Rapid crop component analysis method based on near infrared spectrum and radial basis function neural network modeling

Abstract

The invention discloses a rapid crop component analysis method based on near infrared spectrum and radial basis function neural network modeling, which comprises the following steps: collecting near infrared spectrum data of a cereal sample, and measuring the content true values of main components such as protein, crude fiber, starch and saccharides in the sample; secondly, filtering and denoising, baseline correction, normalization and other preprocessing and feature extraction are carried out on the spectral data; thirdly, a radial basis function neural network model is constructed, the network center is determined through K-means clustering, the width is calculated through an empirical formula, the weight of an output layer is solved through a least square method, and model parameters are optimized in combination with cross validation; and finally, inputting the preprocessed spectral data of the to-be-detected crop grains into the optimized model to realize component content evaluation. The method can achieve excellent effects in component estimation of different crop grains such as corn and wheat, has the advantages of high accuracy, efficient modeling and high practicability, and can be widely applied to component detection of a crop grain production and processing whole chain.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of crop grain quality detection and spectral analysis, and in particular to a crop component rapid analysis method based on near-infrared spectroscopy and a radial basis neural network modeling. BACKGROUND

[0002] The crop grain component content is a core index for evaluating its nutritional value, processing suitability and market value. Traditional detection methods such as Kjeldahl nitrogen determination (protein) and acid-base washing method (crude fiber) have high detection precision, but have defects such as complicated operation, long time (several hours for single detection), large consumption of chemical reagents, and strong sample destructiveness, which cannot meet the rapid detection needs of crop production, processing and circulation links.

[0003] The near-infrared spectroscopy technology has become one of the mainstream technologies for crop component detection due to its advantages of "fast, non-destructive and environmentally friendly". Its principle is to detect the characteristic absorption of C-H, O-H, N-H and other hydrogen-containing groups in grains to grain components. However, the existing modeling methods based on near-infrared spectroscopy still have three major bottlenecks: 1. Limitations of linear modeling methods: partial least squares (PLS) is the most widely used modeling method at present, but it is a linear model and is difficult to capture the complex nonlinear relationship between spectral data and grain components due to "spectral peak overlap" and "matrix effect". When the sample component distribution range is wide, the estimation accuracy significantly decreases (R 2 is usually less than 0.85).

[0004] 2. Complexity of traditional neural networks: although deep neural networks and other methods have nonlinear fitting capabilities, they rely on multi-layer network stacking (usually more than 5 layers), a large number of training samples (more than thousands of samples) and high-performance computing resources, and are prone to overfitting, slow convergence (iteration number often exceeds 500) and other problems, with high modeling cost and poor practicability.

[0005] 3. "Dimension disaster" of high-dimensional spectral data: near-infrared spectroscopy usually contains hundreds to thousands of wavelength points (such as the wave number range of 4000-10000 cm -1 , the spectral resolution is 4 cm -1 , and there are about 1500 dimensions), although high-dimensional data more completely characterizes the characteristics of the data, but at the same time, it not only increases the modeling complexity, but also causes the model stability to decrease due to the interference of redundant information.

[0006] Therefore, it is of great significance to develop a modeling method that can effectively handle nonlinear relationships, simplify network structure, and solve the problem of high-dimensional data redundancy, in order to improve the near-infrared detection precision and practicability of crop components. SUMMARY

[0007] The main purpose of the present application is to provide a crop component rapid analysis method based on near-infrared spectroscopy and radial basis function neural network modeling, aiming to solve the above technical problems.

[0008] To achieve the above purpose, the present application provides a crop component rapid analysis method based on near-infrared spectroscopy and radial basis function neural network modeling, which comprises: (1) Sample data acquisition and construction: Collecting multiple crop samples covering different varieties, production places or growth conditions, collecting near-infrared spectrum data of each crop grain sample, and determining the true value of the content of protein, crude fiber, starch and sugar in each crop grain sample by standard chemical analysis method, and forming a sample data set corresponding to the near-infrared spectrum data and the true value of the component content; (2) Spectral data preprocessing: The near-infrared spectrum data in the sample data set is sequentially subjected to filtering denoising, baseline correction and normalization processing; (3) Spectral data feature extraction: The key features of the spectrum are extracted by principal component analysis to reduce the data dimension and remove redundant information; (4) RBF neural network model construction: A three-layer radial basis function neural network modeling architecture containing input layer, hidden layer and output layer is constructed, wherein the number of input layer nodes is consistent with the dimension of the spectral features extracted in step (3), and the output layer nodes are protein, crude fiber, starch and sugar respectively; Selecting an appropriate radial basis function as the hidden layer activation function, using K-means clustering algorithm to determine the hidden layer node center, and calculating the hidden layer node width by empirical formula, and finally solving the output layer weight by least square method; (5) Model training and optimization: The spectral features extracted in step (3) and the corresponding component content true value are divided into training set, validation set and test set, the RBF neural network model is trained using the training set, the model performance is evaluated by the validation set and the parameters are adjusted, the model parameters are further optimized by cross-validation, and the optimized RBF neural network model is obtained; (6) Estimation of the component of the crop to be measured: Repeating the spectrum collection in step (1), the spectrum data preprocessing in step (2) and the spectrum feature extraction in step (3) for the crop grain sample to be measured, inputting the processed spectrum feature data into the optimized RBF neural network model in step (5), and outputting the estimated value of the content of protein, crude fiber, starch and sugar in the crop grain sample to be measured.

[0009] In an embodiment, in step (1), the step of collecting near-infrared spectrum data of each crop grain sample is: The spectral collection wave number range is set to 4000-10000 cm -1 , and the spectral resolution is 4 cm -1Spectrum collection is performed on each sample of crop grains, and the near-infrared spectrum data of the sample is obtained by averaging 32 scanning times.

[0010] In an embodiment, the step of determining the true values of the contents of protein, crude fiber, starch and sugar in each sample of crop grains by using standard chemical analysis methods in step (1) is specifically as follows: The content of protein in the crop grains is determined by using the Kjeldahl method; The content of crude fiber in the crop grains is determined by using the acid-base washing method; The content of amylopectin in the crop grains is determined by using the iodine colorimetric method; The content of sugar in the crop grains is determined by using the high-performance liquid chromatography method; The collected near-infrared spectrum data is associated with the corresponding component data to construct a sample data set in which the near-infrared spectrum data and the true values of the component contents correspond to each other.

[0011] In an embodiment, the step of spectrum data preprocessing in step (2) is specifically as follows: The filter denoising uses a Savitzky-Golay filter algorithm or a wavelet denoising algorithm, the filter window size of the Savitzky-Golay filter algorithm is 5, and the polynomial order is 2; The baseline correction uses a linear baseline correction method or a polynomial baseline correction method; The normalization processing unifies the spectrum data to the numerical range of [0, 1].

[0012] In an embodiment, in step (3), the number of principal components of the principal component analysis method is 95%.

[0013] In an embodiment, in step (4), the number of hidden layer nodes of the RBF neural network model is optimized and determined in the range of 10-100 through experiments.

[0014] In an embodiment, in step (4), the radial basis function uses a Gaussian function, and the expression is as follows: ; wherein x is a preprocessed near-infrared spectrum data input vector, c i is the center of the i-th hidden layer node, σ i is the width of the i-th hidden layer node, and ||x-c i || represents the Euclidean distance.

[0015] In an embodiment, in step (4), the empirical formula expression is as follows: ; wherein d maxThe maximum distance between all cluster centers, K is the number of hidden layer nodes.

[0016] In an embodiment, the step (5) is performed by using the following formula: The division ratio of the training set, the validation set and the test set is 7:2:1 or 6:2:2; The model performance evaluation index includes accuracy, root mean square error and determination coefficient, and the model overfitting is prevented by adjusting the number of hidden layer nodes, the center, the width and the weight parameter; The cross-validation is 5-fold cross-validation.

[0017] In the technical scheme of the present application, the RBF neural network model is constructed to realize the non-destructive component estimation of crop kernels by near-infrared spectroscopy. Compared with the prior art, the following significant advantages are obtained: (1) Strong non-linear fitting ability and high estimation accuracy: the RBF neural network can effectively capture the non-linear relationship between the spectrum and the component through the local response characteristic of the Gaussian kernel function, and the model accuracy is greatly improved after combining with PCA for de-redundancy.

[0018] (2) Simple network structure and high modeling efficiency: the RBF neural network only needs a three-layer network architecture, does not need to be stacked in multiple layers, and the output layer weight is directly solved by the least square method, so the training time is significantly shortened compared with the traditional deep neural network.

[0019] (3) PCA dimensionality reduction and de-redundancy, good model stability: the PCA reduces the dimension of the near-infrared spectrum data from hundreds to 10-15, retains the main features, relieves the "dimension disaster", and effectively reduces the noise and redundant information interference during the training of the RBF neural network model.

[0020] (4) Convenient operation and strong practicability: using the optimized RBF neural network model combined with the non-destructive characteristics of near-infrared spectroscopy, the entire detection process does not need chemical reagents, the single detection time is ≤5 minutes, and it can be directly applied to crop purchase, processing workshop and other on-site scenes to realize "rapid screening and real-time monitoring". BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical scheme in the embodiments of the present application or the prior art, the drawings needed in the following embodiment or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings from the structures shown in the drawings without creating any creative labor.

[0022] Figure 1 It is the overall flowchart of the method of the present application; Figure 2 It is the radial basis neural network architecture diagram of the present application.

[0023] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0025] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0026] Furthermore, in this invention, descriptions involving "first," "second," etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0027] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0028] This invention provides a rapid method for crop component analysis based on near-infrared spectroscopy and radial basis function neural network modeling.

[0029] The rapid crop component analysis method based on near-infrared spectroscopy and radial basis function neural network modeling provided in this invention includes: Step (1) Sample data collection and construction To ensure the model's universality and robustness, a widely representative sample dataset needs to be constructed. Specific steps include: Step (1a), Sample selection: Collect crop grain samples covering different varieties, origins, growth conditions, and storage periods, with a total sample size of no less than 200 (preferably more than 400) to cover a reasonable range of fluctuations in component content.

[0030] Step (1b), Spectral Acquisition: Spectral data were acquired using a high-precision near-infrared spectrometer, with the spectrometer parameters set as follows: spectral acquisition wavenumber range of 4000-10000 cm⁻¹. -1 (Covering the characteristic absorption bands of hydrogen-containing groups), spectral resolution 4cm -1 The instrument was scanned 32 times and the average value was taken to ensure that the instrument parameters (such as temperature, optical path length, integration time) were completely consistent during the acquisition of all samples, thus avoiding errors introduced by instrument drift. The spectrum of each sample was collected 3 times and the average value was taken as the final spectral data of that sample.

[0031] Step (1c), determination of true composition values: The true values ​​of the component content of each sample were determined using chemical analysis methods recommended by national or industry standards: protein content was determined by the Kjeldahl method, crude fiber content was determined by the acid-base washing method, amylopectin content was determined by the iodine colorimetric method, and carbohydrate content was determined by high performance liquid chromatography.

[0032] Step (1d), Dataset Construction: Each sample is assigned a unique number, and the spectral data is associated with the corresponding true values ​​of the four components one by one according to the number, forming a structured sample dataset of "sample number-spectral data-component true value", which is stored in CSV or MAT format for later use.

[0033] Step (2) Spectral data preprocessing The purpose of preprocessing is to eliminate interference such as spectral noise and baseline drift, while retaining effective component information. The specific process is as follows: Step (2a), Filtering and denoising: Savitzky-Golay filtering (window size 5-11 points, polynomial order 2-3) is used to smooth the spectrum by fitting local polynomials; if the noise contains impulses or complex interference, wavelet denoising is added (using db4 or sym5 wavelet basis, decomposition level 3-5, soft thresholding).

[0034] Step (2b), Baseline Correction: Baseline correction is used to eliminate baseline shifts caused by instrument light source fluctuations and sample scattering. Linear baseline correction is suitable for scenarios where the baseline drifts linearly; it involves fitting a straight line to wavelengths with no characteristic absorption at the two ends of the spectrum and subtracting the baseline to obtain the corrected spectrum. Polynomial baseline correction is suitable for scenarios where the baseline drifts nonlinearly; it involves fitting a 3rd-5th order polynomial to the baseline and then subtracting the corrected spectrum. Apply the baseline correction steps according to the specific circumstances.

[0035] Step (2c), Normalization: The corrected spectral data is mapped to the [0,1] range to eliminate spectral intensity fluctuations caused by differences in sample concentration and particle size. The calculation formula is as follows:

[0036] Where x is the original corrected spectral value, x min x max These are the minimum and maximum values ​​of the full-band spectrum, respectively.

[0037] Step (3) Spectral data feature extraction Principal component analysis (PCA) is used to extract key spectral features and address the problem of redundancy in high-dimensional data. Specific operations include: Step (3a) Data centralization The preprocessed spectral matrix (with dimensions of "number of samples × number of wavelengths") is centered by subtracting the mean value of each wavelength from its spectral value, thus eliminating the influence of dimensions.

[0038] Step (3b): Principal component solution The covariance matrix of the spectral matrix is ​​calculated by singular value decomposition (SVD) to obtain eigenvalues ​​and eigenvectors. The eigenvectors are then sorted in descending order of eigenvalues ​​to form the principal components (PCs).

[0039] Step (3c) Principal component screening The number of principal components is selected based on the principle that the cumulative contribution rate is not less than a threshold (such as 95%). For example, after PCA processing of the corn kernel spectrum with 1501 wavelength points, the first 10-15 principal components can cover more than 95% of the spectral information, reducing the data dimension from 1501 to 10-15, and significantly reducing the modeling complexity.

[0040] Step (3d), Feature Matrix Construction The selected principal components are arranged in the order of the samples to form a spectral feature matrix of "number of samples × number of principal components", which serves as the input data for subsequent RBF neural network modeling or estimation.

[0041] Step (4) Construction of RBF Neural Network Model A three-layer RBF neural network architecture is constructed to achieve a non-linear mapping from the feature matrix to the component content, specifically including: Step (4a), Network Structure Design: It consists of an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is the same as the number of principal components selected in step (3c) (e.g., 12 principal components, then 12 nodes in the input layer), and it is responsible for receiving spectral feature data. The number of nodes in the hidden layer is determined through experimental optimization (range approximately 10-100, preferably 40-60), and local nonlinear mapping is achieved through Gaussian radial basis functions. The number of nodes in the output layer is the same as the number of component categories determined in step (1c), corresponding to the estimated content values ​​of protein, crude fiber, starch, and sugar, respectively.

[0042] Step (4b), Activation Function Selection: The hidden layer uses a Gaussian radial basis function, expressed as:

[0043] Where x is the preprocessed near-infrared spectral data input vector, c i Let σ be the center of the i-th hidden layer node. i Let xc be the width of the i-th hidden layer node. i ∥ represents the Euclidean distance. This function has a local response characteristic of "near activation and far suppression," which can accurately capture nonlinear relationships.

[0044] Step (4c), core parameter determination: The core parameters of the RBF neural network model include the node center c. i Node width σ i Output layer weights w ij In step (3d), the K-means clustering algorithm is used on the spectral feature matrix. The number of clusters is equal to the number of hidden layer nodes, and the cluster centers are c. i This ensures that the center represents the sample feature distribution; the node width is calculated using an empirical formula. , where d max Let K be the maximum distance between all cluster centers, and K be the number of hidden layer nodes, ensuring that the effective range of the RBF function matches the feature distribution. The least squares method is used to minimize the sum of squared errors between the network output and the true values ​​of the components. The calculation formula is as follows: ; Where Φ is the hidden layer output matrix and t is the component truth vector, this avoids the time-consuming problem of iterative optimization in traditional neural networks.

[0045] Step (5), Model Training and Optimization Improving model performance through dataset partitioning, parameter tuning, and cross-validation involves the following steps: Step (5a), Dataset Partitioning: The structured sample dataset described in step (1d) is divided into a training set, a validation set, and a test set in a ratio (7:2:1 or 6:2:2) to be used for model parameter learning, monitoring the training process to prevent overfitting, and evaluating the model's generalization ability, respectively. The spectral data in the structured sample dataset described in step (1d) is replaced with the spectral feature matrix data from step (3d).

[0046] Step (5b), Model Training: The training set data is input into the RBF neural network model constructed in step (4), and the output error is calculated through forward propagation. The weights are solved once by the least squares method, without the need for backpropagation iteration. The training convergence speed is significantly improved compared with the traditional BP neural network.

[0047] Step (5c), parameter optimization: The model performance was evaluated using the validation set data, and model parameters were adjusted. This included selecting the number of hidden layer nodes that minimized the root mean square error (RMSE) on the validation set; excessive principal components could lead to overfitting, so the cumulative contribution rate threshold could be appropriately reduced (e.g., gradually decreasing from 95% to 90%); if the model bias was large, K-means clustering could be re-executed.

[0048] Step (5d), Cross-validation: Further optimization was achieved by using 5-fold cross-validation, which involves dividing the training set into 5 parts, using 4 parts for training and 1 part for validation each time, repeating this process 5 times, and then taking the average performance index to ensure model stability.

[0049] Step (6) Estimation of the composition of the grain to be tested Composition estimation is achieved by using the principle of "same process and same model prediction" for the samples to be tested. Step (6a), Sample preparation: Repeat the spectral acquisition in step (1) (with the same instrument parameters), the spectral data preprocessing in step (2) (with the same denoising, correction, and normalization methods), and the spectral data feature extraction in step (3) (using the PCA parameters of the training set) for the crop grain samples to be tested. Step (6a), Component Prediction: The extracted spectral features to be tested are input into the optimized RBF neural network model in step (5). The model output is the estimated content of protein, crude fiber, starch and sugar in the sample to be tested. The time taken for a single prediction is ≤1 second.

[0050] This invention leverages the local response characteristics and powerful nonlinear approximation capabilities of radial basis functions to effectively address the shortcomings of traditional partial least squares (PLS) nonlinear fitting, and the low training efficiency and overfitting issues caused by the reliance on multi-layer structures in traditional neural networks. This improves the learning efficiency of network parameters and the stability of the model. Experimental results demonstrate that this method achieves excellent results in component estimation for various crops such as corn and wheat, exhibiting advantages such as high accuracy, efficient modeling, and strong practicality. It can be widely applied to quality inspection throughout the entire crop production and processing chain.

[0051] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. All equivalent structural transformations made under the concept of the present invention using the contents of the present invention specification and drawings, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A rapid crop component analysis method based on near-infrared spectroscopy and radial basis function neural network modeling, characterized in that, The rapid crop component analysis method based on near-infrared spectroscopy and radial basis neural network modeling includes: (1) Sample data collection and construction: Collect a variety of crop grain samples covering different varieties, origins or growth conditions, collect near-infrared spectral data of each crop grain sample, and use standard chemical analysis methods to determine the true values ​​of protein, crude fiber, starch and sugar content in each crop grain sample, and form a sample dataset with a one-to-one correspondence between near-infrared spectral data and the true values ​​of component content. (2) Spectral data preprocessing: The near-infrared spectral data in the sample dataset are sequentially filtered for noise reduction, baseline correction and normalization; (3) Spectral data feature extraction: Key spectral features are extracted through principal component analysis to reduce data dimensionality and remove redundant information; (4) RBF neural network model construction: Construct a three-layer radial basis neural network modeling architecture containing an input layer, a hidden layer and an output layer. The number of nodes in the input layer is consistent with the dimension of the spectral features extracted in step (3), and the nodes in the output layer are protein, crude fiber, starch and sugar, respectively. We select the appropriate radial basis function as the activation function of the hidden layer, use the K-means clustering algorithm to determine the center of the hidden layer node, calculate the width of the hidden layer node by empirical formula, and finally use the least squares method to solve the weight of the output layer. (5) Model training and optimization: The spectral features and corresponding true values ​​of component content extracted in step (3) are divided into training set, validation set and test set. The RBF neural network model is trained using the training set, the model performance is evaluated and the parameters are adjusted using the validation set, and the model parameters are further optimized using cross-validation to obtain the optimized RBF neural network model. (6) Estimation of the components of the crop to be tested: Repeat the spectral acquisition in step (1), the spectral data preprocessing in step (2), and the spectral feature extraction in step (3) of the crop grain sample to be tested. Input the processed spectral feature data into the optimized RBF neural network model in step (5) and output the estimated values ​​of the content of protein, crude fiber, starch, and sugar in the crop sample to be tested.

2. The rapid crop component analysis method based on near-infrared spectroscopy and radial basis function neural network modeling according to claim 1, characterized in that, In step (1), the step of collecting near-infrared spectral data of each crop grain sample is as follows: The spectral acquisition wavenumber range is set to 4000-10000 cm⁻¹. -1 Spectral resolution 4cm -1 Spectral data of each crop sample were acquired by scanning 32 times and the average value was taken to obtain the near-infrared spectral data of the sample.

3. The rapid crop component analysis method based on near-infrared spectroscopy and radial basis function neural network modeling according to claim 2, characterized in that, The specific steps in step (1) for determining the true values ​​of protein, crude fiber, starch, and sugar content in each crop grain sample using standard chemical analysis methods are as follows: The protein content of crop grains was determined using the Kjeldahl method. The crude fiber content of crop grains was determined using an acid-base washing method. The amylopectin content in crop grains was determined using the iodine colorimetric method. High performance liquid chromatography was used to determine the sugar content of crop grains; The collected near-infrared spectral data are correlated with the corresponding component data to construct a sample dataset that establishes a one-to-one correspondence between near-infrared spectral data and the true values ​​of component content.

4. The rapid crop component analysis method based on near-infrared spectroscopy and radial basis function neural network modeling according to claim 1, characterized in that, The specific steps of spectral data preprocessing in step (2) are as follows: The filtering and denoising uses either the Savitzky-Golay filtering algorithm or the wavelet denoising algorithm. The Savitzky-Golay filtering algorithm has a filtering window size of 5 and a polynomial order of 2. The baseline correction employs either a linear baseline correction method or a polynomial baseline correction method. The normalization process unifies the spectral data to the numerical range of [0,1].

5. The rapid crop component analysis method based on near-infrared spectroscopy and radial basis function neural network modeling according to claim 1, characterized in that, In step (3), the threshold for the contribution rate of principal components in the principal component analysis method is 95%.

6. The rapid crop component analysis method based on near-infrared spectroscopy and radial basis function neural network modeling according to claim 1, characterized in that, In step (4), the number of hidden layer nodes based on the RBF neural network model is determined experimentally within the range of 10-100.

7. The rapid crop component analysis method based on near-infrared spectroscopy and radial basis function neural network modeling according to claim 1, characterized in that, In step (4), the radial basis function is a Gaussian function, expressed as: ； Where x is the preprocessed near-infrared spectral data input vector, c i Let σ be the center of the i-th hidden layer node. i Let xc be the width of the i-th hidden layer node. i ∥ represents Euclidean distance.

8. The rapid crop component analysis method based on near-infrared spectroscopy and radial basis function neural network modeling according to claim 1, characterized in that, In step (4), the empirical formula expression is: ； Where d max K represents the maximum distance between all cluster centers, and K is the number of hidden layer nodes.

9. The rapid crop component analysis method based on near-infrared spectroscopy and radial basis function neural network modeling according to claim 1, characterized in that, In step (5), The ratio of the training set, validation set, and test set is 7:2:1 or 6:2:2; The model performance evaluation metrics include accuracy, root mean square error, and coefficient of determination. Overfitting of the model is prevented by adjusting the number of hidden layer nodes, center, width, and weight parameters. The cross-validation is 5-fold cross-validation.

Images