A spectral feature selection method combined with empirical mode decomposition

Abstract

The application discloses a spectrum feature selection method combined with empirical mode decomposition and relates to the technical field of hyperspectral data inversion. The spectrum feature selection method combined with empirical mode decomposition comprises the following steps: data acquisition, data preprocessing, empirical mode decomposition, correlation analysis and estimation model construction. The application enhances the physical interpretability of feature selection through physical decomposition of empirical mode decomposition, improves the adaptability to complex spectrum features by using multi-scale analysis, and screens sensitive features by combining with correlation analysis of Pearson correlation coefficients, so that an efficient and robust new paradigm of feature selection is constructed, and the convenience, efficiency and robustness of the spectrum feature selection method are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of hyperspectral data inversion technology, specifically to a method for selecting spectral features that combines empirical mode decomposition. Background Technology

[0002] Hyperspectral data inversion technology has developed rapidly and has become a hot topic in the field of remote sensing. However, the matrix, composition, and structure of soil are very complex. Therefore, the hyperspectral reflectance of soil is not the result of a single factor, but a comprehensive reflection of the complex interactions between soil organic matter (SOM) and other soil components. Furthermore, hyperspectral data typically contains hundreds of narrow bands, which can provide rich soil information, but the high redundancy between adjacent bands can lead to difficulties in data storage and transmission in practical applications, and the data processing process is quite complex. Therefore, how to effectively select sensitive bands closely related to SOM content from a large amount of hyperspectral data has become one of the key issues in the construction of SOM inversion models. To address this issue, many researchers have proposed different data preprocessing methods to select and extract the most representative bands from hyperspectral data. These bands can better reflect the changes in SOM, thereby improving the accuracy and efficiency of the inversion model.

[0003] Existing feature selection methods can be categorized into three types: wrapper-style, embedded, and filtering. However, all of them have significant limitations in hyperspectral scenarios. Wrapper-style methods (such as Recursive Feature Elimination (RFE) and genetic algorithms) evaluate the performance of feature combinations by iteratively training the model. While they can capture feature interactions, their computational complexity increases exponentially with dimensionality. For example, some literature shows that when performing feature selection, wrapper-style methods need to traverse all possible feature subsets, causing the search space to expand exponentially with the number of features, resulting in huge computational overhead. To improve practical feasibility, heuristic algorithms (such as greedy strategies and particle swarm optimization) are usually introduced to narrow the search range and reduce computational complexity. However, these strategies are essentially local sampling of the solution space. Although they significantly reduce computational costs, they greatly increase the risk of getting trapped in local optima due to limited search paths. Embedded methods (such as Lasso and Random Forest) select features through built-in model regularization. Although they are more computationally efficient than wrapper-style methods, their selection results are highly dependent on model assumptions and it is difficult to interpret the physical meaning of the features. In contrast, filtering methods (such as Pearson correlation coefficient (PCC) and mutual information) have become the mainstream choice for hyperspectral analysis due to their independence from the prediction model and computational efficiency. PCC is a classic filtering method for measuring linear relationships between variables and is widely used in hyperspectral remote sensing data processing. However, traditional PCC only measures linear correlation, cannot characterize complex nonlinear relationships, and ignores synergistic effects between features.

[0004] Based on the above, this invention proposes a simple, efficient, and robust method for selecting spectral features by combining empirical mode decomposition. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention proposes a simple, efficient, and robust spectral feature selection method that combines empirical mode decomposition (EMD). This method effectively enhances the physical interpretability of feature selection through the physical decomposition of EMD. It not only improves the adaptability to complex spectral features by utilizing multi-scale analysis, but also strengthens the Pearson correlation coefficient algorithm to enhance the screening ability of feature interactions while maintaining computational efficiency, thereby constructing a new paradigm for simple, efficient, and robust feature selection.

[0006] A spectral feature selection method combining empirical mode decomposition includes the following steps: Data acquisition involves obtaining soil hyperspectral data and corresponding measured data of soil organic matter content through sensors and elemental analyzers. Data preprocessing involves wavelet denoising of the acquired soil hyperspectral data to obtain preprocessed soil hyperspectral data, and missing and outlier processing of the acquired measured soil organic matter content inversion data to obtain preprocessed measured soil organic matter content inversion data. Empirical mode decomposition (EMD) is used to decompose the measured soil organic matter content data at the sample scale of the preprocessed sample-scale inversion data and the soil hyperspectral data at the corresponding measured points in the preprocessed soil hyperspectral data, and extract the intrinsic mode functions containing different frequency components. Correlation analysis is performed based on the Pearson correlation coefficient to analyze the correlation between the intrinsic mode functions of different frequency components, and bands with correlations higher than a preset threshold are selected as sensitive features. The estimation model is constructed by normalizing the acquired sensitive features to obtain normalized sensitive features. The normalized sensitive features are then input into the support vector machine model, and the data is mapped to a high-dimensional space using a kernel function. Finally, the soil organic matter content estimation model is constructed by optimizing the penalty coefficient and kernel function parameters.

[0007] As a preferred aspect of the invention, the specific steps for performing wavelet denoising on the acquired soil hyperspectral data to obtain preprocessed soil hyperspectral data are as follows: Based on the characteristics of soil hyperspectral data, a suitable wavelet basis function was selected, and the optimal number of decomposition layers was determined through experiments. Multi-scale wavelet decomposition of soil hyperspectral data was performed using selected wavelet basis functions to obtain wavelet coefficients at different scales, including low-frequency wavelet coefficients and high-frequency wavelet coefficients. The high-frequency wavelet coefficients obtained by decomposition are thresholded and noise is removed by the soft thresholding method to obtain the thresholded high-frequency wavelet coefficients. Wavelet reconstruction is performed based on the low-frequency wavelet coefficients and the high-frequency wavelet coefficients after thresholding to obtain the preprocessed soil hyperspectral data.

[0008] As a preferred aspect of the invention, the specific steps for thresholding and removing noise from the high-frequency wavelet coefficients obtained by the soft thresholding method to obtain the thresholded high-frequency wavelet coefficients are as follows: The noise standard deviation of the high-frequency wavelet coefficients is calculated, and the discrimination threshold is obtained by calculating the discrimination threshold using the following formula:

[0009] in Indicates the discrimination threshold. Indicates the standard deviation of noise. Indicates the signal length of soil hyperspectral data; A soft threshold formula is set based on the discrimination threshold, and the soft threshold formula is applied to each high-frequency wavelet coefficient to obtain the high-frequency wavelet coefficients after thresholding. The specific soft threshold formula is as follows:

[0010] in This represents the high-frequency wavelet coefficients after thresholding. This represents the high-frequency wavelet coefficients before thresholding.

[0011] As a preferred aspect of the invention, the specific steps for processing the acquired measured data of inverted soil organic matter content to remove missing and outlier values, thereby obtaining preprocessed measured data of inverted soil organic matter content, are as follows: Missing value handling involves processing the empty or missing values ​​in the measured data of soil organic matter content inversion. Linear interpolation or polynomial interpolation methods are used to fill the missing values, or the mean or median of the data is used directly to fill the missing values. Outlier handling involves identifying outliers in the measured soil organic matter content data that do not conform to the expected pattern using the Z-Score or IQR methods. Outliers are then removed and replaced with the mean or median of the data, or interpolation methods are used to repair them.

[0012] As a preferred aspect of the invention, the specific steps of decomposing the preprocessed sample-scale inverted soil organic matter content measured data and the preprocessed soil hyperspectral data corresponding to the measured points using the empirical mode decomposition method to extract intrinsic mode functions containing different frequency components are as follows: S1: Data extraction. The preprocessed measured soil organic matter content data and preprocessed soil hyperspectral data were interpolated and expanded using the Kriging interpolation method, and the preprocessed sample-scale measured soil organic matter content data and the corresponding sample-scale soil hyperspectral data of the measured points were extracted from them. S2: Initialization: Use the preprocessed sample-scale inverted soil organic matter content measured data and the corresponding measured point soil hyperspectral data in the preprocessed soil hyperspectral data as input signals and initialize the counter. S3: Envelope generation, identifying local maxima and minima in the input signal, using cubic spline interpolation to connect maxima to form the upper envelope and connect minima to form the lower envelope; S4: Residual extraction and conditional discrimination. Calculate the average value of the upper and lower envelopes, and subtract the average value of the upper and lower envelopes from the input signal to obtain the residual. Check whether the residuals meet the conditions of IMF, that is, whether the number of extreme points and zero crossover points are equal or differ by 1, and whether the average value of the upper and lower envelopes is zero. S5: Iterative update and termination. If the IMF condition is met, the residual is used as the new input signal, and steps S3 and S4 are repeated. If the IMF condition is not met, the residual is used as the new input signal to continue decomposition. When the residual signal becomes a monotonic function, the decomposition stops and multiple eigenmode functions containing different frequency components are obtained.

[0013] As a preferred aspect of the invention, the specific steps of analyzing the correlation between eigenmode functions of different frequency components based on the Pearson correlation coefficient and selecting bands with correlation higher than a preset threshold as sensitive features are as follows: For each intrinsic mode function corresponding to the soil hyperspectral data and each intrinsic mode function corresponding to the inverted measured soil organic matter content data, the Pearson correlation coefficient between them is calculated. The specific calculation formula is as follows:

[0014] in This represents the Pearson correlation coefficient. and These represent the first and second data points corresponding to the soil hyperspectral data and the measured data of soil organic matter content retrieved from the inversion database. One eigenmode function and These represent the average values ​​of the intrinsic mode functions corresponding to the soil hyperspectral data and the inverted measured soil organic matter content data, respectively. This represents the total number of intrinsic mode functions; Obtain all calculated Pearson correlation coefficients and plot their distribution histograms; dynamically determine the preset threshold based on the distribution of the Pearson correlation coefficients. Spectral bands with Pearson correlation coefficients higher than a preset threshold were selected; these spectral bands are the characteristic bands that are sensitive to changes in soil organic matter content. The selected characteristic bands are deduplicated, and only one sensitive characteristic band is retained for each repeated band, so as to obtain the sensitive characteristics in the measured soil organic matter content inversion data at the sample scale.

[0015] As a preferred aspect of the invention, the specific steps for normalizing the acquired sensitive features to obtain normalized sensitive features are as follows: Iterate through all the acquired sensitive features and find the minimum and maximum values ​​for each feature dimension; Based on the minimum and maximum values ​​of each feature dimension, and using a normalization formula, all obtained sensitive features are normalized to [value missing]. Between these, the normalized sensitive features are obtained.

[0016] As a preferred aspect of the invention, the specific steps of inputting the normalized sensitive features into the support vector machine model, mapping the data to a high-dimensional space using a kernel function, and constructing a soil organic matter content estimation model by optimizing the penalty coefficient and kernel function parameters are as follows: The normalized sensitive features are presented in matrix form, where each row of the matrix represents a sample and each column represents a feature. Initially, radial basis function kernels and linear kernels were selected as kernel functions. The grid search method was used to optimize the penalty coefficients and kernel function parameters within a certain range to find the parameter combination that minimizes the verification error. The model performance was evaluated using cross-validation, and the mean squared error and coefficient of determination were calculated. A scatter plot of the predicted values ​​and the true values ​​was drawn to evaluate the fit. By comparing the differences in model performance between radial basis function kernels and linear kernels, a better kernel function was selected to construct the final soil organic matter content estimation model.

[0017] The present invention has the following advantages: 1. This invention enhances the physical interpretability of feature selection through physical decomposition of empirical mode decomposition, improves the adaptability to complex spectral features through multi-scale analysis, and screens sensitive features by combining correlation analysis of Pearson correlation coefficient, thereby constructing a new paradigm of efficient and robust feature selection, which can significantly improve the convenience, efficiency and robustness of this spectral feature selection method.

[0018] 2. This invention uses wavelet denoising to denoise the acquired soil hyperspectral data, resulting in preprocessed soil hyperspectral data. This effectively reduces the noise level in the soil hyperspectral data, improves the signal-to-noise ratio, and enhances the clarity and identifiability of spectral features. This facilitates more accurate extraction of sensitive bands related to soil organic matter content and improves the accuracy and stability of subsequent feature selection and estimation models. As a result, it provides a more reliable data foundation for the hyperspectral inversion of soil organic matter content and enhances the robustness of this spectral feature selection method.

[0019] 3. This invention uses a soft thresholding method to threshold the high-frequency wavelet coefficients obtained from decomposition and remove noise, resulting in thresholded high-frequency wavelet coefficients. This not only removes noise while preserving important features and details in the signal and avoiding feature loss due to excessive smoothing, but also makes the preserved signal features smoother and more stable through scaling, thereby improving the accuracy and reliability of subsequent feature selection and estimation models and enhancing the robustness of this spectral feature selection method. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a spectral feature selection method combining empirical mode decomposition used in an embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of this invention.

[0022] Example 1: A spectral feature selection method combining empirical mode decomposition, such as... Figure 1 As shown, it includes the following steps: Data acquisition involves obtaining soil hyperspectral data and corresponding measured data of soil organic matter content through sensors and elemental analyzers. Data preprocessing involves wavelet denoising of the acquired soil hyperspectral data to obtain preprocessed soil hyperspectral data, and missing and outlier processing of the acquired measured soil organic matter content inversion data to obtain preprocessed measured soil organic matter content inversion data. Empirical mode decomposition (EMD) is used to decompose the measured soil organic matter content data at the sample scale of the preprocessed sample-scale inversion data and the soil hyperspectral data at the corresponding measured points in the preprocessed soil hyperspectral data, and extract the intrinsic mode functions containing different frequency components. Correlation analysis is performed based on the Pearson correlation coefficient to analyze the correlation between the intrinsic mode functions of different frequency components, and bands with correlations higher than a preset threshold are selected as sensitive features. The estimation model is constructed by normalizing the acquired sensitive features to obtain normalized sensitive features. The normalized sensitive features are then input into the support vector machine model, and the data is mapped to a high-dimensional space using a kernel function. Finally, the soil organic matter content estimation model is constructed by optimizing the penalty coefficient and kernel function parameters.

[0023] The above steps enhance the physical interpretability of feature selection through physical decomposition of empirical mode decomposition, improve adaptability to complex spectral features through multi-scale analysis, and screen sensitive features by combining correlation analysis of Pearson correlation coefficient, thereby constructing a new paradigm for efficient and robust feature selection, which can significantly improve the convenience, efficiency and robustness of this spectral feature selection method.

[0024] The specific steps for performing wavelet denoising on the acquired soil hyperspectral data to obtain preprocessed soil hyperspectral data are as follows: Based on the characteristics of soil hyperspectral data, a suitable wavelet basis function was selected, and the optimal number of decomposition layers was determined through experiments. Multi-scale wavelet decomposition of soil hyperspectral data was performed using selected wavelet basis functions to obtain wavelet coefficients at different scales, including low-frequency wavelet coefficients and high-frequency wavelet coefficients. The high-frequency wavelet coefficients obtained by decomposition are thresholded and noise is removed by the soft thresholding method to obtain the thresholded high-frequency wavelet coefficients. Wavelet reconstruction is performed based on the low-frequency wavelet coefficients and the high-frequency wavelet coefficients after thresholding to obtain the preprocessed soil hyperspectral data.

[0025] The above steps, by performing wavelet denoising on the acquired soil hyperspectral data, yield preprocessed soil hyperspectral data. This effectively reduces the noise level in the soil hyperspectral data, improves the signal-to-noise ratio, and enhances the clarity and identifiability of spectral features. This facilitates more accurate extraction of sensitive bands related to soil organic matter content and improves the accuracy and stability of subsequent feature selection and estimation models. As a result, it provides a more reliable data foundation for the hyperspectral inversion of soil organic matter content and enhances the robustness of this spectral feature selection method.

[0026] The specific steps for thresholding and removing noise from the high-frequency wavelet coefficients obtained by soft thresholding to obtain the thresholded high-frequency wavelet coefficients are as follows: The noise standard deviation of the high-frequency wavelet coefficients is calculated, and the discrimination threshold is obtained by calculating the discrimination threshold using the following formula:

[0027] in Indicates the discrimination threshold. Indicates the standard deviation of noise. Indicates the signal length of soil hyperspectral data; A soft threshold formula is set based on the discrimination threshold, and the soft threshold formula is applied to each high-frequency wavelet coefficient to obtain the high-frequency wavelet coefficients after thresholding. The specific soft threshold formula is as follows:

[0028] in This represents the high-frequency wavelet coefficients after thresholding. This represents the high-frequency wavelet coefficients before thresholding.

[0029] The above steps use a soft thresholding method to threshold the high-frequency wavelet coefficients obtained from the decomposition and remove noise, resulting in thresholded high-frequency wavelet coefficients. This not only removes noise while preserving important features and details in the signal and avoiding feature loss due to excessive smoothing, but also makes the retained signal features smoother and more stable through scaling, thereby improving the accuracy and reliability of subsequent feature selection and estimation models and enhancing the robustness of this spectral feature selection method.

[0030] The specific steps for processing the acquired measured data of soil organic matter content by removing missing and outlier values ​​to obtain preprocessed measured data of soil organic matter content are as follows: Missing value handling involves processing the empty or missing values ​​in the measured data of soil organic matter content inversion. Linear interpolation or polynomial interpolation methods are used to fill the missing values, or the mean or median of the data is used directly to fill the missing values. Outlier handling involves identifying outliers in the measured soil organic matter content data that do not conform to the expected pattern using the Z-Score or IQR methods. Outliers are then removed and replaced with the mean or median of the data, or interpolation methods are used to repair them.

[0031] The specific steps for decomposing the preprocessed sample-scale inverted soil organic matter content measured data and the preprocessed soil hyperspectral data at the corresponding measured points using the empirical mode decomposition method to extract intrinsic mode functions containing different frequency components are as follows: S1: Data extraction. The preprocessed measured soil organic matter content data and preprocessed soil hyperspectral data were interpolated and expanded using the Kriging interpolation method, and the preprocessed sample-scale measured soil organic matter content data and the corresponding sample-scale soil hyperspectral data of the measured points were extracted from them. S2: Initialization: Use the preprocessed sample-scale inverted soil organic matter content measured data and the corresponding measured point soil hyperspectral data in the preprocessed soil hyperspectral data as input signals and initialize the counter. S3: Envelope generation, identifying local maxima and minima in the input signal, using cubic spline interpolation to connect maxima to form the upper envelope and connect minima to form the lower envelope; S4: Residual extraction and conditional discrimination. Calculate the average value of the upper and lower envelopes, and subtract the average value of the upper and lower envelopes from the input signal to obtain the residual. Check whether the residuals meet the conditions of IMF, that is, whether the number of extreme points and zero crossover points are equal or differ by 1, and whether the average value of the upper and lower envelopes is zero. S5: Iterative update and termination. If the IMF condition is met, the residual is used as the new input signal, and steps S3 and S4 are repeated. If the IMF condition is not met, the residual is used as the new input signal to continue decomposition. When the residual signal becomes a monotonic function, the decomposition stops and multiple eigenmode functions containing different frequency components are obtained.

[0032] The specific steps for analyzing the correlation between intrinsic mode functions of different frequency components based on Pearson correlation coefficient and selecting bands with correlation higher than a preset threshold as sensitive features are as follows: For each intrinsic mode function corresponding to the soil hyperspectral data and each intrinsic mode function corresponding to the inverted measured soil organic matter content data, the Pearson correlation coefficient between them is calculated. The specific calculation formula is as follows:

[0033] in This represents the Pearson correlation coefficient. and These represent the first and second data points corresponding to the soil hyperspectral data and the measured data of soil organic matter content retrieved from the inversion database. One eigenmode function and These represent the average values ​​of the intrinsic mode functions corresponding to the soil hyperspectral data and the inverted measured soil organic matter content data, respectively. This represents the total number of intrinsic mode functions; Obtain all calculated Pearson correlation coefficients and plot their distribution histograms; dynamically determine the preset threshold based on the distribution of the Pearson correlation coefficients. Spectral bands with Pearson correlation coefficients higher than a preset threshold were selected; these spectral bands are the characteristic bands that are sensitive to changes in soil organic matter content. The selected characteristic bands are deduplicated, and only one sensitive characteristic band is retained for each repeated band, so as to obtain the sensitive characteristics in the measured soil organic matter content inversion data at the sample scale.

[0034] The specific steps for normalizing the acquired sensitive features to obtain normalized sensitive features are as follows: Iterate through all the acquired sensitive features and find the minimum and maximum values ​​for each feature dimension; Based on the minimum and maximum values ​​of each feature dimension, and using a normalization formula, all obtained sensitive features are normalized to [value missing]. Between these, the normalized sensitive features are obtained.

[0035] The specific steps for inputting the normalized sensitive features into the support vector machine model, mapping the data to a high-dimensional space using a kernel function, and constructing a soil organic matter content estimation model by optimizing the penalty coefficient and kernel function parameters are as follows: The normalized sensitive features are presented in matrix form, where each row of the matrix represents a sample and each column represents a feature. Initially, radial basis function kernels and linear kernels were selected as kernel functions. The grid search method was used to optimize the penalty coefficients and kernel function parameters within a certain range to find the parameter combination that minimizes the verification error. The model performance was evaluated using cross-validation, and the mean squared error and coefficient of determination were calculated. A scatter plot of the predicted values ​​and the true values ​​was drawn to evaluate the fit. By comparing the differences in model performance between radial basis function kernels and linear kernels, a better kernel function was selected to construct the final soil organic matter content estimation model.

[0036] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims. Parts not described in detail in this specification are prior art known to those skilled in the art.

Claims

1. A spectral feature selection method combining empirical mode decomposition, characterized in that, Includes the following steps: Data acquisition involves obtaining soil hyperspectral data and corresponding measured data of soil organic matter content through sensors and elemental analyzers. Data preprocessing involves wavelet denoising of the acquired soil hyperspectral data to obtain preprocessed soil hyperspectral data, and missing and outlier processing of the acquired measured soil organic matter content inversion data to obtain preprocessed measured soil organic matter content inversion data. Empirical mode decomposition (EMD) is used to decompose the measured soil organic matter content data at the sample scale of the preprocessed sample-scale inversion data and the soil hyperspectral data at the corresponding measured points in the preprocessed soil hyperspectral data, and extract the intrinsic mode functions containing different frequency components. Correlation analysis is performed based on the Pearson correlation coefficient to analyze the correlation between the intrinsic mode functions of different frequency components, and bands with correlations higher than a preset threshold are selected as sensitive features. The estimation model is constructed by normalizing the acquired sensitive features to obtain normalized sensitive features. The normalized sensitive features are then input into the support vector machine model, and the data is mapped to a high-dimensional space using a kernel function. Finally, the soil organic matter content estimation model is constructed by optimizing the penalty coefficient and kernel function parameters.

2. The spectral feature selection method combining empirical mode decomposition according to claim 1, characterized in that, The specific steps for performing wavelet denoising on the acquired soil hyperspectral data to obtain preprocessed soil hyperspectral data are as follows: Based on the characteristics of soil hyperspectral data, a suitable wavelet basis function was selected, and the optimal number of decomposition layers was determined through experiments. Multi-scale wavelet decomposition of soil hyperspectral data was performed using selected wavelet basis functions to obtain wavelet coefficients at different scales, including low-frequency wavelet coefficients and high-frequency wavelet coefficients. The high-frequency wavelet coefficients obtained by decomposition are thresholded and noise is removed by the soft thresholding method to obtain the thresholded high-frequency wavelet coefficients. Wavelet reconstruction is performed based on the low-frequency wavelet coefficients and the high-frequency wavelet coefficients after thresholding to obtain the preprocessed soil hyperspectral data.

3. The spectral feature selection method combining empirical mode decomposition according to claim 2, characterized in that, The specific steps for thresholding and removing noise from the high-frequency wavelet coefficients obtained by soft thresholding to obtain the thresholded high-frequency wavelet coefficients are as follows: The noise standard deviation of the high-frequency wavelet coefficients is calculated, and the discrimination threshold is obtained by calculating the discrimination threshold using the following formula: ， in Indicates the discrimination threshold. Indicates the standard deviation of noise. Indicates the signal length of soil hyperspectral data; A soft threshold formula is set based on the discrimination threshold, and the soft threshold formula is applied to each high-frequency wavelet coefficient to obtain the high-frequency wavelet coefficients after thresholding. The specific soft threshold formula is as follows: ， in This represents the high-frequency wavelet coefficients after thresholding. This represents the high-frequency wavelet coefficients before thresholding.

4. The spectral feature selection method combining empirical mode decomposition according to claim 3, characterized in that, The specific steps for processing the acquired measured data of soil organic matter content by removing missing and outlier values ​​to obtain preprocessed measured data of soil organic matter content are as follows: Missing value handling involves processing the empty or missing values ​​in the measured data of soil organic matter content inversion. Linear interpolation or polynomial interpolation methods are used to fill the missing values, or the mean or median of the data is used directly to fill the missing values. Outlier handling involves identifying outliers in the measured soil organic matter content data that do not conform to the expected pattern using the Z-Score or IQR methods. Outliers are then removed and replaced with the mean or median of the data, or interpolation methods are used to repair them.

5. The spectral feature selection method combining empirical mode decomposition according to claim 4, characterized in that, The specific steps for decomposing the preprocessed sample-scale inverted soil organic matter content measured data and the preprocessed soil hyperspectral data at the corresponding measured points using the empirical mode decomposition method to extract intrinsic mode functions containing different frequency components are as follows: S1: Data extraction. The preprocessed measured soil organic matter content data and preprocessed soil hyperspectral data were interpolated and expanded using the Kriging interpolation method, and the preprocessed sample-scale measured soil organic matter content data and the corresponding sample-scale soil hyperspectral data of the measured points were extracted from them. S2: Initialization: Use the preprocessed sample-scale inverted soil organic matter content measured data and the corresponding measured point soil hyperspectral data in the preprocessed soil hyperspectral data as input signals and initialize the counter. S3: Envelope generation, identifying local maxima and minima in the input signal, using cubic spline interpolation to connect maxima to form the upper envelope and connect minima to form the lower envelope; S4: Residual extraction and conditional discrimination. Calculate the average value of the upper and lower envelopes, and subtract the average value of the upper and lower envelopes from the input signal to obtain the residual. Check whether the residuals meet the conditions of IMF, that is, whether the number of extreme points and zero crossover points are equal or differ by 1, and whether the average value of the upper and lower envelopes is zero. S5: Iterative update and termination. If the IMF condition is met, the residual is used as the new input signal, and steps S3 and S4 are repeated. If the IMF condition is not met, the residual is used as the new input signal to continue decomposition. When the residual signal becomes a monotonic function, the decomposition stops and multiple eigenmode functions containing different frequency components are obtained.

6. The spectral feature selection method combining empirical mode decomposition according to claim 5, characterized in that, The specific steps for analyzing the correlation between intrinsic mode functions of different frequency components based on Pearson correlation coefficient and selecting bands with correlation higher than a preset threshold as sensitive features are as follows: For each intrinsic mode function corresponding to the soil hyperspectral data and each intrinsic mode function corresponding to the inverted measured soil organic matter content data, the Pearson correlation coefficient between them is calculated. The specific calculation formula is as follows: ， in This represents the Pearson correlation coefficient. and These represent the first and second data points corresponding to the soil hyperspectral data and the measured data of soil organic matter content retrieved from the inversion database. One eigenmode function and These represent the average values ​​of the intrinsic mode functions corresponding to the soil hyperspectral data and the inverted measured soil organic matter content data, respectively. This represents the total number of intrinsic mode functions; Obtain all calculated Pearson correlation coefficients and plot their distribution histograms; dynamically determine the preset threshold based on the distribution of the Pearson correlation coefficients. Spectral bands with Pearson correlation coefficients higher than a preset threshold were selected; these spectral bands are the characteristic bands that are sensitive to changes in soil organic matter content. The selected characteristic bands are deduplicated, and only one sensitive characteristic band is retained for each repeated band, so as to obtain the sensitive characteristics in the measured soil organic matter content inversion data at the sample scale.

7. The spectral feature selection method combining empirical mode decomposition according to claim 6, characterized in that, The specific steps for normalizing the acquired sensitive features to obtain normalized sensitive features are as follows: Iterate through all the acquired sensitive features and find the minimum and maximum values ​​for each feature dimension; Based on the minimum and maximum values ​​of each feature dimension, and using a normalization formula, all obtained sensitive features are normalized to [value missing]. Between these, the normalized sensitive features are obtained.

8. The spectral feature selection method combining empirical mode decomposition according to claim 7, characterized in that, The specific steps for inputting the normalized sensitive features into the support vector machine model, mapping the data to a high-dimensional space using a kernel function, and constructing a soil organic matter content estimation model by optimizing the penalty coefficient and kernel function parameters are as follows: The normalized sensitive features are presented in matrix form, where each row of the matrix represents a sample and each column represents a feature. Initially, radial basis function kernels and linear kernels were selected as kernel functions. The grid search method was used to optimize the penalty coefficients and kernel function parameters within a certain range to find the parameter combination that minimizes the verification error. The model performance was evaluated using cross-validation, and the mean squared error and coefficient of determination were calculated. A scatter plot of the predicted values ​​and the true values ​​was drawn to evaluate the fit. By comparing the differences in model performance between radial basis function kernels and linear kernels, a better kernel function was selected to construct the final soil organic matter content estimation model.

Images