A high-precision spatial prediction method for soil heavy metal concentrations
By using multi-spectral hyperspectral remote sensing images and environmental co-ordinated variables of chlorophyll fluorescence intensity in induced by daylight, combined with environmental similarity calculation and weighted averaging method, the problem of low spatial prediction accuracy of soil heavy metal concentration is solved, and high in the case of a small number of samples is achieved Precision prediction.
Patent Information
- Application Number
- CN202210914663.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-08-01
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2042-08-01
AI Technical Summary
The existing spatial prediction methods for soil heavy metal concentrations have problems such as low accuracy, dependence on data integrity and sufficient sample size, especially in small areas, which cannot truly reflect the spatial changes in heavy metal concentration.
Multi-spectral hyperspectral remote sensing images and sunlight-induced chlorophyll fluorescence intensity are used as environmental covariates, and high-precision spatial prediction of heavy metal concentration in soil is achieved through environmental similarity calculation and weighted averaging method.
The accuracy of heavy metal concentration prediction is improved, the dependence of ordinary interpolation methods is overcome, and high-precision spatial prediction can be achieved with a small number of samples in the region and unevenly distributed.
Smart Images

Figure CN115128053B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of ecological environment protection, and in particular to a high-precision spatial prediction method for soil heavy metal concentration. Background Art
[0002] The spatiotemporal variation of soil heavy metal pollution is mainly affected and controlled by two types of factors: 1) geological factors. Under natural conditions, the soil heavy metal background pollution caused by geological factors is relatively low; 2) human activities. The changes in heavy metal concentration directly reflect the intensity and degree of human activities, and the heavy metal concentration changes are relatively large. The changes in heavy metal concentration are affected by the ionic state of heavy metals themselves, the interaction (adsorption, migration, exchange) with other substances (soil, plant roots, microorganisms), and other factors. However, no matter what the source of heavy metal pollution is and what the exchange rate with other substances is, there are three characteristics in the spatial variation: 1) The closer the spatial position is, the closer the heavy metal concentration is (spatial autocorrelation); 2) When the distance between spatial positions reaches a certain threshold, the heavy metal concentration varies greatly (spatial heterogeneity); 3) Different sources of heavy metals can be characterized by relevant geographical elements, such as the parent rock type can be indicated by the lithology distribution in geological data.
[0003] Summarizing the above characteristics, the spatial estimation methods of soil heavy metal pollution can be divided into three categories: 1) Spatial autocorrelation method, that is, the soil heavy metal concentration is based on the principle that the closer the geographical location, the more relevant it is. It relies on the data information of the experimental sample points in the regional space and is realized through interpolation. This type of method simply uses the known sample information and infers the unknown point information based on the geographical distance. It has strict requirements on the quantity, quality and spatial distribution of the sample points, and does not consider other influencing factors, so the accuracy is low. 2) Based on the factor correlation method, that is, a comprehensive analysis of various geographical variables that affect the heavy metal concentration, and a mapping relationship between heavy metal concentration and related factors is established through a mathematical model, which is finally used for the prediction of unknown areas. This type of method considers the impact of related geographical variables on heavy metal concentration, and achieves higher accuracy through data-driven machine learning methods, but has the disadvantage of being heavily dependent on the accuracy and completeness of the input data set. 3) A method that combines spatial autocorrelation and factor correlation, and applies geographical distance and other related geographical variables to spatial estimation at the same time. This method effectively utilizes the advantages of the above two methods and achieves complementary advantages, but there are still two shortcomings: 1) The geographical variables used are highly spatially similar in a small area, while the heavy metal concentrations are highly spatially heterogeneous, resulting in the spatial variation of geographical variables not being able to truly reflect the spatial variation of heavy metal concentrations; 2) Relying on data-driven algorithms to improve accuracy cannot effectively quantify the degree of heavy metal pollution. In view of the above shortcomings, it is urgent to find a better method for the spatial prediction of heavy metal concentrations. Summary of the invention
[0004] In order to solve the above technical problems, the present invention provides a high-precision spatial prediction method for soil heavy metal concentration with simple algorithm and wide application range.
[0005] The technical solution of the present invention to solve the above technical problem is: a high-precision spatial prediction method for soil heavy metal concentration, comprising the following steps:
[0006] Step 1: Data acquisition: including field sampling in the region, preparation of multispectral and hyperspectral remote sensing images, and acquisition of geographic environmental variables;
[0007] Step 2: Data preprocessing: laboratory testing and routine statistical analysis of field samples to detect abnormal heavy metal locations; radiometric correction, atmospheric correction and band selection of images are performed to calibrate the images for later use; geographic environmental variables are converted and transformed to unify coordinates, spatial resolution and data types;
[0008] Step 3: Data inversion and analysis: Calculate various vegetation indices and sunlight-induced chlorophyll fluorescence intensity, combine various geographic variables with sample analysis point data for geographic detection, and determine the environmental geographic synergistic variables involved in the prediction;
[0009] Step 4: Prediction of heavy metal concentration: Based on the obtained environmental geographic synergistic variables, geographic similarity calculation is performed, and combined with the heavy metal concentration values of known sampling points, the heavy metal concentration of unknown points is predicted to obtain the predicted spatial distribution of heavy metal concentration;
[0010] Step 5, accuracy estimation: calculate the accuracy of the predicted spatial distribution of heavy metal concentrations.
[0011] In the above-mentioned high-precision spatial prediction method for soil heavy metal concentration, in step 3, the sunlight-induced chlorophyll fluorescence intensity is obtained by inversion using the Fraunhofer dark line method;
[0012] SIF inversion algorithm FLD based on Fraunhofer dark lines: Due to the absorption of certain elements in the solar atmosphere, the irradiance spectrum of solar radiation forms many fine dark lines with a band width of 0.1-10nm after reaching the surface, namely Fraunhofer absorption dark lines. At the Fraunhofer absorption dark lines, the reflected light of vegetation is weak and the fluorescence effect is obvious. Therefore, the Fraunhofer dark lines are used to invert the sunlight-induced chlorophyll fluorescence value SIF;
[0013] Assuming that the earth's surface is isotropic, ignoring the atmospheric effect and proximity effect between canopy sensors, the surface apparent reflectance radiance received by the sensor is the sum of the solar reflectance radiance of the surface vegetation and the sunlight-induced chlorophyll fluorescence reflectance, as shown in formula (1):
[0014]
[0015] Where λ is the wavelength, L(λ) is the apparent reflectance radiance of the surface observed by the sensor, r(λ) is the surface reflectivity, E(λ) is the surface incident irradiance, and SIF(λ) is the chlorophyll fluorescence intensity emitted by vegetation;
[0016] According to the solar irradiance spectrum in the Fraunhofer absorption dark line range and the radiance spectrum observed by the sensor, it can be concluded that each parameter satisfies formula (2):
[0017]
[0018] Among them, L in is the wavelength λ in The radiance at L out is the wavelength λ out The radiance at I in is the wavelength λ in Solar irradiance at out is the wavelength λ out Solar irradiance at in is the wavelength λ in The sunlight-induced chlorophyll fluorescence intensity at out is the wavelength λ out The sunlight-induced chlorophyll fluorescence intensity at R in The wavelength is λ in The surface reflectivity at out is the wavelength λ out The surface reflectivity at
[0019] The standard FLD algorithm assumes that the vegetation reflectance and fluorescence intensity of the Fraunhofer absorption dark line and its adjacent spectral region do not change with wavelength within a very narrow band, that is:
[0020] R in =R out =R FLD f in =f out =f FLD (3)
[0021] Among them, R FLD is the surface reflectivity calculated using the FLD algorithm; f FLD is the sunlight-induced chlorophyll fluorescence intensity calculated using the FLD algorithm;
[0022] Combining equations (2) and (3), we can obtain:
[0023]
[0024] For the GF-5 hyperspectral image, the corresponding Fraunhofer absorption dark line band and its adjacent bands are selected.
[0025] The core of the above-mentioned high-precision spatial prediction method for soil heavy metal concentration is the relationship between the soil property value reflected by each individual soil sample point and its environment. Each individual sample point is used as a substitute for a location with similar environmental conditions, and is used to predict the soil property values of those locations that meet the requirements of environmental similarity with the sample point. It can be specifically summarized into three processes:
[0026] (1) Describe the environmental characteristics of each point; use multiple environmental covariates such as elevation, slope, aspect, and normalized difference moisture index (NDWI) 1240 , Normalized Differential Chlorophyll Index NPCI, Modified Simple Ratio Vegetation Index mSRI-2, Solar Induced Chlorophyll Fluorescence SIF and Normalized Red Edge Vegetation Index NDVI 705 Describe the environmental characteristics related to the target soil properties and form an environmental vector e at that location:
[0027] e=(e1,e2,e3,...,e m ) (5)
[0028] Among them, e1, e2, ..., e m They refer to the first, second, ..., mth environmental covariates considered respectively, and these m environmental covariates together form an m-dimensional environmental vector e;
[0029] (2) Environmental similarity assessment of each unknown point and each soil sample point;
[0030] Calculate the environmental similarity between the unknown point and the sample point:
[0031] S i,j =P(E(e 1i , e 1j ), E(e 2i , e 2j ), ..., E(e vi , e vj ), ..., E(e mi , e mj )) (6)
[0032] Among them, S i,j represents the environmental similarity between the unknown point i and the sample point j in the study area, e vi and e vj Respectively represent the properties of the vth environmental covariate at the i-th unknown point and the j-th sample point, e mi and e mjdenote the properties of the mth environmental covariate at the i-th unknown point and the j-th sample point, respectively. E(·) and P(·) are functions at the environmental covariate scale and soil sample scale, respectively. Function E(·) is used to calculate the environmental similarity at the environmental covariate scale. For a given m environmental covariates, there are environmental similarities from the first environmental covariate to the mth environmental covariate, expressed as E(e 1i , e 1j )、E(e 2i , e 2j )、......、E(e mi , e mj ), E(·) describes how the environmental similarity between the unknown point and the sample point changes when the environmental covariate value of the unknown point deviates from the environmental covariate value of the sample point; based on this, E(·) is defined as:
[0033]
[0034] in, is the standard deviation of the vth environmental covariate in the study area, is the square root of the average deviation of the vth environmental covariate between the unknown point i and the sample point j;
[0035]
[0036] In formula (8), k is the number of unknown points;
[0037] Function P(·) is used to integrate the environmental similarity of the environmental covariate scale and the soil sample scale, using the minimum operator method, that is, the minimum value of all environmental covariate similarities is taken as the environmental similarity at the soil sample level;
[0038] The environmental similarity between each unknown point and all soil sample points is organized into an environmental similarity vector S for each unknown point. i :
[0039] S i =(S i,1 , S i,2 , ..., S i,j ) (9)
[0040] (3) Quantification of prediction uncertainty at each unknown point and estimation of target soil property values;
[0041] If the environmental similarity between the unknown point and the sample points used does not meet the requirements, the existing sample points cannot be used to represent the unknown point. Therefore, it is necessary to calculate the prediction uncertainty to evaluate the reliability of estimating the unknown point using the existing sample points. The uncertainty is inversely proportional to the environmental similarity. The uncertainty calculation method is shown in formula (10):
[0042] Uncertainty i =1-max(S i,1 , S i,2 , ..., S i,n ) (10)
[0043] Among them, Uncertainty i is the uncertainty of unknown point i, n is the number of unknown points used to estimate the target soil property value, max(S i,1 , S i,2 , ..., S i,n ) represents the similarity S i,1 , S i,2 ,......,S i,n The maximum value of
[0044] Set a user-defined uncertainty threshold; if the uncertainty of an unknown point is greater than the user-defined uncertainty threshold, the unknown point is set to an invalid value; if the uncertainty of the unknown point is less than or equal to the user-defined uncertainty threshold, the weighted average method is used to estimate the target attribute value of the unknown point, as shown in formula (11):
[0045]
[0046] Among them, C i is the predicted soil heavy metal concentration of unknown point i, n′ is the number of selected sample points, and its environmental similarity with unknown point i is greater than “1-uncertainty threshold”. Generally, the uncertainty threshold can be set to 1% to 30%. This study selects 30% as the uncertainty threshold; C j is the soil heavy metal concentration at sample point j.
[0047] In the above high-precision spatial prediction method for soil heavy metal concentration, in step 5, the prediction accuracy of the heavy metal concentration at the unknown point i is calculated:
[0048] Assume that the heavy metal concentration at sample point j is C j , the geographical similarity of sample point j is S j , the geographical similarity of unknown point i is S i , C i is the predicted concentration of soil heavy metals at unknown point i, then the similarity error E of the unknown point i As shown in formula (12):
[0049] E i =1-S i (12).
[0050] The beneficial effects of the present invention are as follows: the present invention introduces SIF which is sensitive to heavy metal concentration as an environmental synergistic variable, which can greatly improve the prediction accuracy of heavy metal concentration, and can overcome the dependence of ordinary interpolation methods on the average representativeness of the overall samples, and effectively utilize the representativeness of individual sample points to overcome the limitations of sufficient sample quantity and uniform distribution, and can also achieve high-precision spatial prediction for a small number of samples and uneven distribution in the region. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 It is a flow chart of the present invention.
[0052] Figure 2 Schematic diagram of the sunlight-induced chlorophyll fluorescence inversion principle.
[0053] Figure 3 This is the distribution map of soil heavy metal sampling points in the experimental area.
[0054] Figure 4 This is the predicted concentration and error spatial distribution map of soil heavy metal cadmium and lead based on the similarity method. DETAILED DESCRIPTION
[0055] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0056] like Figure 1 As shown, a high-precision spatial prediction method for soil heavy metal concentration includes the following steps:
[0057] Step 1: Data acquisition: including field sampling in the area, preparation of multispectral and hyperspectral remote sensing images, and acquisition of geographic environmental variables.
[0058] Step 2: Data preprocessing: laboratory testing and routine statistical analysis of field samples to detect abnormal heavy metal locations; radiation correction, atmospheric correction and band selection of images to obtain corrected images for use; geographic environmental variables are converted and transformed to unify coordinates, spatial resolution and data types.
[0059] Step 3: Data inversion and analysis: Calculate various vegetation indices and sunlight-induced chlorophyll fluorescence intensity, combine various geographic variables with sample analysis point data for geographic detection, and determine the environmental geographic synergistic variables involved in the prediction.
[0060] The Fraunhofer dark line method was used to invert the sunlight-induced chlorophyll fluorescence intensity.
[0061] SIF inversion algorithm FLD based on Fraunhofer dark lines: Due to the absorption of certain elements in the solar atmosphere, the irradiance spectrum of solar radiation forms many fine dark lines with a band width of 0.1-10nm after reaching the surface, namely Fraunhofer absorption dark lines. At the Fraunhofer absorption dark lines, the reflected light of vegetation is weak and the fluorescence effect is obvious. Therefore, the Fraunhofer dark lines are used to invert the sunlight-induced chlorophyll fluorescence value SIF;
[0062] Assuming that the earth's surface is isotropic, ignoring the atmospheric effect and proximity effect between canopy sensors, the surface apparent reflectance radiance received by the sensor is the sum of the solar reflectance radiance of the surface vegetation and the sunlight-induced chlorophyll fluorescence reflectance, as shown in formula (1):
[0063]
[0064] Where λ is the wavelength, L(λ) is the apparent reflectance radiance of the surface observed by the sensor, r(λ) is the surface reflectivity, E(λ) is the surface incident irradiance, and SIF(λ) is the chlorophyll fluorescence intensity emitted by vegetation.
[0065] According to the solar irradiance spectrum in the Fraunhofer absorption dark line range and the radiance spectrum observed by the sensor, it can be concluded that each parameter satisfies formula (2):
[0066]
[0067] Among them, L in is the wavelength λ in The radiance at L out is the wavelength λ out The radiance at I in is the wavelength λ in Solar irradiance at out is the wavelength λ out Solar irradiance at in is the wavelength λ in The sunlight-induced chlorophyll fluorescence intensity at out is the wavelength λ out The sunlight-induced chlorophyll fluorescence intensity at R in The wavelength is λ in The surface reflectivity at out is the wavelength λ out The surface reflectivity at .
[0068] The standard FLD algorithm assumes that the vegetation reflectance and fluorescence intensity of the Fraunhofer absorption dark line and its adjacent spectral region do not change with wavelength within a very narrow band in the present invention, that is:
[0069] R in =R out =RFLD f in =f out =f FLD (3)
[0070] Among them, R FLD is the surface reflectivity calculated using the FLD algorithm; f FLD is the sunlight-induced chlorophyll fluorescence intensity calculated using the FLD algorithm.
[0071] Combining equations (2) and (3), we can obtain:
[0072]
[0073] For the GF-5 hyperspectral image, the corresponding Fraunhofer absorption dark line band and its adjacent bands are selected.
[0074] Step 4: Prediction of heavy metal concentration: Based on the obtained environmental geographic synergistic variables, geographical similarity calculation is performed, and combined with the heavy metal concentration values of known sampling points, the heavy metal concentration of unknown points is predicted to obtain the predicted spatial distribution of heavy metal concentration.
[0075] The core of step 4 is the relationship between the soil property values reflected by each individual soil sample point and its environment. Each individual sample point is used as a substitute for a location with similar environmental conditions, and is used to predict the soil property values of locations that meet the requirements of environmental similarity with the sample point. This method is abbreviated as iPSM method; it can be summarized into three steps:
[0076] (1) Describe the environmental characteristics of each point; use multiple environmental covariates such as elevation, slope, aspect, and normalized difference moisture index (NDWI) 1240 , Normalized Differential Chlorophyll Index NPCI, Modified Simple Ratio Vegetation Index mSRI-2, Solar Induced Chlorophyll Fluorescence SIF and Normalized Red Edge Vegetation Index NDVI 705 Describe the environmental characteristics related to the target soil properties and form an environmental vector e at that location:
[0077] e=(e1,e2,e3,...,e m ) (5)
[0078] Among them, e1, e2, ..., e m They refer to the first, second, ..., mth environmental covariates considered respectively, and these m environmental covariates together form an m-dimensional environmental vector e.
[0079] (2) Environmental similarity assessment of each unknown point and each soil sample point;
[0080] Calculate the environmental similarity between the unknown point and the sample point:
[0081] S i,j =P(E(e 1i , e 1j ), E(e 2i , e 2j ), ..., E(e vi , e vj ), ..., E(e mi , e mj )) (6)
[0082] Among them, S i,j represents the environmental similarity between the unknown point i and the sample point j in the study area, e vi and e vj Respectively represent the properties of the vth environmental covariate at the i-th unknown point and the j-th sample point, e mi and e mj denote the properties of the mth environmental covariate at the i-th unknown point and the j-th sample point, respectively. E(·) and P(·) are functions at the environmental covariate scale and soil sample scale, respectively. Function E(·) is used to calculate the environmental similarity at the environmental covariate scale. For a given m environmental covariates, there are environmental similarities from the first environmental covariate to the mth environmental covariate, expressed as E(e 1i , e 1j )、E(e 2i , e 2j )、......、E(e mi , e mj ), E(·) describes how the environmental similarity between the unknown point and the sample point changes when the environmental covariate value of the unknown point deviates from the environmental covariate value of the sample point; based on this, E(·) is defined as:
[0083]
[0084] in, is the standard deviation of the vth environmental covariate in the study area, is the square root of the average deviation of the vth environmental covariate between the unknown point i and the sample point j;
[0085]
[0086] In formula (8), k is the number of unknown points;
[0087] Function P(·) is used to integrate the environmental similarity of the environmental covariate scale and the soil sample scale, using the minimum operator method, that is, the minimum value of all environmental covariate similarities is taken as the environmental similarity at the soil sample level;
[0088] The environmental similarity between each unknown point and all soil sample points is organized into an environmental similarity vector S for each unknown point. i :
[0089] S i =(S i,1 , S i,2 , ..., S i,j ) (9)
[0090] (3) Quantification of prediction uncertainty at each unknown point and estimation of target soil property values;
[0091] If the environmental similarity between the unknown point and the sample points used does not meet the requirements, the existing sample points cannot be used to represent the unknown point. Therefore, it is necessary to calculate the prediction uncertainty to evaluate the reliability of estimating the unknown point using the existing sample points. The uncertainty is inversely proportional to the environmental similarity. The uncertainty calculation method is shown in formula (10):
[0092] Uncertainty i =1-max(S i,1 , S i,2 , ..., S i,n ) (10)
[0093] Among them, Uncertainty i is the uncertainty of unknown point i, n is the number of unknown points used to estimate the target soil property value, max(S i,1 , S i,2 , ..., S i,n ) represents the similarity S i,1 , S i,2 ,......,S i,n The maximum value of .
[0094] Set a user-defined uncertainty threshold; if the uncertainty of an unknown point is greater than the user-defined uncertainty threshold, the unknown point is set to an invalid value; if the uncertainty of the unknown point is less than or equal to the user-defined uncertainty threshold, the weighted average method is used to estimate the target attribute value of the unknown point, as shown in formula (11):
[0095]
[0096] Among them, C i is the predicted soil heavy metal concentration of unknown point i, n′ is the number of selected sample points, and its environmental similarity with unknown point i is greater than “1-uncertainty threshold”. Generally, the uncertainty threshold can be set to 1% to 30%. This study selects 30% as the uncertainty threshold; C j is the soil heavy metal concentration at sample point j.
[0097] Step 5, accuracy estimation: calculate the accuracy of the predicted spatial distribution of heavy metal concentrations.
[0098] Assume that the heavy metal concentration at sample point j is C j , the geographical similarity of sample point j is S j , the geographical similarity of unknown point i is S i , C i is the predicted concentration of soil heavy metals at unknown point i, then the similarity error E of the unknown point i As shown in formula (12):
[0099] E i =1-S i (12).
[0100] Example
[0101] The digital elevation data (DEM) uses the 30-meter resolution data released by the National Aeronautics and Space Administration (NASA); the image data uses my country's National High-resolution No. 5 (GF-5) hyperspectral image, which was acquired on December 22, 2019. The specific image is: GF5_AHSI_E111.64_N29.83_20191222_008627_L10000068607.
[0102] Based on the above two types of data, eight environmental factors were extracted using the ArcGIS 10.8 software platform: elevation, slope, aspect, normalized difference water index (NDWI1240), normalized difference chlorophyll index (NPCI), modified simple ratio vegetation index mSRI-2, sunlight-induced chlorophyll fluorescence (SIF), and normalized red edge vegetation index (NDVI705). λ The spectral reflectance at wavelength λ is shown. The extracted environmental factors were tested for collinearity. If the variance expansion factor (VIF) between two environmental factors is less than 10, it means that the collinearity between the two environmental factors is weak and can be regarded as the best factor. Table 1 lists the variables used to construct the model for predicting the concentration of heavy metals cadmium and lead and the calculation formulas of some of the variables.
[0103] Table 1 Collinearity test of geographical covariates
[0104]
[0105] Note: Environmental co-variables with VIF values less than 10 can be used for spatial prediction of heavy metal concentrations.
[0106] Environmental covariates that pass the collinearity test are divided into two groups:
[0107] (1) The first group of environmental covariates includes: elevation, slope, aspect, NDWI, and mSRI-2;
[0108] (2) The second group of environmental covariates includes: elevation, slope, aspect, NDWI, mSRI-2, and SIF.
[0109] The difference between the above two groups of variables lies only in whether sunlight-induced chlorophyll fluorescence parameters are considered for predicting soil heavy metal concentrations, in order to verify the effect of SIF on improving the predicted concentrations of soil heavy metals.
[0110] iPSM models of cadmium and lead concentrations were constructed respectively to perform spatial prediction of heavy metal cadmium and lead concentrations in the entire region.
[0111] Randomly select any 3 sample points from the 14 sample points as the validation sample set, and the remaining 11 sample points are used for spatial prediction. In order to improve the accuracy of the validation model, the 14 sample points are randomly selected 3 times to obtain three different sets of data, and then the three selected different sets of data are used for spatial prediction and model validation.
[0112] In Table 2, Cd(variable group 1), Cd(variable group 2), Pb(variable group 1), and Pb(variable group 2) respectively represent the following meanings:
[0113] (1) Cd (variable group 1): a cadmium concentration prediction model constructed with the first group of environmental covariates as factors;
[0114] (2) Cd (variable group 2): a cadmium concentration prediction model constructed with the second group of environmental covariates as factors;
[0115] (3) Pb (variable group 1): lead concentration prediction model constructed with the first group of environmental covariates as factors;
[0116] (4) Pb (variable group 2): lead concentration prediction model constructed with the second group of environmental covariates as factors;
[0117] The model parameters are shown in the following table:
[0118] Comparing the determination coefficient (R2) and root mean square error (RMSE) of Cd (variable group 1) and Cd (variable group 2), it can be seen that the accuracy and stability of the cadmium concentration prediction model Cd (variable group 2) constructed by adding SIF environmental factors as environmental covariates are better than those of the Cd (variable group 1) model; similarly, comparing the determination coefficient (R2) and root mean square error (RMSE) of Pb (variable group 1) and Pb (variable group 2), the accuracy and stability of the Pb (variable group 2) model containing SIF environmental variables are better than those of the Pb (variable group 1) model without SIF environmental variables. However, the factor combination of adding SIF environmental factors has a more obvious effect on improving the accuracy and stability of the cadmium concentration iPSM model, while the improvement of the stability of the lead concentration iPSM model is not very obvious.
[0119] In addition, although the stability of the cadmium concentration iPSM model constructed without SIF environmental factors as environmental covariates is better than that of the lead concentration iPSM model, its accuracy is not as good as that of the lead concentration iPSM model. The cadmium concentration iPSM model with SIF environmental factors as environmental covariates is significantly better than that of the lead concentration iPSM model in terms of accuracy and stability. The specific results are shown in Table 2.
[0120] Table 2 Evaluation of the effect of predicting cadmium and lead concentrations based on different geographical covariate combinations
[0121]
[0122] Note: The geographic covariates included in combination 1 are: elevation, slope, aspect, normalized water index, ratio vegetation index; the geographic covariates included in combination 2 are: elevation, slope, aspect, normalized water index, ratio vegetation index, chlorophyll fluorescence intensity. The first, second and third groups represent the evaluation indicators obtained by randomly dividing the sampling points in the study area into three groups according to the prediction points and verification points for prediction and verification.
[0123] Model parameters are an indicator for evaluating model accuracy. At the same time, the accuracy of the prediction results of the iPSM model constructed by different factor combinations can also be analyzed through the effective prediction range. The iPSM model mainly determines whether the sample point data can be used to predict the unknown point by evaluating the environmental similarity between the unknown point and the sample point in the study area. The user-defined uncertainty threshold is an indicator to measure whether the environmental similarity between the unknown point and the sample point meets the requirements. In this study, the uncertainty threshold was set to 0.4. When the uncertainty of the unknown point is less than 0.4, it can be predicted by the sample point. The number of pixels in the study area is 93,570. As shown in Table 3, the prediction model established by the first group of environmental factor combinations, that is, without adding the SIF factor, has a prediction range of 46.73% for the entire region, while the prediction model constructed by adding the SIF environmental factor as the environmental variable has a prediction range of 29.82% for the entire region. The reduction in scope does not mean that the model precision and accuracy are reduced, but that the model constructed without SIF as an environmental covariate has a prediction uncertainty of less than 0.4 between about 46.73% of the areas in the entire study area and the sample points. After adding the SIF environmental covariate, only 29.82% of the areas in the entire region have a prediction uncertainty of less than 0.4 between the sample points. The remaining areas are not suitable for spatial prediction through sample data due to excessively high prediction uncertainty. This shows that the addition of the sunlight-induced chlorophyll fluorescence (SIF) environmental covariate makes the assessment of the environmental similarity between unknown points and sample points more rigorous and accurate, and at the same time confirms that the prediction results of the iPSM model constructed by adding the sunlight-induced chlorophyll fluorescence environmental covariate will be more reasonable and reliable.
[0124] Table 3 Comparison of the spatial range of heavy metal concentrations predicted based on different combinations of geographical covariates
[0125]
[0126] Description: Description: The geographic covariates included in combination 1 are: elevation, slope, aspect, normalized water index, ratio vegetation index; the geographic covariates included in combination 2 are: elevation, slope, aspect, normalized water index, ratio vegetation index, chlorophyll fluorescence intensity. The first, second and third groups represent the evaluation indicators obtained by randomly dividing the sampling points in the study area into three groups according to the prediction points and verification points for prediction and verification.
[0127] The study area is located to the north of Jinshi, Changde City, Hunan Province, with developed transportation. It has a mid-subtropical humid monsoon climate with cold winters and hot summers; it has the following climatic characteristics: hot and humid summers, cold and wet winters, and developed monsoons. The study area is divided into north and south by the Lishui River. The terrain slopes from south to north. North of the Lishui River is the Liyang Plain, the border area of the middle and lower reaches of the Yangtze River. It has a flat terrain and crisscrossing rivers and lakes. From the global 30-meter land cover data product, it is known that the study area is mainly composed of cultivated land and artificial surfaces. Cultivated land runs through the entire study area, accounting for about 68.5%, and artificial surfaces are built along the Lishui River, accounting for about 22%. The abundant land resources provide favorable conditions for agricultural development. Jinshi is rich in agricultural and sideline products such as grain, cotton, and oil. It has developed a number of production sites such as high-quality rice, cotton, oil, and fisheries. It was designated as a major comprehensive oil county in Hunan Province in 2020. After screening, such as Figure 3 As shown in the figure, the number of samples used to estimate the heavy metal concentrations in the study area is 14, distributed in the area south of the Lishui River.
[0128] The sample points were collected in the plains and cultivated areas of the region, where the terrain is relatively undulating and the altitude is within the range of 20-30 meters. After preliminary field investigations, on-site experiments, and sample sampling in the study area, 14 surface soil samples were obtained. Irregular random sampling was used in the study area, and each point in the spatial data formed included basic information such as geographic information, geological information, soil type, and the degree of human activity interference; the collection depth of soil samples was 0-30 cm. The heavy metal content in the soil solid samples was detected by acid digestion and quantitative analysis by plasma mass spectrometry / spectrometer, and the detection limit met the detection requirements of "GB15618-1995 / 2008 Soil Environmental Quality Standard" in units of μg / g. After inspection, the concentration data of 8 kinds of soil heavy metal elements at 14 sample points were obtained, and their spatial distribution Figure 3 shown.
[0129] Figure 4 a) and b) show that the concentrations of heavy metals cadmium and lead in the study area are generally low. High concentrations mainly appear in the plains north of the Lishui River, while low concentrations are mainly found in the south of the Lishui River. The areas where high concentrations of cadmium and lead appear are relatively consistent. Generally, places with high concentrations of heavy metal cadmium also have high concentrations of lead. In addition, according to the prediction results, there are more high concentrations of lead in the study area than high concentrations of cadmium, indicating that lead pollution in the study area is relatively more serious. Figure 4 c) shows that the uncertainty in rivers, lakes and other parts of the region is very high, indicating that the environmental similarity between these places and the sample points is very low (which is in line with the actual situation, as the sample points are all located in cultivated areas). Therefore, the iPSM model does not make spatial predictions on the heavy metal concentrations in these areas.
Claims
1. A high-precision spatial prediction method for soil heavy metal concentration, characterized in that: The following steps are involved: Step 1: Data acquisition: including field sampling in the region, preparation of multispectral and hyperspectral remote sensing images, and acquisition of geographic environmental variables; Step 2: Data preprocessing: laboratory testing and routine statistical analysis of field samples to detect abnormal heavy metal locations; radiometric correction, atmospheric correction and band selection of images are performed to calibrate the images for later use; geographic environmental variables are converted and transformed to unify coordinates, spatial resolution and data types; Step 3: Data inversion and analysis: Calculate various vegetation indices and sunlight-induced chlorophyll fluorescence intensity, combine various geographic variables with sample analysis point data for geographic detection, and determine the environmental geographic synergistic variables involved in the prediction; Step 4: Prediction of heavy metal concentration: Based on the obtained environmental geographic synergistic variables, geographic similarity calculation is performed, and combined with the heavy metal concentration values of known sampling points, the heavy metal concentration of unknown points is predicted to obtain the predicted spatial distribution of heavy metal concentration; Step 5, accuracy estimation: calculate the accuracy of the predicted spatial distribution of heavy metal concentrations; In the step 3, the sunlight-induced chlorophyll fluorescence intensity is obtained by using the Fraunhofer dark line method. SIF inversion algorithm FLD based on Fraunhofer dark lines: Due to the absorption of certain elements in the solar atmosphere, the irradiance spectrum of solar radiation forms many fine dark lines with a band width of 0.1-10nm after reaching the surface, namely Fraunhofer absorption dark lines. At the Fraunhofer absorption dark lines, the reflected light of vegetation is weak and the fluorescence effect is obvious. Therefore, the Fraunhofer dark lines are used to invert the sunlight-induced chlorophyll fluorescence value SIF; Assuming that the earth's surface is isotropic, ignoring the atmospheric effect and proximity effect between canopy sensors, the surface apparent reflectance radiance received by the sensor is the sum of the solar reflectance radiance of the surface vegetation and the sunlight-induced chlorophyll fluorescence reflectance, as shown in formula (1): Where λ is the wavelength, L(λ) is the apparent reflectance radiance of the surface observed by the sensor, r(λ) is the surface reflectivity, E(λ) is the surface incident irradiance, and SIF(λ) is the chlorophyll fluorescence intensity emitted by vegetation; According to the solar irradiance spectrum in the Fraunhofer absorption dark line range and the radiance spectrum observed by the sensor, it can be concluded that each parameter satisfies formula (2): Among them, L in is the wavelength λ in The radiance at L out is the wavelength λ out The radiance at I in is the wavelength λ in Solar irradiance at out is the wavelength λ out Solar irradiance at in is the wavelength λ in The sunlight-induced chlorophyll fluorescence intensity at out is the wavelength λ out The sunlight-induced chlorophyll fluorescence intensity at R in The wavelength is λ in The surface reflectivity at out is the wavelength λ out The surface reflectivity at The standard FLD algorithm assumes that the vegetation reflectance and fluorescence intensity of the Fraunhofer absorption dark line and its adjacent spectral region do not change with wavelength within a very narrow band, that is: R in =R out =R FLD f in =f out =f FLD (3) Among them, R FLD is the surface reflectivity calculated using the FLD algorithm; f FLD is the sunlight-induced chlorophyll fluorescence intensity calculated using the FLD algorithm; Combining equations (2) and (3), we can obtain: Select the corresponding Fraunhofer absorption dark line band and its adjacent bands for the GF-5 hyperspectral image; The core of step 4 is the relationship between the soil property value reflected by each individual soil sample point and its environment. Each individual sample point is used as a substitute for a location with similar environmental conditions, and is used to predict the soil property values of those locations that meet the requirements of environmental similarity with the sample point. It can be specifically summarized into three processes: (1) Describe the environmental characteristics of each point; use multiple environmental covariates such as elevation, slope, aspect, and normalized difference moisture index (NDWI) 1240 , Normalized Differential Chlorophyll Index NPCI, Modified Simple Ratio Vegetation Index mSRI-2, Solar Induced Chlorophyll Fluorescence SIF and Normalized Red Edge Vegetation Index NDVI 705 Describe the environmental characteristics related to the target soil properties and form an environmental vector e at that location: and=(e1,e2,e3,…,e m ) (5) Among them, e1, e2, ..., e m They refer to the first, second, ..., mth environmental covariates considered respectively, and these m environmental covariates together form an m-dimensional environmental vector e; (2) Environmental similarity assessment of each unknown point and each soil sample point; Calculate the environmental similarity between the unknown point and the sample point: S i,j =P(E(e1,e 1j ),Sign in 2i ,have been 2j ),...,Sign in vi ,have been vj ),...,Sign in mi ,have been mj )) (6) Among them, S i,j represents the environmental similarity between the unknown point i and the sample point j in the study area, e vi and e vj denote the properties of the vth environmental covariate at the i-th unknown point and the j-th sample point, respectively, e mi and e mj denote the properties of the mth environmental covariate at the i-th unknown point and the j-th sample point, respectively. E(·) and P(·) are functions at the environmental covariate scale and soil sample scale, respectively. Function E(·) is used to calculate the environmental similarity at the environmental covariate scale. For a given m environmental covariates, there are environmental similarities from the first environmental covariate to the mth environmental covariate, expressed as E(e 1i , e 1j )、E(e 2i , e 2j )、......、E(e mi , e mj ), E(·) describes how the environmental similarity between the unknown point and the sample point changes when the environmental covariate value of the unknown point deviates from the environmental covariate value of the sample point; based on this, E(·) is defined as: in, is the standard deviation of the vth environmental covariate in the study area, is the square root of the average deviation of the vth environmental covariate between the unknown point i and the sample point j; In formula (8), k is the number of unknown points; Function P(·) is used to integrate the environmental similarity of the environmental covariate scale and the soil sample scale, using the minimum operator method, that is, the minimum value of all environmental covariate similarities is taken as the environmental similarity at the soil sample level; The environmental similarity between each unknown point and all soil sample points is organized into an environmental similarity vector S for each unknown point. i : S i =(S i,1 ,S i,2 ,...,S i,j ) (9) (3) Quantification of prediction uncertainty at each unknown point and estimation of target soil property values; If the environmental similarity between the unknown point and the sample points used does not meet the requirements, the existing sample points cannot be used to represent the unknown point. Therefore, it is necessary to calculate the prediction uncertainty to evaluate the reliability of estimating the unknown point using the existing sample points. The uncertainty is inversely proportional to the environmental similarity. The uncertainty calculation method is shown in formula (10): Uncertainty i =1-max(S i,1 ,S i,2 ,...,S i,n ) (10) Among them, Uncertainty i is the uncertainty of unknown point i, n is the number of unknown points used to estimate the target soil property value, max(S i,1 , S i,2 , ..., S i,n ) represents the similarity S i,1 , S i,2 ,……,S i,n The maximum value of Set a user-defined uncertainty threshold; if the uncertainty of an unknown point is greater than the user-defined uncertainty threshold, the unknown point is set to an invalid value; if the uncertainty of the unknown point is less than or equal to the user-defined uncertainty threshold, the weighted average method is used to estimate the target attribute value of the unknown point, as shown in formula (11): Among them, C i is the predicted soil heavy metal concentration of unknown point i, n′ is the number of selected sample points, and its environmental similarity with unknown point i is greater than "1-uncertainty threshold", and 30% is selected as the uncertainty threshold; C j is the soil heavy metal concentration at sample point j.
2. The high-precision spatial prediction method for soil heavy metal concentration according to claim 1 is characterized in that: In step 5, the prediction accuracy of heavy metal concentration at unknown point i is calculated: Assume that the heavy metal concentration at sample point j is C j , the geographical similarity of sample point j is S j , the geographical similarity of unknown point i is S i , C i is the predicted concentration of soil heavy metals at unknown point i, then the similarity error E of the unknown point i As shown in formula (12): E i =1-S i (12)。
