A Non-destructive Detection Method for Agricultural Metabolites Based on Two-Dimensional Correlation Spectroscopy

By using a two-dimensional correlation spectrum-based method combined with deep feature encoding and nonlinear regression models, the problem of insufficient detection accuracy and robustness in the detection of agricultural product metabolites was solved, achieving non-destructive and efficient detection of agricultural product metabolites.

CN123065602APending Publication Date: 2026-11-10JIANGNAN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611253920.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-18
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies for detecting metabolites in agricultural products suffer from low sensitivity, long processing time, high cost, and significant destructiveness, making it difficult to meet the non-destructive, high-throughput detection requirements for large-scale screening and online monitoring. Furthermore, traditional one-dimensional spectral modeling methods lack sufficient detection accuracy and robustness.

Method used

A two-dimensional correlation spectrum-based approach was adopted, which constructs synchronous and asynchronous two-dimensional correlation spectra through different spectral preprocessing methods. Combined with a deep feature encoding module and a nonlinear regression model, non-destructive detection of agricultural product metabolites was achieved.

Benefits of technology

It improves the accuracy and robustness of the detection of metabolites in agricultural products, enables rapid, non-destructive, high-throughput detection, and enhances the accuracy and stability of the detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123065602A_ABST
    Figure CN123065602A_ABST
Patent Text Reader

Abstract

This application discloses a non-destructive detection method for agricultural product metabolites based on two-dimensional correlation spectroscopy, relating to the field of hyperspectral technology. This method heterogeneously preprocesses effective spectral data using different spectral preprocessing methods, and then combines 2T2D-COS to obtain synchronous and asynchronous two-dimensional correlation spectra. These spectra are then input into a deep feature encoding module for feature encoding to obtain a low-dimensional deep feature vector of the agricultural product to be tested. This enhances weak spectral response information while preserving cross-band correlation structures, enabling the deep feature encoding module to learn nonlinear features that are difficult to extract using traditional PCA. Finally, the results are input into a nonlinear regression prediction model to obtain the metabolite detection results, achieving non-destructive detection of agricultural product metabolites. This method can effectively improve the prediction accuracy and model robustness of complex metabolites, enabling rapid and non-destructive detection of agricultural product quality.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of hyperspectral technology, and in particular to a non-destructive detection method for agricultural metabolites based on two-dimensional correlation spectroscopy. Background Technology

[0002] The composition and content of key metabolites such as carbohydrates, amino acids, and lipids in agricultural products are important indicators for assessing their nutritional function, storage stability, and processing suitability. With the upgrading of consumer demand for high-quality, functional agricultural products, establishing efficient and accurate metabolite detection methods has become a key technological requirement in the fields of quality evaluation, germplasm resource exploration, and intelligent processing.

[0003] Currently, the analysis of metabolites in agricultural products mainly relies on chromatography-mass spectrometry (GC-MS / HPLC-MS). While this type of method offers high detection sensitivity, it suffers from limitations such as complex sample preparation, long analysis time, high reagent costs, and destructive nature, making it difficult to meet the urgent need for non-destructive, high-throughput detection in large-scale screening of breeding populations and online monitoring of production lines.

[0004] In recent years, hyperspectral imaging technology has been widely used in agricultural product quality testing as a rapid and non-destructive detection method. However, the weak vibrational absorption intensity and severe band overlap of agricultural product metabolites in the near-infrared region limit the detection accuracy of traditional one-dimensional spectral modeling methods based on the entire or empirical bands. Secondly, the relationship between metabolite concentration and spectral response is usually significantly nonlinear, while traditional linear dimensionality reduction methods such as principal component analysis (PCA) can only preserve the global linear structure of the data, making it difficult to effectively extract the nonlinear information, resulting in insufficient feature extraction and inadequate generalization ability of the detection model. This leads to the current low accuracy and robustness of hyperspectral imaging-based agricultural product metabolite detection. Summary of the Invention

[0005] This application addresses the aforementioned problems and technical needs by proposing a non-destructive detection method for agricultural product metabolites based on two-dimensional correlation spectroscopy. The technical solution of this application is as follows: A non-destructive detection method for agricultural product metabolites based on two-dimensional correlation spectroscopy, comprising: Obtain effective spectral data of the agricultural products to be tested; Two different spectral preprocessing methods were used to preprocess the effective spectral data as reference spectra and spectra to be measured, and the 2T2D-COS method was used to obtain synchronous two-dimensional correlation spectra and asynchronous two-dimensional correlation spectra. After the synchronous two-dimensional correlation spectrum and the asynchronous two-dimensional correlation spectrum are concatenated into a dual-channel correlation spectrum image, the image is input into a trained deep feature encoding module for feature encoding to obtain a low-dimensional deep feature vector of the agricultural product to be tested. The low-dimensional deep feature vector is input into the trained nonlinear regression prediction model to obtain the metabolite detection results of the agricultural product to be tested.

[0006] The further technical solution is that the spectral preprocessing method used for the effective spectral data is one or more of the following combined processing methods: SG smoothing, standard normal variable transformation (SNV), multivariate scattering correction (MSC), and first derivative transformation (D1).

[0007] A further technical solution is that the non-destructive detection method for agricultural product metabolites also includes: Obtain spectral data of agricultural product samples; Construct multiple different combinations of spectral preprocessing methods, each combination including two different spectral preprocessing methods; Two spectral preprocessing methods from one of the spectral preprocessing methods were used to preprocess a set of sample spectral data as reference spectrum and spectrum to be measured. Then, the synchronous two-dimensional correlation spectrum and asynchronous two-dimensional correlation spectrum were obtained by using the 2T2D-COS method and stitched together to form a two-channel correlation spectrum image. The dual-channel correlation spectrum image of each agricultural product sample is input into the encoder of the convolutional autoencoder to obtain a low-dimensional deep feature vector. The decoder of the convolutional autoencoder is then used to reconstruct the dual-channel correlation spectrum image based on the low-dimensional deep feature vector. The mean square error between the dual-channel correlation spectrum image of each agricultural product sample and the corresponding reconstructed dual-channel correlation spectrum image is used as the loss function to train the convolutional autoencoder, and the encoder in the trained convolutional autoencoder is used as the trained deep feature encoding module.

[0008] The further technical solution is as follows: the encoder in the convolutional autoencoder uses a 1×1 convolutional layer to achieve information fusion and dimensionality adjustment of the input feature channels of the input dual-channel correlation spectrum image. Multiple 3×3 convolutional layers are used to perform progressive downsampling and gradually increase the number of feature channels to extract local correlation features and high-order spatial features in the dual-channel correlation spectrum image to obtain a high-dimensional feature tensor. Each convolutional layer uses the SELU activation function. The high-dimensional feature tensor is converted into a high-dimensional feature vector through a flattening operation, and then the high-dimensional feature vector is mapped into a low-dimensional depth feature vector through a fully connected layer.

[0009] The further technical solution is as follows: the dual-channel correlation spectrum image is 510×510×2. The deep feature encoding module uses five 3×3 convolutional layers with a stride of 2 to perform step-by-step downsampling, compressing the dual-channel correlation spectrum image to a high-dimensional feature tensor of 16×16×256. After the high-dimensional feature tensor is flattened into a high-dimensional feature vector, it is mapped to a 20-dimensional latent space through a fully connected layer to obtain the low-dimensional deep feature vector of the agricultural product to be tested.

[0010] The further technical solution is that the decoder in the convolutional autoencoder uses a fully connected layer to map the low-dimensional depth feature vector output by the encoder to a high-dimensional feature vector. Then, it is converted into a 16×16×256 three-dimensional feature tensor through a reconstruction operation. The tensor is then passed through five transposed convolutional layers for progressive upsampling and channel number compression. Finally, a reconstructed dual-channel correlation spectrum image with the same size as the dual-channel correlation spectrum image is obtained through a 1×1 convolutional layer and a size reconstruction operation.

[0011] A further technical solution involves stitching the synchronous two-dimensional correlation spectrum and the asynchronous two-dimensional correlation spectrum into a dual-channel correlation spectrum image, which also includes: After performing independent Z-score normalization on the synchronous two-dimensional correlation spectrum and the asynchronous two-dimensional correlation spectrum, the two-channel correlation spectrum image is obtained by stitching them together along the channel dimension.

[0012] The further technical solution involves obtaining effective spectral data of the agricultural product to be tested, including: After acquiring hyperspectral images of the area where the agricultural product to be tested is located and performing standardized correction processing, the agricultural product to be tested is segmented and the region of interest is extracted. The average value of the spectral response value of each pixel in the region of interest is calculated to obtain the effective spectral data.

[0013] A further technical solution is that the nonlinear regression prediction model adopts a random forest regression model.

[0014] The further technical solution is that the agricultural product to be tested is corn kernels, and the metabolite detection results of the agricultural product to be tested are lipid metabolites of corn kernels. The lipid metabolites of corn kernels include at least one of 9-linolenic acid, pinemanic acid, N-palmitoylethanolamine, N-stearoylethanolamine, N-oleoylethanolamine and propionic acid-related lipids.

[0015] The beneficial technical effects of this application are: This application discloses a non-destructive detection method for agricultural product metabolites based on two-dimensional correlation spectroscopy. This method heterogeneously preprocesses effective spectral data using different spectral preprocessing methods, and then combines 2T2D-COS to obtain synchronous and asynchronous two-dimensional correlation spectra. These spectra are then input into a deep feature encoding module for feature encoding to obtain a low-dimensional deep feature vector of the agricultural product to be tested. This enhances weak spectral response information while preserving cross-band correlation structures, enabling the deep feature encoding module to learn nonlinear features that are difficult to extract using traditional PCA. Finally, the results are input into a nonlinear regression prediction model to obtain the metabolite detection results, achieving non-destructive detection of agricultural product metabolites. This method can effectively improve the prediction accuracy and model robustness of complex metabolites, enabling rapid and non-destructive detection of agricultural product quality. Attached Figure Description

[0016] Figure 1This is a flowchart illustrating a method for non-destructive testing of agricultural product metabolites in one embodiment of this application.

[0017] Figure 2 This is a diagram of the network structure of the convolutional autoencoder used when training the deep feature encoding module. Detailed Implementation

[0018] The specific embodiments of this application will be further described below with reference to the accompanying drawings.

[0019] This application discloses a non-destructive detection method for agricultural product metabolites based on two-dimensional correlation spectroscopy. Please refer to [link / reference]. Figure 1 The non-destructive testing method for agricultural product metabolites includes the following steps: Step 110: Obtain the effective spectral data of the agricultural product to be tested.

[0020] The first step is to acquire hyperspectral images of the region where the agricultural product to be tested is located. The hyperspectral images contain images of the region of interest and the irrelevant background. Each hyperspectral image contains 512 bands in the spectral range of 900~1700nm, and the spatial resolution of each band is 2048×1088.

[0021] Then, the hyperspectral image is standardized and corrected to make the acquired image as accurate as possible. For example, the classic black and white plate correction is used, which will not be elaborated here.

[0022] Finally, the agricultural product to be tested is segmented and the region of interest is extracted. The average value of the spectral response of each pixel in the region of interest is calculated to obtain the effective spectral data of the agricultural product to be tested.

[0023] Step 120: Two different spectral preprocessing methods are used to preprocess the effective spectral data as reference spectra and spectra to be measured, and the 2T2D-COS method is used to obtain synchronous two-dimensional correlation spectra and asynchronous two-dimensional correlation spectra.

[0024] The spectral preprocessing methods used for the effective spectral data are one or more combinations of SG smoothing, standard normal variable transformation (SNV), multivariate scattering correction (MSC), and first derivative transformation (D1).

[0025] This application employs two different spectral preprocessing methods to preprocess the effective spectral data, using it as both a reference spectrum and the spectrum to be measured, to construct a 2T2D-COS dual-trace input. By calculating the synchronous and asynchronous correlation strengths between different wavelength variables, synchronous two-dimensional correlation spectral matrices and asynchronous two-dimensional correlation spectral matrices are obtained, respectively. The synchronous two-dimensional correlation spectral matrix describes the consistency of the changing trends of different bands during the perturbation process, reflecting the cooperative changing relationship between different wavelength variables; the asynchronous two-dimensional correlation spectral matrix describes the differences in the response sequence of different bands, reflecting the temporal variation characteristics between different wavelength variables.

[0026] Step 130: After concatenating the synchronous two-dimensional correlation spectrum and the asynchronous two-dimensional correlation spectrum into a dual-channel correlation spectrum image, input it into the trained deep feature encoding module for feature encoding to obtain the low-dimensional deep feature vector of the agricultural product to be tested.

[0027] Please refer to Figure 1 The deep feature encoding module first uses a 1×1 convolutional layer to fuse and adjust the dimensionality of the input dual-channel correlation spectrum image. Then, it uses multiple 3×3 convolutional layers to progressively downsample and increase the number of feature channels to extract local correlation features and higher-order spatial features from the dual-channel correlation spectrum image, resulting in a high-dimensional feature tensor. Finally, the high-dimensional feature tensor is flattened into a high-dimensional feature vector, which is then mapped to a low-dimensional deep feature vector through a fully connected layer. Each convolutional layer in the deep feature encoding module uses the SELU activation function instead of ReLU, fully preserving the physical meaning of the negative values ​​in the correlation spectrum matrix.

[0028] The synchronous two-dimensional correlation spectrum and the asynchronous two-dimensional correlation spectrum obtained in step 120 above have a strict spatial alignment relationship in the wavelength dimension. They carry the covariance information and response time sequence information of the same set of wavenumber pairs. Therefore, this application inputs the synchronous two-dimensional correlation spectrum and the asynchronous two-dimensional correlation spectrum as dual-channel correlation spectrum images in the wavelength dimension into the deep feature encoding module for joint encoding. The joint input enables the deep feature encoding module to automatically learn the optimal fusion weights of the two types of correlation structures through the first 1×1 convolutional layer, avoiding the limitations of manually preset fusion strategies.

[0029] The synchronous and asynchronous two-dimensional correlation spectra output by 2T2D-COS are large-size, high-dimensional correlation spectrum matrices, typically 510×510 pixels, far exceeding the input size of the deep feature encoding module (usually below 128×128). Therefore, for a 510×510×2 dual-channel correlation spectrum image, this embodiment designs an adapted deep feature encoding module structure: the deep feature encoding module uses five 3×3 convolutional layers with a stride of 2 for progressive downsampling, compressing the dual-channel correlation spectrum image to a high-dimensional feature tensor of 16×16×256. After flattening the high-dimensional feature tensor into a high-dimensional feature vector, it is mapped to a 20-dimensional latent space through a fully connected layer to obtain a low-dimensional deep feature vector of the agricultural product to be tested, forcing the deep feature encoding module to retain only the most discriminative structural features in the correlation spectrum.

[0030] In addition, considering that the spectral intensity of different agricultural products may fluctuate significantly due to their inherent characteristics, in order to reduce the interference caused by such fluctuations, the synchronous two-dimensional correlation spectrum and the asynchronous two-dimensional correlation spectrum are first subjected to independent Z-score normalization, and then the channel dimensions are spliced ​​to obtain the dual-channel correlation spectrum image input to the depth feature encoding module.

[0031] This step requires the use of a pre-trained deep feature encoding module, therefore it also includes methods for training the deep feature encoding module, including: (1) Obtain the sample spectral data of agricultural products. This is similar to step 110 and will not be repeated here.

[0032] When calculating the average spectral response value of each pixel in the region of interest for each agricultural product sample, the average spectral response value of each pixel in a circular region with a fixed radius centered on the centroid of the region of interest is calculated as the effective spectral data of the agricultural product to be tested, so as to ensure the consistency of the spectral extraction region.

[0033] (2) Construct multiple different combinations of spectral preprocessing methods, each combination of spectral preprocessing methods including two different spectral preprocessing methods.

[0034] Multiple different spectral preprocessing methods can be constructed by selecting one or more combinations from SG smoothing, Standard Normal Transform (SNV), Multivariate Scatter Correction (MSC), and First Derivative Transform (D1). Different pairwise combinations of these methods can then yield various spectral preprocessing method combinations. For example, in one instance, one spectral preprocessing method combination includes spectral preprocessing method 1 (SG smoothing) and spectral preprocessing method 2 (Multivariate Scatter Correction (MSC) + First Derivative Transform (D1)). Another spectral preprocessing method combination includes spectral preprocessing method 1 (SG smoothing + Standard Normal Transformation (SNV)) and spectral preprocessing method 2 (First Derivative Transform (D1)), and so on.

[0035] (3) Use two spectral preprocessing methods in one of the spectral preprocessing methods to preprocess a set of sample spectral data as reference spectrum and spectrum to be measured, and use the 2T2D-COS method to obtain synchronous two-dimensional correlation spectrum and asynchronous two-dimensional correlation spectrum and then stitch them together to form a dual-channel correlation spectrum image.

[0036] (4) Input the dual-channel correlation spectrum image of each agricultural product sample into the encoder of the convolutional autoencoder to perform feature encoding to obtain a low-dimensional deep feature vector, and use the decoder of the convolutional autoencoder to obtain the reconstructed dual-channel correlation spectrum image based on the low-dimensional deep feature vector.

[0037] Please refer to the network structure of the convolutional autoencoder. Figure 2 The encoder uses the same Figure 1 The network structure is the same as that of the medium-depth feature encoding module. The decoder uses a fully connected layer to first map the low-dimensional depth feature vector output by the encoder to a high-dimensional feature vector, and then converts it into a 16×16×256 three-dimensional feature tensor through a reconstruction operation. Then, it passes through five transposed convolutional layers for progressive upsampling and channel compression. Finally, it obtains a reconstructed two-channel correlation spectrum image with the same size as the two-channel correlation spectrum image through a 1×1 convolutional layer and a size reconstruction operation.

[0038] (5) The mean square error between the dual-channel correlation spectrum image of each agricultural product sample and the corresponding reconstructed dual-channel correlation spectrum image is used as the loss function to train the convolutional autoencoder.

[0039] Due to the scattering effect and overlapping absorption bands in the near-infrared spectra of agricultural metabolites, multiple combinations of different spectral preprocessing methods were constructed during training to provide diverse input bases for the reference and target spectra. During training, the Adam optimization algorithm was used to iteratively update the network parameters, with a training batch size of 8 and a maximum training epoch of 200. By minimizing the loss function, the convolutional autoencoder learned the latent feature structure in the two-dimensional correlated spectrum and output a low-dimensional deep feature vector.

[0040] (6) Finally, the encoder in the trained convolutional autoencoder is used as the trained deep feature encoding module.

[0041] Step 140: Input the low-dimensional deep feature vector into the trained nonlinear regression prediction model to obtain the metabolite detection results of the agricultural product to be tested. In one embodiment, the nonlinear regression prediction model adopts a random forest regression model.

[0042] In one application example, the agricultural product to be tested and the agricultural product sample are both corn kernels. The metabolite detection results of the agricultural product to be tested are lipid metabolites of corn kernels. The lipid metabolites of corn kernels include at least one of 9-linolenic acid, pinemanic acid, N-palmitoylethanolamine, N-stearoylethanolamine, N-oleoylethanolamine and propionic acid-related lipids.

[0043] The above descriptions are merely preferred embodiments of this application, and this application is not limited to the above embodiments. It is understood that other improvements and variations that can be directly derived or conceived by those skilled in the art without departing from the spirit and concept of this application should be considered to be included within the protection scope of this application.

Claims

1. A non-destructive detection method for agricultural product metabolites based on two-dimensional correlation spectroscopy, characterized in that, The non-destructive testing method for agricultural product metabolites includes: Obtain effective spectral data of the agricultural products to be tested; The effective spectral data were preprocessed using two different spectral preprocessing methods to serve as reference and target spectra, and synchronous two-dimensional correlation spectra and asynchronous two-dimensional correlation spectra were obtained using the 2T2D-COS method. After the synchronous two-dimensional correlation spectrum and the asynchronous two-dimensional correlation spectrum are concatenated into a dual-channel correlation spectrum image, the image is input into a trained deep feature encoding module for feature encoding to obtain a low-dimensional deep feature vector of the agricultural product to be tested. The low-dimensional deep feature vector is input into the trained nonlinear regression prediction model to obtain the metabolite detection results of the agricultural product to be tested.

2. The non-destructive testing method for agricultural product metabolites according to claim 1, characterized in that, The spectral preprocessing methods used for the effective spectral data are one or more combinations of SG smoothing, standard normal variable transformation (SNV), multivariate scattering correction (MSC), and first derivative transformation (D1).

3. The non-destructive testing method for agricultural product metabolites according to claim 1, characterized in that, The non-destructive testing method for agricultural product metabolites also includes: Obtain spectral data of agricultural product samples; Construct multiple different combinations of spectral preprocessing methods, each combination including two different spectral preprocessing methods; Two spectral preprocessing methods from one of the spectral preprocessing methods were used to preprocess a set of sample spectral data as reference spectrum and spectrum to be measured. Then, the synchronous two-dimensional correlation spectrum and asynchronous two-dimensional correlation spectrum were obtained by using the 2T2D-COS method and stitched together to form a two-channel correlation spectrum image. The dual-channel correlation spectrum image of each agricultural product sample is input into the encoder of the convolutional autoencoder to obtain a low-dimensional deep feature vector. The decoder of the convolutional autoencoder is then used to reconstruct the dual-channel correlation spectrum image based on the low-dimensional deep feature vector. The mean square error between the dual-channel correlation spectrum image of each agricultural product sample and the corresponding reconstructed dual-channel correlation spectrum image is used as the loss function to train the convolutional autoencoder, and the encoder in the trained convolutional autoencoder is used as the trained deep feature encoding module.

4. The non-destructive testing method for agricultural product metabolites according to claim 3, characterized in that, The encoder in the convolutional autoencoder uses a 1×1 convolutional layer to fuse and adjust the dimensionality of the input feature channels of the dual-channel correlation spectrum image. It uses multiple 3×3 convolutional layers to perform progressive downsampling and gradually increase the number of feature channels to extract local correlation features and high-order spatial features from the dual-channel correlation spectrum image to obtain a high-dimensional feature tensor. Each convolutional layer uses the SELU activation function. The high-dimensional feature tensor is flattened into a high-dimensional feature vector, and then the high-dimensional feature vector is mapped into a low-dimensional depth feature vector through a fully connected layer.

5. The non-destructive testing method for agricultural product metabolites according to claim 3, characterized in that, The dual-channel correlation spectrum image is 510×510×2. The deep feature encoding module uses five 3×3 convolutional layers with a stride of 2 to perform step-by-step downsampling, compressing the dual-channel correlation spectrum image to a high-dimensional feature tensor of 16×16×256. After flattening the high-dimensional feature tensor into a high-dimensional feature vector, it is mapped to a 20-dimensional latent space through a fully connected layer to obtain the low-dimensional deep feature vector of the agricultural product to be tested.

6. The non-destructive testing method for agricultural product metabolites according to claim 5, characterized in that, The decoder in the convolutional autoencoder uses a fully connected layer to map the low-dimensional depth feature vector output by the encoder to a high-dimensional feature vector. Then, it is converted into a 16×16×256 three-dimensional feature tensor through a reconstruction operation. The tensor is then passed through five transposed convolutional layers for progressive upsampling and channel compression. Finally, a 1×1 convolutional layer and a size reconstruction operation are used to obtain a reconstructed dual-channel correlation spectrum image with the same size as the dual-channel correlation spectrum image.

7. The non-destructive testing method for agricultural product metabolites according to claim 1, characterized in that, The process of stitching synchronous two-dimensional correlation spectra and asynchronous two-dimensional correlation spectra into a two-channel correlation spectrum image also includes: After performing independent Z-score normalization on the synchronous two-dimensional correlation spectrum and the asynchronous two-dimensional correlation spectrum, the two-channel correlation spectrum image is obtained by stitching them together along the channel dimension.

8. The non-destructive testing method for agricultural product metabolites according to claim 1, characterized in that, Obtaining effective spectral data for the agricultural products to be tested includes: After acquiring hyperspectral images of the area where the agricultural product to be tested is located and performing standardized correction processing, the agricultural product to be tested is segmented and the region of interest is extracted. The average value of the spectral response value of each pixel in the region of interest is calculated to obtain the effective spectral data.

9. The non-destructive testing method for agricultural product metabolites according to claim 1, characterized in that, The nonlinear regression prediction model uses a random forest regression model.

10. The non-destructive testing method for agricultural product metabolites according to claim 1, characterized in that, The agricultural product to be tested is corn kernels, and the metabolite detection results of the agricultural product to be tested are lipid metabolites of corn kernels. The lipid metabolites of corn kernels include at least one of 9-linolenic acid, pinemanic acid, N-palmitoylethanolamine, N-stearoylethanolamine, N-oleoylethanolamine and propionic acid-related lipids.