSTM module structure in a 1D-CNN-LSTM network.

[0025] Figure 3 This refers to the Connection module structure in a 1D-CNN-LSTM network.

[0026] Figure 4(a) shows the image of the original spectrum;

[0027] Figure 4(b) shows the image after multivariate scattering correction of the original spectrum;

[0028] Figure 4(c) shows the image after standard normal transformation of the original spectrum;

[0029] Figure 4(d) shows the image after mean centering of the original spectrum;

[0030] Figure 4(e) shows the image after normalization of the original spectrum;

[0031] Figure 5 These are the training steps for the prediction model;

[0032] Figure 6 This is a visualization of the protein content predicted using this invention;

[0033] Figure 7 This is a visualization result of the oil content predicted using this invention;

[0034] Figure 8 This is a schematic diagram of the structure of the present invention. Detailed Implementation

[0035] The following detailed embodiments are provided to explain the technical solutions of the claims of this invention, so that those skilled in the art can understand the claims. The scope of protection of this invention is not limited to the following specific embodiments. Any modifications made by those skilled in the art that incorporate the technical solutions of the claims but differ from the following detailed embodiments are also within the scope of protection of this invention.

[0036] Rapeseed is one of the world's three major oilseed crops and has long been the most widely used oilseed crop in my country, holding an important position in the domestic vegetable oil market. Studies have shown that rapeseed quality has significant reference value for oil yield. Rapid and accurate detection of rapeseed-related components is crucial for assisting experts in seed selection and breeding, guiding farmers' production, and increasing crop yield. Rapeseed quality evaluation, classification, and quality testing are important processes in researching high rapeseed yields. This application uses rapeseed, a widely used crop seed, as an example to illustrate a deep learning-based method for crop seed quality detection, including the following steps:

[0037] Construct a crop seed quality prediction model based on convolutional neural networks;

[0038] The model is trained using multiple sets of infrared spectral data of rapeseed and their corresponding quality indicators as a training set to obtain a trained rapeseed quality prediction model. The quality indicators are used to evaluate the quality of rapeseed. The crop seed quality prediction model is used to predict the quality of rapeseed based on its infrared spectral data.

[0039] The technical solution described above includes the following methods for collecting infrared spectral data and corresponding quality indicators of rapeseed:

[0040] 1) Rapeseed quality and near-infrared spectral data acquisition

[0041] a. Using Antaris II The FT-NIR analyzer was used to collect infrared spectra of rapeseed. The wavelength range of the analyzer is 830–2630 nm (12000–3800 c⁻¹). The sampler is a transparent cylinder. To prevent light scattering, at least one layer of rapeseed should be spread in the sample cup to cover the sampler of the near-infrared spectrometer. Each rapeseed sample weighing about 14 grams was placed in the sampler, which was positioned at the light aperture of the near-infrared spectrometer. The RAPESEED model corresponding to the spectrometer was used to extract the spectral information of the rapeseed. A total of 180 sets of spectral data were extracted, and each set of spectral data was used as a sample. The samples were arranged and saved according to their numbers. Samples 1–12 were of the Huayouza 62 variety, samples 13–66 were a mixture of other mixed varieties, samples 67–90 were high-yield rapeseed, samples 91–120 were rapeseed harvested from the experimental field, and samples 121–180 were mixed samples of various varieties.

[0042] b. Sample quality determination. Physicochemical values ​​were measured using traditional methods by the Oil Crops Research Institute of the Chinese Academy of Agricultural Sciences. The quality determination included the determination of oil content and crude protein content of rapeseed.

[0043] (1) Determination of oil content: In this embodiment, the residual method is used to determine the oil content of oilseeds. First, samples are taken according to GB5491, dried in a drying oven, passed through a 0.42mm straight sieve, and then placed in ground glass bottles for later use. Petri dishes and paper packets are prepared according to standards. After soaking and extracting the sample packets in anhydrous ether in a fat extractor, the sample packets are removed and placed in a ventilated place to allow the ether to evaporate. Then, the sample packets are weighed, dried, and weighed again. The oil content (w) of rapeseed is calculated using the following formula:

[0044]

[0045] In the formula:

[0046] m represents the moisture content of rapeseed, determined according to the national standard GB / T 5497, and the unit is mass fraction (%).

[0047] a, b, and c represent the paper package mass, the paper package mass plus the dried sample mass, and the paper package mass plus the sample mass after sampling, respectively. All units are grams (g), and the calculations are rounded to two decimal places.

[0048] (2) Determination of crude protein content: According to GB5009.5-2016, the traditional method for determining the crude protein content of solid samples such as grains, beans, milk powder, rice flour, and protein powder with a protein content of 10g / 100g or higher should use the combustion method, which involves burning the sample at a high temperature of 900℃-1200℃. The mixture of gases produced during combustion (NO) x The nitrogen gas is transported to the reduction furnace by a carrier gas such as nitrogen. After reduction at 800°C, the nitrogen gas is then passed through a thermal conductivity detector (TCD) to detect its content.

[0049] The above technical solution also includes preprocessing the infrared spectral data of the multiple crop seeds to remove noise and baseline drift signals present in the original spectra. The data preprocessing methods used in this embodiment include the following five: Savitsky-Golay smoothing (SG), multivariate scattering correction (MSC), standard normal transformation (SNV), mean centering (CT), and standardization (SS). This invention does not limit the execution order of each preprocessing method.

[0050] The Savitsky-Golay (SG) smoothing method is used for preprocessing in spectral analysis. It performs a k-th order polynomial fitting on data points within a certain window to obtain the fitted result. Using the Savitsky-Golay method for smoothing filtering can improve the smoothness of the spectrum and reduce noise interference.

[0051] Multivariate scattering correction (MSC) algorithm is used for the preprocessing of hyperspectral data. MSC can effectively eliminate spectral differences caused by different scattering levels, thereby enhancing the correlation between spectra and data. This method corrects baseline shifts and offsets in spectral data using ideal spectra.

[0052] The Standard Normal Transform (SNV) is used to correct spectral errors caused by scattering between samples.

[0053] Standardization (SS) ensures that each column of data (i.e., each feature data) in the dataset is standardized uniformly. Specifically, Z-Score normalization is used to guarantee that the mean of each feature dimension is 0 and the variance is 1.

[0054] Save the preprocessed data. The original spectral image and the preprocessed spectral image are shown below. Figures 4(a) to 4(e) As shown:

[0055] In the above technical solution, due to errors caused by human factors and interference from environmental factors during the spectral data acquisition process, the spectra of some samples deviate from the population sample set. These sample spectra contain a lot of irrelevant information and carry fewer data features, resulting in poor prediction performance of the established model for unknown samples. Therefore, before data preprocessing, the infrared spectral data of the multiple crop seeds are further processed to remove environmental interference that causes the spectra of some samples to deviate from the population sample set, thus ensuring the accuracy of the established model. The removal method used in this invention is the Mahalanobis distance algorithm, which identifies outliers based on the Mahalanobis distance between samples and removes the samples corresponding to the outliers.

[0056] After removing samples, to ensure the effectiveness and accuracy of the model, the sample set needs to be partitioned. In this embodiment, the sample set is divided into a training set and a validation set in a 4:1 ratio, with a final training set of 140 samples and a validation set of 35 samples. Rapeseed oil content prediction models and protein content prediction models based on 1D-CNN-LSTM (1D Convolutional Neural-LongShort-Term Memory Networks) are established respectively. The two prediction models are trained using the training set, and then validated and evaluated using the validation set samples.

[0057] In the above technical solution, the convolutional neural network includes an Inception module, which is located in one or more convolutional layers in the convolutional neural network. It is used to convert each convolutional kernel in the same convolutional layer into multiple convolutional kernels of different sizes. The size of the converted convolutional kernels is smaller than the size of the original convolutional kernels, so as to extract more spectral information features.

[0058] In the above technical solution, the convolutional neural network also includes an LSTM module, which is located between the fully connected layer and the stretching layer in the convolutional neural network, and is used to extract spectral information features to enhance the connection between spectral sampling points.

[0059] like Figure 1The rapeseed oil content and protein content prediction model consists of a convolutional neural network and a recurrent neural network: it is a quantitative prediction model 1D-CNN-LSTM based on a one-dimensional convolutional neural network, combining CNN and LSTM. The convolutional neural network model features sparse connections, parameter sharing, pooling, and a deep hierarchical structure. The input to the prediction model is the original spectrum of the rapeseed sample, which is a 1*1556 one-dimensional vector. The output is the predicted oil content and crude protein content of the rapeseed. The entire network model adopts a basic design of a one-dimensional convolutional neural network. The convolutional layers mainly extract features through convolution operations. The convolutional kernel slides along the one-dimensional vector and performs convolution operations to generate feature maps. An Inception module is used between convolutional layers 2 and 3, with a structure... Figure 3 As shown, two 1*1 convolutional kernels, one 3*1 convolutional kernel, and one 5*1 convolutional kernel are set in convolutional layer 3 and connected in parallel. This allows the prediction model to extract features of different sizes in the same convolutional layer, increasing the width and depth of the spectral analysis model and enriching the extracted spectral information. The extracted feature information can better utilize its temporal information through the LSTM network. The spectral data undergoes continuous data transformation in the prediction model through the operations of the convolutional layers and the LSTM network, and finally the prediction result is output through the output layer.

[0060] Compared to fully connected models, convolutional neural networks significantly reduce the number of parameters through local connections between two layers of neurons. LTTM networks can extract temporal information from spectral data, making predictions more accurate. The structure of a single LSTM is as follows: Figure 2 As shown, the calculation formula is as follows.

[0061]

[0062] In the formula:

[0063] h t and h t-1 These represent the outputs of the t-th layer and the (t-1)-th layer, respectively.

[0064] i t for h t-1 With x t The product is then processed by the activation function σ and output.

[0065] f t Forgotten Gate;

[0066] tanh is the activation function;

[0067] o t for h t-1 With x tThe output after convolution and dot product operations, and then through the activation function σ.

[0068] W xi W xf W xo W xc They represent x respectively i x f x o x c The weights are given by i, f, o, and c, which correspond to the input i, the forget gate f, o, and c, respectively. t c t ;

[0069] W hi W hf W ho W hc They represent h respectively i h f h o h c The weights;

[0070] W ci W cf W co They represent c respectively i c f c o The weights;

[0071] c t-1 and c t These represent the output of the previous layer t-1 and the output of the current layer t, respectively.

[0072] b i b f b o b c These represent the biases corresponding to i, f, o, and c, respectively.

[0073] x t This is the input for layer t.

[0074] Model training methods such as Figure 5 The diagrams illustrate training the model using the traditional PLS method and training it using the 1D-CNN-LSTM-based model described in this invention, respectively. The training steps are described below using the 1D-CNN-LSTM-based model:

[0075] a. Construction of the experimental platform.

[0076] The experimental platform consisted of a Windows 10 64-bit operating system and CUDA version 10.0. A one-dimensional convolutional neural network was built using PyTorch as the learning framework and Python code. The primary prediction parameters were protein and oil content.

[0077] b. Data Processing and Model Training. The input data for the 1D-CNN-LSTM model is the entire spectrum, with a total of 1557 wavenumbers. Therefore, a 1*1557 vector is input. After the spectral data is input, it passes through convolutional layers, pooling layers, stretching layers, an LSTM module, and a fully connected layer. Finally, the output layer outputs the prediction result. The last layer uses the softmax activation function, while the others use the ReLU activation function. The predicted values ​​are compared with the true values, and MAE (mean square error) is used as the loss function. Backpropagation is performed to update the weights until the loss function converges. The specific hyperparameter settings are shown in Table 1.

[0078]

[0079] Table 11 Hyperparameter Settings for D-CNN-LSTM Network

[0080] c. Model Testing

[0081] In training, this invention employs multiple iterations, selecting the optimal model from different iteration counts for testing. During model testing, the predicted results are compared with the actual results, and the root error prediction RMSEP and correlation coefficient R are calculated. 2 .

[0082] Evaluation criteria

[0083] To quantify the effectiveness of this invention, the root mean square error prediction (RMSEP) and correlation coefficient (R) were used. 2 As an evaluation indicator.

[0084] The formula for calculating the root mean square error prediction (RMSEP) is as follows:

[0085]

[0086] Correlation coefficient R 2 The calculation formula is as follows:

[0087]

[0088] After training, it was found that the loss tended to stabilize after 150 iterations. Therefore, the training model was finally obtained at epoch=150. The model at this point was saved, and protein content and oil content were predicted according to the evaluation indicators. The prediction results are as follows. Figure 6 and Figure 7As shown in Table 3, the model results for protein and oil content under different preprocessing conditions are presented. The results indicate that when the standard normal transformation (SNV) preprocessing method is used, the R-value of the protein 1D-CNN-LSTM model is significantly improved. 2 The maximum index is 0.94, while the minimum mean squared error (RMSE) is 0.21, indicating the best model performance. When using SG smoothing preprocessing, the R-value of the 1D-CNN-LSTM model for oil content is [value missing]. 2 The maximum index is 0.78, and the minimum mean squared error (RMSE) is 0.29.

[0089]

[0090] Table 2 Comparison of various pretreatment methods

[0091] To compare with traditional methods, this invention uses PLS, the best-performing traditional method, for comparative experiments. By comparing the 1D-CNN-LSTM network model and the PLS model, we can conclude that, under the same preprocessing method, the one-dimensional convolutional neural network has higher prediction accuracy. This is because the spectrum of rapeseed seeds is a 1*1557 one-dimensional vector with a large number of sampling points. The PLS-based algorithm extracts less feature information and cannot fully reflect the seed information. In contrast, the one-dimensional convolutional neural network can extract deeper feature information, and we expanded the receptive field through the Inception module, making the extracted spectral information more diverse and the prediction results more accurate. Compared with the PLS algorithm, as shown in Table 3, the convolutional neural network used in this invention shows significant improvements in performance metrics, proving that the proposed method is superior to the best traditional algorithm in non-destructive testing of rapeseed.

[0092]

[0093] Table 3 Comparison of results between the present invention and the PLS algorithm.

[0094] from Figure 6 and Figure 7 The quality prediction visualization results shown indicate that the 1D-CNN-LSTM has a small error and can achieve accurate non-destructive detection of rapeseed quality.

[0095] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0096] This application also includes a crop seed quality prediction system, such as... Figure 8 As shown, it includes a convolutional neural network construction module, an infrared spectral data acquisition module, a quality indicator acquisition module, and a model training module;

[0097] The convolutional neural network building module is used to build crop seed quality prediction models based on convolutional neural networks;

[0098] The infrared spectral data acquisition module is used to acquire infrared spectral data from multiple crop seeds;

[0099] The quality index acquisition module is used to obtain quality indexes for multiple crop seeds.

[0100] The model training module is used to train a crop seed quality prediction model; the crop seed quality prediction model is used to predict the quality of crop seeds based on their infrared spectral data.

[0101] In the above technical solution, the convolutional neural network includes an Inception module and / or an LSTM module. The Inception module is located in one or more convolutional layers in the convolutional neural network and is used to convert each convolutional kernel in the same convolutional layer into multiple convolutional kernels of different sizes. The size of the converted convolutional kernels is smaller than the size of the original convolutional kernels, so as to extract more spectral information features. The LSTM module is located between the fully connected layer and the stretching layer in the convolutional neural network and is used to extract spectral information features to enhance the connection between spectral sampling points.

[0102] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A method for predicting crop seed quality based on deep learning, characterized in that, include: Construct a crop seed quality prediction model based on convolutional neural networks; The model is trained using infrared spectral data of multiple crop seeds and their corresponding quality indicators as a training set to obtain a trained crop seed quality prediction model. The quality indicators are used to evaluate the quality of crop seeds. The crop seed quality prediction model is used to predict the quality of crop seeds based on their infrared spectral data. The convolutional neural network includes an Inception module, which is located in one or more convolutional layers of the convolutional neural network. It is used to convert each convolutional kernel in the same convolutional layer into multiple convolutional kernels of different sizes. The size of the converted convolutional kernels is smaller than the size of the original convolutional kernels, so as to extract more spectral information features. The convolutional neural network also includes an LSTM module, which is located between the fully connected layers and the stretching layers in the convolutional neural network, and is used to extract spectral information features to enhance the connection between spectral sampling points.

2. The crop seed quality prediction method based on deep learning as described in claim 1, characterized in that, It also includes data removal of the infrared spectral data of the multiple crop seeds to eliminate environmental factors that cause some sample spectra to deviate from the population sample set.

3. The crop seed quality prediction method based on deep learning as described in claim 2, characterized in that, It also includes preprocessing the infrared spectral data of the multiple crop seeds to remove noise signals and baseline drift signals present in the original spectral images.

4. The crop seed quality prediction method based on deep learning as described in claim 3, characterized in that, The preprocessing methods include Savitsky-Golay smoothing and / or multivariate scattering correction and / or standard normal transformation and / or mean centering and / or normalization.

5. The crop seed quality prediction method based on deep learning as described in claim 1, characterized in that, The output layer of the convolutional neural network uses the softmax activation function.

6. The crop seed quality prediction method based on deep learning as described in claim 1, characterized in that, The convolutional layers, stretching layers, pooling layers, and fully connected layers of the convolutional neural network use the ReLU activation function.

7. A crop seed quality prediction system employing the deep learning-based crop seed quality prediction method described in claim 1, characterized in that, It includes a convolutional neural network construction module, an infrared spectral data acquisition module, a quality indicator acquisition module, and a model training module; The convolutional neural network building module is used to build a crop seed quality prediction model based on convolutional neural networks; The infrared spectral data acquisition module is used to acquire infrared spectral data of multiple crop seeds; The quality index acquisition module is used to obtain quality indexes for multiple crop seeds. The model training module is used to train the crop seed quality prediction model using infrared spectral data of multiple crop seeds and corresponding quality indicators as a training set, thereby obtaining a trained crop seed quality prediction model; the crop seed quality prediction model is used to predict the quality of crop seeds based on the infrared spectral data of crop seeds.

Images