Soil element hyperspectral inversion method based on CARS-IRIV waveband selection

Abstract

The invention relates to a soil element hyperspectral inversion method based on CARS-IRIV waveband selection, and belongs to the field of soil hyperspectral prediction. The method comprises the following steps: firstly, collecting a soil sample, measuring visible near-infrared hyperspectral data and chromium concentration of the soil sample, removing wavebands with relatively large noise in the hyperspectral data, dividing a data set into a training set and a verification set, preprocessing the training set and the verification set, and then extracting characteristic wavebands by adopting CARS-IRIV; and then establishing a soil chromium concentration prediction model of support vector regression, training the model by using a training set, evaluating prediction precision of the model by using a verification set in combination with evaluation indexes, and finally drawing a soil element chromium concentration spatial distribution diagram in a sampling area. According to the method, the soil elements can be quickly and accurately inverted, and powerful technical support is provided for quantitative inversion of the concentration of the heavy metal chromium in the soil.

Description

Technical Field

[0001] The present invention relates to a hyperspectral inversion method for soil elements based on CARS-IRIV band selection, belonging to the field of soil hyperspectral prediction. Background Art

[0002] Chromium (Cr) is a common pollutant in soil. Since chromium is toxic, persistent, and easily absorbed by plants, it can enter the human body through the food chain and may cause serious harm to human health. Therefore, monitoring and predicting the concentration of heavy metals in farmland soil is of great significance for preventing these harmful substances from causing harm to human health through the food chain and promoting the protection and treatment of the soil environment.

[0003] Hyperspectral technology has become one of the effective new technologies for rapidly assessing soil heavy metal pollution due to its advantages such as non-destructive detection and simple operation. Hyperspectral data can provide rich spectral information due to its high spectral resolution, but redundant bands will affect the accuracy of soil inversion. Therefore, the selection of hyperspectral data dimensionality reduction methods is crucial for improving the accuracy and computational efficiency of the model. Commonly used band selection methods include the competitive adaptive reweighted sampling (CARS) algorithm and the iterative retained informative variables (IRIV) algorithm, etc. In the study of soil heavy metal hyperspectral inversion, it is crucial to reasonably select spectral characteristic bands closely related to soil parameter inversion. The CARS algorithm has the advantage of quickly screening key bands, which can effectively remove redundant bands, reduce the feature dimension, and thus improve the computational efficiency and prediction performance of the model. At the same time, the IRIV algorithm can further refine band variables, extract strong and weak band information of soil heavy metals, reveal the mechanism of soil heavy metal inversion, and further improve the prediction accuracy and stability of the model. The applications of the CARS and IRIV methods in feature band screening are relatively extensive and each has its own advantages.

[0004] However, the computational running time of the IRIV algorithm is slow and cannot meet the future large-scale real-time monitoring requirements. The CARS algorithm can quickly select sensitive band information of soil parameters, but the characteristic bands selected by CARS are usually more, and there is still a problem of redundant information, which to a certain extent affects the efficiency and accuracy of the model. Therefore, combining the efficient screening ability of CARS and the variable refinement advantage of IRIV may obtain more robust and highly predictive features in the band selection process. And the existing research on combining CARS and IRIV for the screening of characteristic spectra of soil heavy metal elements is relatively few. Therefore, a hyperspectral inversion method for heavy metal elements in farmland soil based on CARS-IRIV band selection is proposed. Summary of the Invention

[0005] The purpose of the present invention is to provide a hyperspectral inversion method for soil elements based on CARS-IRIV band selection, aiming to solve the technical problems of slow calculation running time, redundant information in band selection, and difficulty in ensuring the running efficiency and prediction stability of the model in the prior art.

[0006] To achieve the above object, the present invention is realized through the following technical solutions: A hyperspectral inversion method for soil elements based on CARS-IRIV band selection. First, soil samples are collected, and their visible near-infrared hyperspectral data and chromium concentration data are measured. Then, the KS method is used to divide the soil sample data set into a training set and a validation set. Next, data preprocessing is performed on the training set and the validation set. Subsequently, the CARS-IRIV method is used to extract characteristic bands from the hyperspectral data. Next, a support vector regression (SVR) model is constructed to predict the chromium concentration, and the training set is used to train the model. Finally, combined with evaluation indicators, the validation set is used to evaluate the accuracy of each prediction model.

[0007] Specifically, it includes the following steps:

[0008] Step1: Collect soil samples and measure the indoor spectral reflectance data of the soil samples and the chromium content of the samples;

[0009] Step2: Eliminate the spectral intervals with large noise in the measured visible near-infrared spectral reflectance data of the soil samples;

[0010] Step3: Divide the spectral reflectance data and chromium concentration of the soil samples after eliminating the spectral intervals with large noise. The soil samples are divided into two parts, a training set and a validation set, according to a ratio;

[0011] Step4: Perform preprocessing on the divided spectral reflectance data. The preprocessing is to use the convolution smoothing method to eliminate spectral noise to obtain the original spectral reflectance data, and use the differential transformation to transform the original spectral reflectance data to highlight spectral features;

[0012] Step5: Use the CARS-IRIV band selection algorithm to screen characteristic bands from the original spectral reflectance data after preprocessing;

[0013] Step6: Use the selected spectral characteristic wavelengths as model variables. Among them, the selected characteristic bands are independent variables, and the chromium content is the dependent variable. The support vector machine SVM method is selected to construct a hyperspectral prediction model for chromium content to predict the soil heavy metal chromium content, and model testing is performed;

[0014] Step7: Use the spatial interpolation method to draw the spatial distribution map of chromium concentration in the study area based on the measured values and predicted values of chromium.

[0015] The Step 1 is specifically as follows:

[0016] Collect several soil samples, air-dry them naturally, grind them and sieve them;

[0017] The processed soil samples were divided into two parts, one part was used to collect soil spectral reflectance data for measurement in a dark room, and the other part was used to determine the heavy metal chromium content.

[0018] The spectral range with large noise in Step 2 is specifically:

[0019] The noise in the range of 350nm-399nm and 2401nm-2500nm in the process of measuring the hyperspectral data is defined as the spectral range with large noise, and the spectral data in the spectral range with large noise is eliminated.

[0020] The transformation using differential transformation in Step 4 is specifically as follows:

[0021]

[0022]

[0023] in, is the wavelength of the ith band, is the wavelength of the i+1th band, is the wavelength of the i-1th band; The wavelength The first-order differential spectrum at , The wavelength The spectral reflectance at The wavelength The spectral reflectance at The wavelength The first-order differential spectrum at , The wavelength The first-order differential spectrum at , The wavelength The second-order differential spectrum at is the wavelength interval between two adjacent bands.

[0024] The Step 5 is specifically as follows:

[0025] Step 5.1: Initially screen the characteristic bands using the CARS algorithm. Take the chromium content data of m measured samples and the spectral reflectance data of n wavelengths of m samples as input data for calculation. Evaluate the importance of each wavelength by calculating the absolute value of the regression coefficient using the regression model each time. Combine the adaptive reweighted sampling method and the exponential decay function. Retain important wavelengths and remove unimportant wavelengths by adjusting the weights. Finally, calculate the root mean square error using cross-validation and select the wavelength subset with the minimum root mean square error using the Monte Carlo sampling method as the optimal characteristic bands.

[0026] Step 5.2: Based on the characteristic bands initially screened by the CARS method, perform a secondary screening using the IRIV algorithm. First, generate different variable combinations by randomly combining variables and considering the interactions between variables. Then, establish a partial least squares (PLS) model for each variable combination and evaluate the model performance by calculating the root mean square error of cross-validation (RMSECV). Finally, select the band with the minimum RMSECV value as the optimal characteristic bands.

[0027] In IRIV, for each wavelength variable, calculate the difference in RMSECV when it is included and excluded in the model, and calculate the P-value using the Mann-Whitney U test method to evaluate the importance of each variable. According to the P-value results, eliminate the uninformative variables and interfering variables with P-values higher than the significance threshold, and retain the variables with P-values lower than the significance threshold. Through multiple iterations, gradually eliminate the uninformative variables and interfering variables until finally retaining all variables with P-values lower than the significance threshold, and then perform backward elimination. The finally obtained wavelength subset is the optimal characteristic band set.

[0028] The beneficial effects of the present invention are as follows: By combining the CARS algorithm and the IRIV algorithm, the present invention performs a secondary dimensionality reduction process on the characteristic wavelength variables, effectively solving the common redundancy and collinearity problems in hyperspectral data. The CARS algorithm can quickly screen out the key bands that have a greater impact on the model performance. Especially when dealing with high-dimensional data, it focuses on the sensitive bands related to soil elements. The IRIV algorithm further optimizes these screened bands, eliminates the uninformative or interfering variables, and accurately selects the optimal characteristic bands. Through the organic combination of CARS and IRIV, both the robustness and prediction accuracy of band selection are improved, and the computational complexity of the inversion model is significantly reduced, enhancing the operating efficiency and prediction stability of the model. The method of the present invention has strong practical application potential and can be widely applied to fields such as soil component analysis and environmental monitoring. Description of the Drawings

[0029] Figure 1 It is a schematic flowchart of the method in the embodiment of the present invention;

[0030] Figure 2 Schematic diagram of sampling points in the research area in the embodiment of the present invention;

[0031] Figure 3 Scatter plot of the predicted values and measured values of soil heavy metal nickel concentration of each prediction model on the validation set in the embodiment of the present invention;

[0032] Figure 4 Spatial distribution map of the measured values and predicted values of chromium element concentration in the research area in the embodiment of the present invention. Specific implementation manners

[0033] The content of the present invention will be further described below in conjunction with the accompanying drawings and specific implementation manners, but this does not limit the scope of the present invention.

[0034] In this embodiment, soil samples are collected from a mountainous farmland area. This area has abundant rainfall, sufficient sunlight, distinct dry and wet seasons, and belongs to the tropical semi-humid mountain monsoon climate. The soil is mostly acidic red soil. Most of the farmlands are distributed in mountainous areas with complex terrain and large height changes, showing a terrain feature of high in the north and low in the south. Due to the effect of the terrain, a small river flowing from north to south is formed in the research area, and farmlands are distributed on both sides of the river, so that the farmland area is naturally divided into three regions: upper, middle, and lower reaches. Due to the special geographical environment and mineral mining activities in this area, it is inevitable that the heavy metal content in the surrounding farmland soil exceeds the standard to varying degrees.

[0035] As Figure 1 shown, a hyperspectral inversion method for soil elements based on CARS-IRIV band selection includes the following steps:

[0036] Step1: Collect soil samples and measure the indoor spectral reflectance data of the soil samples and the chromium content of the samples;

[0037] Specifically, weeds and gravel around the sampling points are cleared before sampling. To ensure the representativeness of the collected samples, the five-point sampling method is used to collect the surface soil (0-20 cm). At the center position of each sampling point, samples are collected equidistantly in four directions from this point, and a total of 5 soil samples are collected. These samples are fully mixed to obtain the mixed soil sample of the quadrat area. In addition, a GPS measuring instrument is used to record the central position coordinates of the quadrat area and record the surrounding environment of the sampling points. After removing stones, plant residues and roots from each sample, about 1 kg of soil sample is collected and sealed for preservation. A total of 121 sampling points are collected. The sampling point distribution map is as Figure 1 shown. After the samples are taken back to the laboratory for natural air-drying treatment, they are ground and sieved. The soil samples are divided into two parts. One part is used for the determination of soil element content, and the other part is used to measure the soil spectral reflectance using the FieldSpec3 portable ground object spectrometer of ASD Company.

[0038] Step 2: Eliminate the spectral intervals with relatively large noise in the visible near-infrared spectral reflectance data of the measured soil samples;

[0039] Specifically, eliminate the spectral intervals with relatively large noise in the visible near-infrared hyperspectral data of the soil samples measured in Step 1. The measured spectral wavelength range is 350 nm to 2500 nm, and the sampling interval is 1 nm. During the spectral acquisition process, due to reasons such as the instrument, the noise is relatively large in the intervals of 350 nm to 399 nm and 2401 nm to 2500 nm. Therefore, the spectral data in these two intervals are eliminated.

[0040] Step 3: Divide the spectral reflectance data and chromium concentration of the soil samples after eliminating the spectral intervals with relatively large noise into two parts: a training set and a validation set according to a ratio;

[0041] Specifically, use the KS algorithm to divide the hyperspectral data and chromium concentration of the 121 soil samples after eliminating the spectral intervals with relatively large noise in Step 2 into a training set and a validation set according to a ratio of 3:1; among them, there are 90 in the training set and 31 in the validation set. The statistical results of the chromium element concentration after dividing the soil samples are shown in Table 1.

[0042] Table 1 Statistical description of chromium element concentration in soil samples

[0043] Step 4: Preprocess the divided spectral reflectance data. The preprocessing is to use the convolution smoothing method to eliminate spectral noise to obtain the original spectral reflectance data, and use differential transformation on the original spectral reflectance data to highlight spectral features;

[0044] Specifically, preprocess the reflectance data processed in Step 3; among them, the data preprocessing uses Savitzky-Golay convolution smoothing to eliminate spectral noise to obtain the original spectral reflectance data, and use the second-order differential to transform the spectral data to highlight spectral features.

[0045]

[0046]

[0047] Among them, is the wavelength of the i-th band, is the wavelength of the (i + 1)-th band, is the wavelength of the (i - 1)-th band; is the wavelength at the first-order differential spectrum, is the spectral reflectance at wavelength , is the spectral reflectance at wavelength , is the first-order differential spectrum at wavelength , is the first-order differential spectrum at wavelength , is the second-order differential spectrum at wavelength , is the wavelength interval between two adjacent bands.

[0048] Step5: Use the CARS-IRIV band selection algorithm to screen the characteristic bands from the preprocessed original spectral reflectance data;

[0049] Specifically, use the CARS, IRIV, and CARS-IRIV band selection algorithms to screen the characteristic bands from the preprocessed spectral reflectance data in Step4. The CARS-IRIV characteristic band selection process is divided into the following steps:

[0050] Step5.1: Use the CARS algorithm to preliminarily screen the characteristic bands. Take the chromium content data of m samples and the spectral reflectance data of n wavelengths of m samples as input data for calculation. Evaluate the importance of each wavelength by calculating the absolute value of the regression coefficient using the regression model each time. Combine the adaptive reweighted sampling method and the exponential decay function to retain important wavelengths and remove unimportant wavelengths by adjusting the weights. Finally, calculate the root mean square error using cross-validation and select the wavelength subset with the smallest root mean square error as the optimal characteristic band using the Monte Carlo sampling method, so as to obtain the best variable combination. The main parameter settings are: the number of cross-validations is 10, the number of iterations is 100, and the ratio of the modeling set to the calibration set during Monte Carlo sampling is 0.8.

[0051] Step5.2: Simultaneously record the characteristic bands and running time obtained during the CARS screening process for each individual record.

[0052] Step5.3: Use IRIV to perform secondary screening on the characteristic bands initially screened by the CARS algorithm. First, generate different variable combinations by randomly combining variables and considering the interactions between variables. Then, establish a partial least squares (PLS) model based on each variable combination and evaluate the model performance by calculating the root mean square error of cross-validation (RMSECV). Finally, select the band with the minimum RMSECV value as the optimal characteristic band. Then, in IRIV, for each wavelength variable, calculate the difference in RMSECV when it is included and excluded from the model, and use the Mann-Whitney U test to calculate the P-value to evaluate the importance of each variable. According to the P-value results, eliminate the uninformative variables and interfering variables with P-values higher than the significance threshold, and retain the variables with P-values lower than the significance threshold. Through multiple iterations, gradually eliminate the uninformative variables and interfering variables until finally retaining all variables with P-values lower than the significance threshold, and then perform backward elimination. The final wavelength subset obtained is the optimal characteristic band set.

[0053] Step5.4: Record the characteristic bands and running time obtained during the single IRIV screening process, as well as the running time of the entire CARS-IRIV band selection method. Finally, the various indicators of the CARS, IRIV, and CARS-IRIV feature selection algorithms are shown in Table 2 below:

[0054] Table 2 Various indicators of the CARS, IRIV, and CARS-IRIV feature selection algorithm results

[0055] Judging from the calculation running time in Table 2, the CARS-IRIV method can significantly shorten the calculation running time of band selection. The calculation time of using the IRIV method alone is relatively long, which is 11319.99 seconds, while the CARS method has extremely high running efficiency, with a running time of 24.4 seconds. However, first perform a preliminary screening of the hyperspectral bands by the CARS method, and then use the IRIV method to refine and optimize the preliminarily screened bands, the calculation running time can reach 90.89 seconds. Judging from the number of characteristic bands extracted, the number of characteristic bands selected based on the CARS-IRIV method is 50; while the number of characteristic bands extracted based on the CARS method is 87, accounting for 4.35% of the entire band. It can be seen that the number of characteristic bands extracted by the CARS-IRIV method is less. Therefore, the CARS-IRIV method gives full play to the high efficiency of CARS and the precision advantage of IRIV, and can further eliminate the redundant band information of heavy metals compared with the individual CARS and IRIV methods.

[0056] Step 6: Use the selected spectral characteristic wavelengths as model variables. Among them, the selected characteristic bands are independent variables, and the chromium content is the dependent variable. Select the support vector machine (SVM) method to construct a hyperspectral prediction model for chromium content to predict the soil heavy metal chromium content, and conduct model tests;

[0057] Specifically, use the spectral characteristic wavelengths selected in Step 5 as model variables, establish a hyperspectral prediction model for chromium content to predict the soil heavy metal chromium content, and conduct model tests. The selected characteristic bands are independent variables, and the chromium content is the dependent variable. Select the support vector machine (SVM) method to construct a hyperspectral prediction model for chromium content; and use the coefficient of determination (R2), root mean squared error (RMSE), and relative analysis error (RPD) to evaluate the model accuracy. The closer R2 is to 1, the better the model fitting degree and stability. The lower the RMSE, the higher the model accuracy. Among them, when RPD > 2.0, it indicates that the model has good prediction ability; when it is in the range of 1.4 < RPD < 2.0, the model prediction ability is average; when RPD < 1.4, the model prediction ability is poor. Finally, the evaluation indexes of the SVR models based on the characteristic bands obtained by CARS, IRIV, and CARS - IRIV on the training set and test set are shown in Table 3 below:

[0058] Table 3 Modeling results based on CARS, IRIV, and CARS - IRIV feature selection combined with SVR

[0059] Judging from the inversion model accuracy indexes in Table 3, in terms of the performance of the model on the validation set, the inversion model accuracy of the CARS - IRIV method is better than that of using CARS or IRIV alone. The modeling accuracies are in the order of: CARS - IRIV > CARS > IRIV. The R 2 values of the model based on the CARS - IRIV strategy on the validation set all reach above 0.93, and the RPD values all reach above 3.84, with the best modeling effect. Followed by the model based on CARS, whose R 2 is 0.88, RMSE is 185.69 mg·kg -1 , RPD is 2.95. Finally, the R 2 value of the model based on IRIV on the validation set is 0.72, RMSE is 291.71 mg·kg -1, the RPD is 1.88. This indicates that the prediction performance based on the CARS-IRIV strategy model is excellent, and there is a significant improvement in the modeling effect compared to the other two models. This is because the CARS method focuses on reducing redundant band features, while the IRIV method pays more attention to the ability to distinguish strong and weak feature bands, thus affecting the inversion effect of the model. In feature band selection, the CARS method can identify the sensitive bands of soil elements in the preliminary screening stage. Especially when dealing with high-dimensional data, it can focus on the feature bands that have a greater impact on the model performance. Combining with the scatter plots of the measured and predicted values of soil element chromium in the validation set of each prediction model, as Figure 3 shown. Among them, the prediction points of the training set and the validation set on the scatter plot of the model based on the CARS-IRIV band selection strategy are closer to the 1:1 line than those of the other two single band selection algorithms.

[0060] Step 7: Use the spatial interpolation method to draw the spatial distribution map of chromium concentration in the study area based on the measured and predicted values of chromium;

[0061] Specifically, use the ordinary Kriging interpolation method to draw the spatial distribution map of chromium concentration in the study area based on the measured and predicted values of chromium, as Figure 4 shown. The spatial distribution map of the content of soil element chromium in the study area is drawn based on the estimated values obtained from the CARS-IRIV band selection strategy model, as Figure 4 shown in b, and at the same time Figure 4 a shows the spatial distribution based on the measured values. Figure 4 a and Figure 4 The spatial distribution maps of heavy metals in b are highly similar, indicating that the spatial distribution of the interpolation results is consistent with the measured values, which shows that the CARS-IRIV band selection strategy model adopted has a good fitting effect and can accurately predict the content of soil heavy metals, thus effectively revealing the spatial variation characteristics of soil heavy metals in the study area.

[0062] Based on the above results, the effectiveness of the CARS-IRIV band selection strategy is demonstrated, and the soil element inversion model established based on it basically meets the requirements of actual use and has the feasibility of application.

[0063] The specific implementation manners of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above implementation manners. Within the scope of knowledge possessed by those of ordinary skill in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A soil element hyperspectral inversion method based on CARS-IRIV band selection, characterized in that: The specific steps include: Step 1: Collect soil samples, and measure the indoor spectral reflectance data of the soil samples and the chromium content of the samples; Step 2: Eliminate the spectral interval with large noise in the visible near-infrared spectral reflectance data of the measured soil samples; Step 3: Divide the spectral reflectance data and chromium concentration of the soil samples after removing the spectral interval with large noise, and divide the soil samples into two parts: training set and validation set according to the proportion; Step 4: preprocessing the divided spectral reflectance data, wherein the preprocessing is to use a convolution smoothing method to eliminate spectral noise, obtain original spectral reflectance data, and transform the original spectral reflectance data using differential transformation to highlight the spectral characteristics; Step 5: Use the CARS-IRIV band selection algorithm to screen the characteristic bands of the preprocessed raw spectral reflectance data; Step 6: Use the selected spectral characteristic wavelengths as model variables, where the selected characteristic bands are independent variables and the chromium content is the dependent variable. Use the support vector machine (SVM) method to construct a chromium content hyperspectral prediction model to predict the soil heavy metal chromium content, and perform model testing. Step 7: Use the spatial interpolation method to draw the spatial distribution map of chromium concentration in the study area based on the measured and predicted values ​​of chromium.

2. The soil element hyperspectral inversion method based on CARS-IRIV band selection according to claim 1 is characterized in that: The Step 1 is specifically as follows: Collect several soil samples, air-dry them naturally, grind them and sieve them; The processed soil samples were divided into two parts, one part was used to collect soil spectral reflectance data for measurement in a dark room, and the other part was used to determine the heavy metal chromium content.

3. The soil element hyperspectral inversion method based on CARS-IRIV band selection according to claim 1 is characterized in that: The spectral range with large noise in Step 2 is specifically: The noise in the range of 350nm-399nm and 2401nm-2500nm in the process of measuring the hyperspectral data is defined as the spectral range with large noise, and the spectral data in the spectral range with large noise is eliminated.

4. The soil element hyperspectral inversion method based on CARS-IRIV band selection according to claim 1 is characterized in that: The transformation using differential transformation in Step 4 is specifically as follows: ； ； in, is the wavelength of the ith band, is the wavelength of the i+1th band, is the wavelength of the i-1th band; The wavelength The first-order differential spectrum at , The wavelength The spectral reflectance at The wavelength The spectral reflectance at The wavelength The first-order differential spectrum at , The wavelength The first-order differential spectrum at , The wavelength The second-order differential spectrum at is the wavelength interval between two adjacent bands.

5. The soil element hyperspectral inversion method based on CARS-IRIV band selection according to claim 1 is characterized in that: The Step 5 is specifically as follows: Step 5.1: Use the CARS algorithm to preliminarily screen the characteristic bands. Take the measured chromium content data of m samples and the spectral reflectance data of n wavelengths of m samples as input data for calculation. The importance of each wavelength is evaluated by calculating the absolute value of the regression coefficient using the regression model each time. Combined with the adaptive reweighted sampling method and the exponential decay function, the weights are adjusted to retain important wavelengths and remove unimportant wavelengths. Finally, cross-validation is used to calculate the root mean square error, and the Monte Carlo sampling method is used to select the wavelength subset with the smallest root mean square error as the optimal characteristic band; Step 5.2: Based on the characteristic bands initially screened by the CARS method, the IRIV algorithm is used for secondary screening. First, different variable combinations are generated by randomly combining variables and combining the interactions between variables. Then, a partial least squares PLS model is established based on each variable combination, and the root mean square error RMSECV is calculated by cross-validation to evaluate the model effect. Finally, the band with the smallest RMSECV value is selected as the optimal characteristic band. Step 5.3: In IRIV, for each wavelength variable, the difference in RMSECV when it is included and excluded in the model is calculated, and the P value is calculated using the Mann-Whitney U test to evaluate the importance of each variable. Based on the P value results, uninformative variables and interference variables with P values ​​higher than the significance threshold are eliminated, and variables with P values ​​lower than the significance threshold are retained. Through multiple iterations, uninformative variables and interference variables are gradually eliminated until all variables with P values ​​lower than the significance threshold are finally retained, and then reverse elimination is performed. The final wavelength subset is the optimal characteristic band set.

Images