Method for inverting heavy metal content in soil by using profile spectrum of undisturbed soil

Abstract

The invention discloses a method for inverting the content of heavy metals in soil by using an undisturbed soil profile spectrum, and belongs to the technical field of soil attribute detection. Comprising the following steps: collecting a plurality of undisturbed soil profile samples; the method comprises the following steps: dividing a soil profile into a plurality of soil layers at equal intervals, measuring visible-near infrared hyperspectral data of each soil layer, and analyzing and determining the standard content of heavy metal elements in a sample in the soil layer; taking the preprocessed hyperspectral data predictive variable, taking the standard soil heavy metal content as the response variable, and adopting different machine learning algorithms to respectively construct hyperspectral inversion models of the four soil heavy metal elements of the undisturbed profile; and determining an optimal inversion model of each soil heavy metal element according to the evaluation indexes. The method provided by the invention can meet the application requirements of rapid, accurate and nondestructive detection of the contents of different heavy metal elements in the soil profile.

Description

Technical Field

[0001] The present invention relates to the technical field of soil property detection, and in particular to a method for inverting the heavy metal content in soil by using the original soil profile spectrum. Background Art

[0002] At present, most of the methods for detecting heavy metal content at home and abroad are mainly based on laboratory physical and chemical analysis, and the heavy metal content in the sample is determined by analytical instruments. The traditional physical and chemical test and analysis methods for heavy metal content in soil mainly include atomic absorption spectrophotometry (AAS), atomic fluorescence spectrometry (AFS), inductively coupled plasma emission spectroscopy (ICP-AES) and inductively coupled plasma mass spectrometry (ICP-MS). Although these technologies have relatively low detection limits for heavy metal content detection, high sensitivity and high accuracy, they all require acid digestion or microwave digestion of samples for pretreatment, which has the problems of high cost, long time consumption, complex operation and easy formation of secondary pollution, and it is difficult to meet the needs of convenient and efficient detection of heavy metal content in a large number of soil samples. Therefore, it is very necessary to develop a convenient and efficient detection technology for heavy metal content in soil, which is of great significance to the in-depth study of soil heavy metal pollution, environmental safety and human health.

[0003] Visible-near infrared reflectance spectroscopy (Vis-NIR) is an emerging near-ground sensing technology that can analyze soil properties using the absorption characteristics of electromagnetic waves. The energy required for the movement of electrons and nuclear transitions in different soil material components is different, and the energy level is related to the wavelength and frequency of the electromagnetic wave. Therefore, electromagnetic waves of different bands produce different spectral characteristics when acting on soil samples. In the visible light band, the absorption spectrum is mainly caused by charged particles and is closely related to the content of iron ions in minerals. In the near-infrared band, the absorption spectrum is mainly caused by OH - 、SO4 2- 、CO3 2- This is due to the different absorption energies of different ionic bonds in the plasma group. Due to the advantages of Vis-NIR, such as fast analysis, non-destructiveness, low cost, and the ability to predict multiple elements at once, it has been widely used in predicting soil properties. Currently, Vis-NIR has been successfully used to predict soil properties such as soil organic matter, nitrogen, phosphorus, potassium, iron oxides, clay, and heavy metal content.

[0004] The invention patent with application number 201810233070.X discloses a method for inverting the heavy metal content of soil by combining indoor and outdoor spectra. First, based on the series of spectral preprocessing of the soil indoor spectrum, a full-factor principal component stepwise regression model for the inversion of soil heavy metal content is constructed, and the characteristic band principal components are extracted; then, the Kennard-Stone algorithm is used to select a suitable transformation set to construct a correlation transformation model of the indoor and outdoor spectra of the soil, and the converted soil outdoor spectrum is also subjected to a series of spectral preprocessing; finally, the characteristic band principal components extracted by the soil indoor spectral inversion model of the heavy metal content are integrated to construct a regression model for the hyperspectral inversion of heavy metal content based on the outdoor spectrum of the soil after conversion.

[0005] The invention patent with application number 202010427959.9 discloses a support vector machine regression method for heavy metal content that combines the bands and ratio combinations of indoor and outdoor spectra, and its steps include: step S1: soil sample collection and soil spectrum measurement; step S2: KS-DS algorithm to build a spectral conversion model; step S3: spectral preprocessing and correlation analysis; step S4: construction of a support vector machine heavy metal content inversion model that integrates indoor and outdoor spectra for detection.

[0006] The invention patent with application number 202111309905.3 discloses a method for estimating the heavy metal content in soil based on hyperspectral remote sensing technology. The implementation steps are as follows: 1) Obtain soil samples from the area to be monitored, and after natural air drying and grinding in the laboratory, measure the spectral reflectance of the soil samples and the contents of four heavy metal elements: chromium, cobalt, copper and arsenic in the soil; 2) After Savitzky-Golay smoothing and spectral resampling of the original spectral data, use different spectral transformation methods to perform correlation analysis on the reflectance data and heavy metal element data to obtain the modeling bands of each heavy metal element; 3) Use different modeling methods to fit the regression equation of the reflectance data and the heavy metal content to construct a hyperspectral estimation model, and evaluate the model for each heavy metal element based on four evaluation indicators: modeling determination coefficient, verification determination coefficient, root mean square error and relative analysis error, to obtain the best estimation model for each heavy metal element.

[0007] Although the methods described in the above inventions can obtain the content of heavy metal elements in soil with high detection accuracy, they all require the soil samples to be air-dried, ground and sieved (for example, 100 mesh, 60 mesh) before being placed in a sample container and then measuring their hyperspectral data, and cannot directly detect the heavy metal content of the original soil profile. Therefore, it is urgent to find a method that can directly measure hyperspectral data on the original soil profile and invert the content of heavy metal elements. Summary of the invention

[0008] In order to solve the technical problems existing in the background technology, the present invention provides a method for inverting the heavy metal content in soil by using the original soil profile spectrum.

[0009] The present invention adopts the following technical scheme: a method for inverting the heavy metal content in soil using the original soil profile spectrum, comprising at least the following steps: S1 Soil profile sample pretreatment: Collect several soil profile samples in the test area, and divide each soil profile sample into several continuous soil layers from the surface to the inside; S2 Effective hyperspectral detection: N measurement points are evenly selected on each soil layer sample, and the spectrum measurement is repeated 5 times at each measurement point to obtain 40 spectrum curves. The average value of the 40 spectrum data is calculated as the effective hyperspectral data of the soil layer; S3 Detection of standard content values ​​of heavy metals: Conduct at least four heavy metal content tests on each soil layer to obtain the standard content value of each heavy metal; S4 Inversion of heavy metal content in unknown soil samples: Firstly, using effective hyperspectral data as the prediction variable and the standard content value of heavy metals as the corresponding variable, several hyperspectral inversion models were established using at least one machine learning algorithm. Based on the evaluation index, the optimal hyperspectral inversion model for each soil heavy metal element was determined. Secondly, the heavy metal content in unknown soil samples is inverted based on the optimal hyperspectral inversion model.

[0010] The specific process of constructing the hyperspectral inversion model is as follows: S41 Modeling set and test set division: The principal component analysis method is used to detect the uniformity of effective hyperspectral data, and the Hotelling statistic test method is used to identify and eliminate abnormal effective hyperspectral data; the Kennard-Stone algorithm is used to divide the retained effective hyperspectral data into a modeling set and a test set, where the ratio of the modeling set to the test set is 7:3; S42 Determine the best parameter combination: Use grid search and ten-fold cross validation method to optimize the parameters of each machine learning algorithm, that is, each algorithm is given an initial value range and a search interval to obtain a corresponding parameter combination; for each heavy metal element, use grid search and ten-fold cross validation method to screen all parameter combinations to obtain the corresponding best parameter combination; S43 is based on the optimal parameter combination, with the effective hyperspectral data as the prediction variable and the standard content value of heavy metals as the response variable, to establish a mapping relationship between the effective hyperspectral data and the standard content values ​​of different heavy metal elements, that is, a hyperspectral inversion model.

[0011] Furthermore, the machine learning algorithm includes at least: PLSR algorithm, Cubist algorithm, GPR algorithm and SVM algorithm; Correspondingly, the several hyperspectral inversion models are respectively a PLSR inversion model, a Cubist prediction model, a GPR prediction model and a SVM prediction model.

[0012] Furthermore, the evaluation index at least includes: determination coefficient, root mean square error and Lin consistency correlation coefficient; the screening process of the optimal hyperspectral inversion model is as follows: The evaluation values ​​of the standard content values ​​of the four heavy metals predicted by different hyperspectral inversion models were judged according to the following criteria:

[0013] Where LCCC represents Lin's consistency correlation coefficient.

[0014] Furthermore, the four soil heavy metal elements are: arsenic (As), copper (Cu), lead (Pb), and zinc (Zn).

[0015] Furthermore, the best parameter combination of the PLSR inversion model is the parameter combination corresponding to the root mean square error value of the ten-fold cross validation no longer changes significantly; The best parameter combination of the Cubist inversion model, the GPR inversion model and the SVM inversion model is the parameter combination corresponding to the minimum root mean square error value of the ten-fold cross validation.

[0016] Furthermore, step S2 also includes preprocessing of effective hyperspectral data, and the specific process is as follows: one or more combinations of apparent absorbance, first-order derivative, second-order derivative, Savitzky-Golay smoothing, Gap-Segment derivative, detrending, standard normal variable transformation, and continuum removal processing methods.

[0017] Beneficial effects of the present invention: The present invention can quickly and accurately predict the content of As, Cu, Pb, and Zn elements in the original soil profile, making up for the shortcomings of traditional laboratory chemical analysis methods and conventional visible-near infrared spectroscopy technology. Through model evaluation and optimization screening, a machine learning model with strong robustness and high prediction accuracy can be promoted and applied to the prediction of different heavy metal contents in the original soil profile of similar farmlands across the country, and used to study the distribution and migration characteristics of different heavy metal elements in the soil profile, and evaluate the degree of heavy metal pollution in the soil. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of the hyperspectral inversion method for heavy metal content in the original soil profile in Example 1; Figure 2 The average spectrum curve of 20 soil layers in one soil profile in Example 1; Figure 3The modeling set R is the four machine learning models in Example 1 that predict the soil As, Cu, Pb and Zn contents respectively. 2 Value box plot; Figure 4 Box plots of RMSE values ​​of the modeling set for predicting soil As, Cu, Pb and Zn contents by the four machine learning models in Example 1, respectively; Figure 5 Box plots of LCCC values ​​of the modeling set predicted by the four machine learning models in Example 1 for soil As, Cu, Pb and Zn contents, respectively; Figure 6 The test set scatter plot of the four machine learning models in Example 1 predicting soil As content; Figure 7 The test set scatter plot of the four machine learning models in Example 1 predicting soil Cu content; Figure 8 It is a scatter plot of the test set of the four machine learning models in Example 1 for predicting soil Pb content; Fig. 9 This is a scatter plot of the test set of the four machine learning models in Example 1 for predicting soil Zn content. DETAILED DESCRIPTION

[0019] Heavy metals usually refer to 60 elements with a density greater than 4.0 or 45 elements with a density greater than 5.0. Heavy metals in soil pollution include Hg, Cd, Pb, Cr and metalloid As with strong biological toxicity, as well as toxic elements such as Zn, Cu, Co, and Ni. Human activities, especially the discharge of waste gas, wastewater, and waste residue during mineral mining and smelting, and the irrational application of fertilizers and pesticides in agricultural farming, have aggravated the accumulation and diffusion of heavy metal pollutants in the soil. Heavy metal pollutants in soil are hidden, highly toxic, difficult to degrade, hysteretic, and bioaccumulative, making them difficult to control and restore. As a high-risk area for heavy metal pollution, the content of heavy metals in farmland soils around mining and smelting areas is relatively high, and the safe production of agricultural products is seriously threatened. Heavy metal pollutants generated during the production process of industrial and mining plants can easily migrate into the surrounding farmland soils, infiltrate into the soil with rainfall, and spread with surface runoff and groundwater.

[0020] like Figure 1 As shown, the present invention provides a method for inverting the content of heavy metals in soil by using the spectrum of the original soil profile, and uses the visible-near infrared effective hyperspectral to quickly detect the content of heavy metals As, Cu, Pb and Zn in the original soil profile of farmland as a specific embodiment for explanation, and the specific process is as follows: S1 Soil profile sample pretreatment: Soil sampling is performed in the detection area at a predetermined depth to obtain several soil profile samples of the detection area; in this embodiment, the detection area is located in the farmland around two typical industrial and mining plants, Guixi Smelter and Dexing Copper Mine in Jiangxi Province, and a gasoline-powered soil sampler is used to collect undisturbed soil profile samples in the above areas: the depth is 100 cm, and the number of samples is 19. Each soil profile sample is cut vertically into two halves with a stainless steel knife, and a relatively flat half-section is selected with a measurement interval of 5 cm, and each half-section is divided into about 20 soil layers from the surface to the bottom.

[0021] S2 Effective hyperspectral data detection: ASD FieldSpec 4 ground feature hyperspectrometer and high-density contact probe are used to measure the reflectance spectrum data of each soil layer. S3 Detection of standard content values ​​of heavy metals: After the reflectance spectrum data is measured as effective hyperspectral data, soil samples of each soil layer are collected as analysis samples. The analysis samples are pre-treated, ground and sieved, and the content of four heavy metals is determined using standard analysis methods to obtain the standard content value of each heavy metal.

[0022] S4 Inversion of unknown heavy metal content in soil: Before constructing the hyperspectral inversion model, spectral preprocessing technology is used to eliminate noise and irrelevant information in the reflectance spectrum data to establish an inversion model with high prediction accuracy and good robustness.

[0023] The grid search method was used to optimize the parameters of the machine learning model, and the optimal parameters were selected and the inversion model was trained based on the ten-fold cross-validation results. At least one machine learning algorithm was used to create several Vis-NIR hyperspectral inversion models of soil heavy metal content. Based on the evaluation indicators, the best prediction model corresponding to the type of soil heavy metals was selected from several hyperspectral inversion models, and the heavy metal content of unknown soil samples was inverted based on the best prediction model.

[0024] In a further embodiment, step S1 uses the Netherlands Eijkelkamp The soil sampling drill collected a total of 19 undisturbed soil profile samples with a length of 100 cm and an inner diameter of approximately 8.4 cm. Each profile sample was numbered, and the longitude, latitude, altitude, collection depth, land use type, and other information of the collection site were recorded in detail. The undisturbed soil samples were packaged with film tape to avoid damage during transportation.

[0025] The preparation process of soil profile samples is as follows: Use a stainless steel knife to cut each collected soil profile sample vertically along the axial direction into two semi-cylindrical profile samples for portable Vis-NIR hyperspectral measurement. Since the collected original soil profile samples contain high moisture, direct hyperspectral measurement will affect the prediction performance of the inversion model. Therefore, the semi-cylindrical soil profile samples are air-dried indoors for 3 to 4 days.

[0026] The effective hyperspectral data in step S2 are obtained by ASD FieldSpec 4 ground feature hyperspectrometer, with a band range of 350-2500 nm, a resolution of 3 nm (350-1000 nm) and 10 nm (1000-2500 nm), and a sampling interval of 1.4 nm (350-1000 nm) and 2 nm (1000-2500 nm). Its high-density contact probe is vertically covered on the surface of the soil sample to measure its reflectance spectrum. The probe has a built-in 6.5 W halogen lamp as a light source, and the front measurement window diameter is about 2 cm. Eight measurement points are evenly selected for each soil layer sample, and the spectrum measurement is repeated 5 times at each point. The arithmetic mean of 40 spectrum curves is calculated as the reflectance spectrum data of the soil layer sample. Before the measurement, the instrument is preheated for 30 min and whiteboard calibration is performed. After the measurement of each soil layer sample is completed, the probe contact window is cleaned with ethanol and sanitary cotton balls, and a whiteboard calibration is performed once after the measurement of each profile sample is completed.

[0027] In a further embodiment, the effective hyperspectral data in step S2 is processed in the following manner: the influence of the band noise at the beginning and end of the spectral band (350-379 nm and 2451-2500 nm) is deleted, and the reflectance spectrum data in the 380-2450 nm band is retained; the hyperspectral data in the 380-2450 nm range is resampled at a sampling interval of 5 nm to reduce the redundancy between the hyperspectral data, and each hyperspectral data contains a total of 415 spectral bands. Figure 2 As shown, 20 hyperspectral curves were obtained for each soil profile sample.

[0028] In a further embodiment, the following preprocessing methods are used in step S4 to reduce the influence of instrument background noise or baseline drift on spectral reflectance: one or more of the processing methods such as apparent absorbance (log(1 / R)), first-order derivative, second-order derivative, Savitzky-Golay smoothing, Gap-Segment derivative, detrending, standard normal variable transformation, and continuum removal. After system comparison, the optimal spectral preprocessing method for soil heavy metals is screened out as follows: first, the apparent absorbance (log(1 / R)) is transformed, and then the first-order derivative is processed.

[0029] In a further embodiment, the four soil heavy metal elements in step S3 are: arsenic (As), copper (Cu), lead (Pb), and zinc (Zn). Specifically, the following analysis and testing process is included: the soil sample is crushed with a wooden stick to pick out impurities such as stones, animal and plant remains, and then it is naturally air-dried and then ground and passed through a 100-mesh sieve. The sieved soil sample is prepared into a test solution using the HF-HNO3-HCIO4 three-acid digestion method. The content of Cu, Pb, and Zn in the soil is determined by atomic absorption spectroscopy, and the content of As is determined by atomic fluorescence spectrometry. The quality of heavy metal analysis is controlled by adding parallel samples and soil composition analysis standard substances. The unit of soil heavy metal content is mg / kg, and the statistical characteristics of the standard content of each heavy metal element are shown in Table 1.

[0030] Table 1 Statistical characteristics of heavy metal contents in undisturbed soil profiles

[0031] In a further embodiment, at least one learning algorithm in step S4 is a PLSR algorithm, a Cubist algorithm, a GPR algorithm and a SVM algorithm. Correspondingly, the several hyperspectral inversion models are a PLSR inversion model, a Cubist inversion model, a GPR inversion model and a SVM inversion model.

[0032] The creation process of the hyperspectral inversion model is as follows: The spectral homogeneity of all soil samples was tested using principal component analysis and Hotelling's T 2 ) test method to identify and remove abnormal samples. Using the Kennard-Stone algorithm, 70% of the 390 soil samples were used as the modeling set (a total of 273 samples), and 30% were used as an independent test set (a total of 117 samples). Each inversion model was repeated 100 times on the modeling set to evaluate the stability and uncertainty of the model.

[0033] Grid search and ten-fold cross validation are used to optimize the parameters of each machine learning algorithm, that is, each algorithm is given an initial value range and search interval to obtain a corresponding parameter combination; for each soil heavy metal element, all parameter combinations are screened to obtain the corresponding optimal parameter combination. In a further embodiment, the PLSR inversion model, the Cubist inversion model, the GPR inversion model and the SVM inversion model are constructed using the "pls", "cubist", "gaussprPoly" and "svmRadial" methods in the "caret" package of the R open source programming language, respectively. The GPR model selects a polynomial as the kernel function, and the SVM model selects a radial basis function (RBF) as the kernel function. For each soil heavy metal element, grid search and ten-fold cross validation are used to screen the optimal parameters of the four inversion models of PLSR, Cubist, GPR and SVM, respectively. The PLSR inversion model takes the number of principal components corresponding to the time when the RMSE value of the ten-fold cross validation no longer changes significantly as the optimal number of principal components of PLSR. The Cubist, GPR and SVM inversion models take the parameter combination corresponding to the minimum RMSE value of the ten-fold cross validation as the optimal parameter combination.

[0034] For example, first, the initial value range and search interval are set for the parameters of each inversion algorithm (as shown in Table 2): PLSR model (ncomp = 1, 2, 3, ..., 20); Cubist model (committees = 1, 2, 3, ..., 100; neighbors = 1, 2, 3, ..., 10); GPR model (degree = 1, 2, 3, ..., 10; scale = (1, 2, 3, ..., 10000) × 10 −5 ); SVM model (sigma=(1, 2, 3, …, 10000)×10 −6 ; C penalty coefficient = 1, 2, 3, ..., 200). Secondly, among all candidate parameter combinations, try each possibility by looping through them, and find the parameter combination corresponding to the smallest ten-fold cross-validation RMSE value as the best parameter of the inversion model. Among them, the ten-fold cross-validation is to randomly divide the sample data into 10 parts, take 9 of them as training sets in turn, and the remaining 1 as a test set for evaluation tests. This process is repeated many times, and the average is calculated, and the smallest RMSE value is selected as an estimate of the accuracy of the algorithm. The optimal parameters for predicting soil As, Cu, Pb and Zn contents by the above four inversion models are shown in Table 2.

[0035] Table 2 Optimal parameter values ​​of inversion models for heavy metals in different soils

[0036] Based on the optimal parameter combination, the preprocessed hyperspectral data was used as the prediction variable and the standard soil heavy metal content was used as the response variable, and a mapping relationship between the Vis-NIR hyperspectral data and the heavy metal content of different soils was established.

[0037] In a further embodiment, the model performance evaluation index in step S4 at least includes: the coefficient of determination (R 2 ), root mean square error (RMSE) and Lin consistency correlation coefficient (LCCC); the evaluation indicators are used to evaluate the prediction performance of different inversion models on the modeling set and the test set, and select the best prediction model, which specifically includes the following steps: compare the soil heavy metal content evaluation values ​​predicted by the visible-near infrared hyperspectral inversion model on the modeling set and the test set, and use the following standards for judgment:

[0038] Based on the above process, the machine learning inversion models of soil As, Cu, Pb and Zn contents were evaluated respectively: Figure 3 , Figure 4 and Figure 5 The performance indicator box plots of the machine learning inversion model with ten-fold cross validation repeated 100 times for the modeling set were plotted respectively; Figure 6 , Figure 7 , Figure 8 and Fig. 9 Scatter plots of measured values ​​and predicted values ​​of soil heavy metals were drawn respectively.

[0039] The model performance evaluation results of the modeling set showed that the prediction performance of the three nonlinear machine learning models, Cubist, GPR and SVM, was better than that of the PLSR model; the GPR inversion model predicted soil As (R 2 =0.55, RMSE=5.24 mg / kg, LCCC=0.72) and Cu (R 2 =0.75, RMSE=16.50 mg / kg, LCCC=0.85) had the highest accuracy, indicating that the established GPR model had a medium ability to predict As and a good ability to predict Cu. The Cubist inversion model predicted soil Pb (R 2 =0.84, RMSE=4.19 mg / kg, LCCC=0.91) and Zn (R 2 =0.94, RMSE=7.27 mg / kg, LCCC=0.97) had the highest accuracy, indicating that the established Cubist model was nearly perfect in predicting Pb and Zn.

[0040] The model performance evaluation results of the modeling set showed that the SVM inversion model showed high prediction accuracy in the test set evaluation results, predicting soil As (R 2=0.64, RMSE=4.06 mg / kg) and Cu (R 2 =0.95, RMSE=8.04 mg / kg) content is more accurate than the other three models. The GPR inversion model predicts soil Pb (R 2 =0.89, RMSE=3.69 mg / kg) content has the highest accuracy, while the Cubist inversion model predicts Zn (R 2 =0.97, RMSE=5.46 mg / kg) had the highest accuracy.

[0041] The present invention has been disclosed above with preferred embodiments, but they are not intended to limit the present invention. Any technical solutions obtained by adopting equivalent replacement or equivalent transformation solutions fall within the protection scope of the present invention.

Claims

1. A method for inverting the content of heavy metals in soil using the spectrum of the original soil profile, characterized in that: The following steps are involved: S1 Soil profile sample pretreatment: Collect several soil profile samples in the test area, and divide each soil profile sample into several continuous soil layers from the surface to the inside; S2 Effective hyperspectral detection: N measurement points are evenly selected on each soil layer sample, and the spectrum measurement is repeated 5 times at each measurement point to obtain 40 spectrum curves. The average value of the 40 spectrum data is calculated as the effective hyperspectral data of the soil layer; S3 Detection of standard content values ​​of heavy metals: Conduct at least four heavy metal content tests on each soil layer to obtain the standard content value of each heavy metal; S4 Inversion of heavy metal content in unknown soil samples: Firstly, using effective hyperspectral data as the prediction variable and the standard content value of heavy metals as the corresponding variable, several hyperspectral inversion models were established using at least one machine learning algorithm. Based on the evaluation index, the optimal hyperspectral inversion model for each soil heavy metal element was determined. Secondly, the heavy metal content in unknown soil samples is inverted based on the optimal hyperspectral inversion model.

2. The method for inverting the heavy metal content in soil using the original soil profile spectrum according to claim 1, characterized in that: The specific process of constructing the hyperspectral inversion model is as follows: S41 Modeling set and test set division: The principal component analysis method is used to detect the uniformity of effective hyperspectral data, and the Hotelling statistic test method is used to identify and eliminate abnormal effective hyperspectral data; the Kennard-Stone algorithm is used to divide the retained effective hyperspectral data into a modeling set and a test set, where the ratio of the modeling set to the test set is 7:3; S42 Determine the best parameter combination: Use grid search and ten-fold cross validation method to optimize the parameters of each machine learning algorithm, that is, each algorithm is given an initial value range and a search interval to obtain a corresponding parameter combination; for each heavy metal element, use grid search and ten-fold cross validation method to screen all parameter combinations to obtain the corresponding best parameter combination; S43 is based on the optimal parameter combination, with the effective hyperspectral data as the prediction variable and the standard content value of heavy metals as the response variable, to establish a mapping relationship between the effective hyperspectral data and the standard content values ​​of different heavy metal elements, that is, a hyperspectral inversion model.

3. The method for inverting the heavy metal content in soil using the original soil profile spectrum according to any one of claims 1 or 2, characterized in that: The machine learning algorithm includes at least: PLSR algorithm, Cubist algorithm, GPR algorithm and SVM algorithm; Correspondingly, the several hyperspectral inversion models are respectively a PLSR inversion model, a Cubist prediction model, a GPR prediction model and a SVM prediction model.

4. The method for inverting the heavy metal content in soil using the original soil profile spectrum according to claim 1, characterized in that: The evaluation index at least includes: determination coefficient, root mean square error and Lin consistency correlation coefficient; the screening process of the optimal hyperspectral inversion model is as follows: The evaluation values ​​of the standard content values ​​of the four heavy metals predicted by different hyperspectral inversion models were judged according to the following criteria: Where LCCC represents Lin's consistency correlation coefficient.

5. The method for inverting the heavy metal content in soil using the original soil profile spectrum according to claim 1, characterized in that: The four soil heavy metal elements are: arsenic (As), copper (Cu), lead (Pb), and zinc (Zn).

6. The method for inverting the heavy metal content in soil using the original soil profile spectrum according to claim 3, characterized in that: The best parameter combination of the PLSR inversion model is the parameter combination corresponding to the root mean square error value of the ten-fold cross validation no longer changes significantly; The best parameter combination of the Cubist inversion model, the GPR inversion model and the SVM inversion model is the parameter combination corresponding to the minimum root mean square error value of the ten-fold cross validation.

7. The method for inverting the heavy metal content in soil using the original soil profile spectrum according to claim 1, characterized in that: Step S2 also includes preprocessing of effective hyperspectral data, and the specific process is as follows: one or more combinations of apparent absorbance, first-order derivative, second-order derivative, Savitzky-Golay smoothing, Gap-Segment derivative, detrending, standard normal variable transformation, and continuum removal processing methods.

Images