Loess profile hyperspectral inversion method and system based on partial least square method

Abstract

The invention relates to the technical field of spectral inversion, and discloses a loess profile hyperspectral inversion method and system based on partial least squares, and the method comprises the following steps: utilizing a loess soil sample, and measuring the particle size of the sample and the content of total organic carbon and macroelements; the method comprises the following steps: collecting soil hyperspectral data, performing spectrum transformation processing, and determining a spectrum characteristic wave band; constructing a loess profile parameter inversion model based on a partial least square method, and performing training and performance evaluation; and performing hyperspectral inversion analysis on an unknown loess soil sample by using the trained and evaluated model. According to the method, the correlation among the spectral data, the loess profile granularity, the total organic carbon and the constant element number can be obtained, and the loess profile parameter inversion model based on the partial least square method is established by taking the selected wave band with a relatively large correlation coefficient as the characteristic wave band; and a new thought and a new technical means are provided for the remote sensing technology in the loess profile environment evolution research in the future.

Description

Technical Field

[0001] The present invention relates to the technical field of spectral inversion, and in particular to a method and system for inverting a loess profile hyperspectral spectrum based on a partial least squares method. Background Art

[0002] The occurrence and development process of loess itself records a wealth of historical information. Understanding the evolutionary characteristics and effects of loess properties and their relationship with changes in environmental conditions provides a basis for us to interpret environmental changes. The methods commonly used to restore the ancient environment of loess profiles include physical and chemical properties, paleontology, paleomagnetism, etc., but these methods have the disadvantages of long research cycles, time-consuming and labor-intensive, and high costs for testing and analyzing environmental indicators. In recent years, with the rapid development of computer and remote sensing technology, the extremely high spectral resolution of hyperspectral remote sensing has a good effect in the study of soil properties. Hyperspectral remote sensing can combine specific soils with bands, and conduct correlation analysis on the spectral characteristics and physical and chemical properties of different types of soils through spectral mathematical transformation, statistical analysis and other methods, and establish an inversion model of soil properties, so as to achieve rapid, quantitative, and low-cost monitoring of soil properties. At present, according to the subtle differences in soil parameters (organic matter content, water content, etc.), hyperspectral remote sensing is widely used in land resources, geological disasters, ecological environment restoration, precision agriculture and other fields. However, in the study of environmental evolution, it is rare to apply hyperspectral analysis of time series soil property changes to paleoenvironmental changes reflected by sedimentary profiles.

[0003] As the loess belt on the edge of the plain, it contains rich information on environmental changes and the evolution of the East Asian monsoon system, and records the succession cycles of the short-term regional paleoclimate in more detail, and plays an important role in the study of environmental evolution. In the study of the physical and chemical properties of loess, particle size is one of the important characteristics of loess sediments. It has the characteristics of simple measurement method, clear environmental indication significance, and sensitivity to climate change; and total organic carbon (TOC) is also a commonly used indicator in environmental evolution. It is the product of sediments under a certain climatic environment, and its content and properties are closely related to the soil-forming environment of the sediments. The geochemical element characteristics of loess can better reveal the changes in the sedimentary environment of the loess profile. In the sedimentary sequence of loess-paleosols, soil formation has a great relationship with the chemical elements in the profile. The chemical elements of loess will produce certain differentiation during the deposition process, and the paleosols are developed and deposited on the basis of this differentiation. Therefore, the present invention provides a hyperspectral inversion method and system for loess profiles based on partial least squares. Summary of the invention

[0004] In view of the problems in the related art, the present invention proposes a loess profile hyperspectral inversion method and system based on partial least squares method to overcome the above-mentioned technical problems existing in the existing related technology.

[0005] To this end, the specific technical solution adopted by the present invention is as follows:

[0006] According to one aspect of the present invention, a loess profile hyperspectral inversion method based on partial least squares method is provided, and the loess profile hyperspectral inversion method based on partial least squares method comprises the following steps:

[0007] The loess soil samples at different levels of the loess profile were collected by interval sampling method, and the particle size, total organic carbon and major element content of loess soil samples at different levels were determined;

[0008] Based on the ground feature spectrum collection specification, soil hyperspectral data of loess soil samples at different levels were collected, and spectral transformation was performed on the soil hyperspectral data. The correlation between the spectral data and the particle size, total organic carbon and major element content of the loess soil was analyzed to determine the spectral characteristic bands.

[0009] With the selected spectral characteristic bands as independent variables and the measured values ​​of loess profile environmental change indicators as dependent variables, a loess profile parameter inversion model based on partial least squares method was constructed and trained and its performance evaluated.

[0010] The loess profile parameter inversion model based on partial least squares method after training and evaluation was used to perform hyperspectral inversion analysis on unknown loess soil samples to predict the particle size, total organic carbon and major element contents of the unknown loess soil samples.

[0011] Furthermore, the method of collecting loess soil samples at different layers of the loess profile by using the interval sampling method includes:

[0012] During sampling, the top disturbance layer of preset thickness and the weathered surface of the loess profile are removed, loess soil samples are collected in the topsoil layer at a first preset interval, loess soil samples are collected in the loess layer and paleosol layer at a second preset interval, and loess soil samples are collected in the transition layer at a third preset interval.

[0013] Furthermore, the particle size of the loess soil sample was measured by a laser particle size analyzer, the total organic carbon was measured by the potassium dichromate method, and the content of major elements was measured by atomic absorption spectrophotometry, wherein the major elements included Fe2O3, CaO and MgO.

[0014] Furthermore, the collecting of soil hyperspectral data of loess soil samples at different levels based on the ground feature spectrum collection specification and performing spectral transformation processing on the soil hyperspectral data includes:

[0015] In a laboratory equivalent to a darkroom, a portable ground feature spectrometer is used to collect soil hyperspectral data of loess soil samples at different levels based on ground feature spectrum collection specifications, and the noise data in the soil hyperspectral data is removed. At the same time, the first-order differential, second-order differential, reciprocal logarithm and envelope-removing spectral transformation of the soil hyperspectral data are performed;

[0016] The spectral curve is drawn according to the soil hyperspectral data, and the jump points of the spectral curve are removed and the spectral curve is smoothed. The arithmetic average is performed on several spectral curves of each loess soil sample.

[0017] Furthermore, the soil hyperspectral data of loess soil samples at different levels collected based on the ground object spectrum collection specification also includes:

[0018] Before measurement, a reference white plate is used for calibration optimization to obtain the absolute reflectivity.

[0019] Furthermore, the correlation between the analysis spectral data and the particle size, total organic carbon and major element content of loess soil is determined to include:

[0020] The smoothed original soil hyperspectral data, first-order differential spectral transformation data, second-order differential spectral transformation data, de-envelope spectral transformation data and reciprocal logarithmic spectral transformation data were correlated with the average particle size, total organic carbon and major element content to determine the spectral characteristic bands.

[0021] Furthermore, the method of using the selected spectral characteristic band as an independent variable and the measured value of the loess profile environmental change index as a dependent variable to construct a loess profile parameter inversion model based on the partial least squares method and to perform training and performance evaluation includes:

[0022] The spectral characteristic bands of the original soil hyperspectral after smoothing, the first-order differential spectral transformation, the second-order differential spectral transformation, the envelope-removed spectral transformation and the reciprocal logarithmic spectral transformation were used as independent variables, and the measured values ​​of the loess profile environmental change indicators were used as dependent variables. The loess profile parameter inversion model based on the partial least squares method was constructed and trained.

[0023] The accuracy and prediction ability of the trained loess profile parameter inversion model based on partial least squares method were evaluated by combining the evaluation indicators of determination coefficient, root mean square error and relative analysis error.

[0024] Furthermore, the calculation formula of the determination coefficient is:

[0025]

[0026] The formula for calculating the root mean square error is:

[0027]

[0028] The calculation formula of relative analysis error is:

[0029] RPD=STDEV(Y i ) / RMSE

[0030] Where n is the number of samples; Y i is the measured value of sample i; X i is the predicted value of sample i; is the average value of Xi; is the mean value of Yi; STDEV is the standard deviation.

[0031] According to another aspect of the present invention, a loess profile hyperspectral inversion system based on partial least squares method is provided, and the loess profile hyperspectral inversion system based on partial least squares method comprises:

[0032] The sample collection and measurement module is used to collect loess soil samples at different levels of the loess profile using the interval sampling method, and to measure the particle size, total organic carbon and major element content of loess soil samples at different levels;

[0033] The module for determining the spectral characteristic bands is used to collect soil hyperspectral data of loess soil samples at different levels based on the ground feature spectrum collection specifications, and to perform spectral transformation on the soil hyperspectral data, analyze the correlation between the spectral data and the particle size, total organic carbon and major element content of the loess soil, and determine the spectral characteristic bands;

[0034] The inversion model building module is used to build a loess profile parameter inversion model based on partial least squares method, with the selected spectral characteristic bands as independent variables and the measured values ​​of loess profile environmental change indicators as dependent variables, and to conduct training and performance evaluation;

[0035] The hyperspectral inversion analysis module is used to perform hyperspectral inversion analysis on unknown loess soil samples using the loess profile parameter inversion model based on partial least squares after training and evaluation, and to predict the particle size, total organic carbon and major element content of the unknown loess soil samples.

[0036] Furthermore, the spectral characteristic band determination module includes a spectral transformation processing module and a correlation analysis module;

[0037] The spectral transformation processing module is used to collect soil hyperspectral data of loess soil samples at different levels in a laboratory equivalent to a darkroom using a portable ground object spectrometer based on ground object spectrum collection specifications, and remove noise data in the soil hyperspectral data, and perform spectral transformation of the first-order differential, second-order differential, reciprocal logarithm and envelope removal on the soil hyperspectral data; it is also used to draw a spectral curve based on the soil hyperspectral data, remove jump points of the spectral curve, smooth the spectral curve, and perform arithmetic averaging on several spectral curves of each loess soil sample;

[0038] The correlation analysis module is used to perform correlation analysis on the smoothed original soil hyperspectral data, first-order differential spectral transformation data, second-order differential spectral transformation data, de-envelope spectral transformation data and reciprocal logarithmic spectral transformation data with the average particle size, total organic carbon and constant element content to determine the spectral characteristic bands.

[0039] The beneficial effects of the present invention are as follows: the present invention takes loess profile as the research object, combines hyperspectral technology, and conducts correlation analysis on the original spectrum after smoothing, the first-order differential, the second-order differential, the envelope-removed line and the reciprocal logarithm with the particle size, total organic carbon and the number of constant elements of the loess profile, thereby obtaining the correlation between the spectral data and the particle size, total organic carbon and the number of constant elements of the loess profile, so that a loess profile parameter inversion model based on the partial least squares method can be established according to the selected bands with larger correlation coefficients as characteristic bands, providing new ideas and technical means for future remote sensing technology in the study of loess profile environmental evolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0041] Figure 1 The present invention is a flowchart of a method for hyperspectral inversion of loess profiles based on partial least squares method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0042] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention and are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these contents, ordinary technicians in the field should be able to understand other possible implementations and advantages of the present invention. The components in the figures are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0043] According to an embodiment of the present invention, a method and system for hyperspectral inversion of loess profile based on partial least squares method are provided.

[0044] The present invention is further described with reference to the accompanying drawings and specific embodiments. Figure 1 As shown, according to one embodiment of the present invention, a loess profile hyperspectral inversion method based on partial least squares method is provided, and the loess profile hyperspectral inversion method based on partial least squares method comprises the following steps:

[0045] S101. Collect loess soil samples at different layers of the loess profile using an interval sampling method, and determine the particle size, total organic carbon and major element content of the loess soil samples at different layers;

[0046] Specifically, in this embodiment, the loess in Zaoshugou, Mangshan, Zhengzhou is used to illustrate as follows:

[0047] The geographical coordinates of the Zaoshugou loess profile in Mangshan, Zhengzhou are: N34°57′5.76″, E113°20′59.03″, with an altitude of 260m. The profile is exposed 2.3m above the ground surface and is divided into 5 layers from bottom to top: Layer ①, 0-20cm from the ground surface, is the topsoil cultivation layer; Layer ②, 20-50cm from the ground surface, is gray-yellow silt soil with many roots and worm holes, and the upper part is 14C dating to 1440±30a.BP; Layer ③, 50-140cm from the ground surface, is light brown-yellow clay silt soil, with white mycelium developed in the lower part, which is the ancient soil layer. The 14C dating of the bottom is 4600±30a.BP, and the top is 2810±30a.BP; Layer ④, 140-170cm from the ground surface, gray-yellow silt sand, with scattered white mycelium, the transition layer between paleosol and loess, the bottom 14C dating is 5400±150a.BP; Layer ⑤, 170-230cm from the ground surface, no bottom is visible, gray-yellow silt sand.

[0048] The loess profile of Zaoshugou in Mangshan, Zhengzhou is 2.3m high. After removing the 20cm top disturbed layer, the actual sampling depth is 2.1m. Loess is loose and porous, and has strong collapsibility. If the sampling interval is too dense, it is easy for the adjacent sampling layers to collapse and be contaminated. Therefore, when sampling, the weathered surface of the profile is removed, and samples are collected at intervals of 3 to 5cm on the fresh surface. The loess and paleosol layers are sampled at intervals of 3-4cm. Because of the presence of calcified limestone in the transition layer, the interval is 4-5cm. A total of 47 samples were collected from the entire profile, including the topsoil layer T S 3 samples, loess layer L 0-1 Samples 7, paleosol layer S 0-1 Samples 19, transition layer L t 6 samples, loess layer L 0-2There are 11 samples. The collected samples are divided into two parts, one for total organic carbon and particle size determination, and the other for spectral experiments. Some samples of total organic carbon were contaminated during the test and were not used in data analysis.

[0049] The particle size test uses Microtrac S3500 laser particle size analyzer, which has a measurement range of 0.25-1500μm and a sample measurement accuracy of 0.6%. The main factor in the change of loess particle size is the change in the intensity of the winter wind, which is often used as an alternative indicator to indicate the intensity of the winter wind change. The particle size parameter average particle size can be used to indicate the concentration trend of the sediment particle size distribution, which reflects the changes in the sediment environment such as the original particle size composition of the sediment source and the average kinetic energy of the transport medium.

[0050] Total organic carbon (TOC) is the product of sediments under certain climatic conditions, and its content and properties are closely related to the soil-forming environment of the sediments. The change of total organic carbon content in the loess-paleosol continuous sedimentary profile can not only indicate the enrichment degree of humus in the loess sediments, but also reflect the environmental conditions of the loess during deposition. The total organic carbon test in this embodiment adopts the traditional potassium dichromate method.

[0051] Changes in climatic conditions are the biggest factor affecting the geochemical characteristics of major elements. Therefore, in the loess-paleosol sedimentary sequence, soil formation has a great relationship with the chemical elements in the profile. CaO and MgO are characteristic elements of arid climates. Among the major elements, Ca is the most mobile element and is very active during chemical weathering, which means it has the strongest migration ability. The level of calcium content in sediments is obviously controlled by climatic conditions, mainly water conditions. Therefore, climate changes can be inverted based on changes in calcium content in sediments. Magnesium is also one of the most active elements in the surface environment, but its activity is lower than that of calcium, but it is closely related to calcium. The content of magnesium in sediments is closely related to weathering. When weathering is strong, magnesium is easily leached, and when weathering is weak, magnesium is relatively enriched. Fe is a variable valence element with extremely strong chemical activity. Changes in the iron oxide content in loess can reveal climate changes. When Fe 3+ A higher content indicates a high temperature climate and a stronger oxidation. Therefore, this embodiment selects three constant elements, Fe2O3, CaO, and MgO, as geodetic elements reflecting the climate environment, and atomic absorption spectrophotometry is used to perform element determination of Fe2O3, CaO, and MgO in the Environmental Standard Laboratory of the Institute of Geography, Henan Academy of Sciences.

[0052] S102, collecting soil hyperspectral data of loess soil samples at different levels based on the ground feature spectrum collection specification, and performing spectral transformation processing on the soil hyperspectral data, analyzing the correlation between the spectral data and the particle size, total organic carbon and major element content of the loess soil, and determining the spectral characteristic bands;

[0053] Specifically, the ASD FieldSpec 4Standard-Res portable ground object spectrometer is used to collect soil hyperspectral data. The measurement range of the ground object spectrometer is 350-2500nm, the spectral sampling interval is 1.4nm when it is 350-1000nm, and the sampling interval is 2nm when it is 1000-2500nm. Its spectral resolution is 3nm@700nm; 10nm@1400nm; 10nm@2100nm.

[0054] Soil spectral data were collected under the ground feature spectrum collection specifications, and the spectral data used were measured in a laboratory equivalent to a darkroom. Before measurement, a reference white board was used for calibration optimization to obtain absolute reflectance. The aluminum box containing the sample was placed on black rubber with a thickness of 3 cm and a reflectivity of approximately 0. A standard DC tin wire quartz halogen lamp was used as the light source. The probe field angle was 25° and the probe distance was 15 cm. In order to reduce the influence of the spectral anisotropy of the soil sample, the sample box was rotated 3 times during the measurement, each time by 90°. Each sample was measured in 4 directions, and 5 spectral curves were collected in each direction. A total of 20 spectral curves were collected. The 20 spectral curves of each sample were arithmetic averaged for subsequent data processing and analysis.

[0055] During the spectrum measurement, the data with large noise influence at the beginning and end of 350-400nm and 2400-2500nm are removed, and then the jump points of the spectrum curve are removed and the spectrum curve is smoothed. Hyperspectral data has a large amount of information and is complex to process, making it difficult to use directly for analysis. Therefore, spectral transformations such as first derivate reflectance (FD), second derivate reflectance (SD), reciprocal logarithm (LOG(1 / R)), and continuum removal (CR) are performed on it.

[0056] The correlation analysis between particle size, total organic carbon and reflectance spectrum is as follows:

[0057] The five spectral transformation data, including original spectrum after smoothing of loess profile, first-order differential, second-order differential, envelope-removed spectrum and reciprocal logarithm, were correlated with the average particle size and TOC.

[0058] By comparing the correlation coefficients of spectral data after different forms of transformation (Table 1), it is found that compared with the number of significantly correlated bands of the original spectral reflectance after transformation, except for the number of significantly correlated bands of the spectral reflectance after the reciprocal logarithm transformation, the number of significantly correlated bands of the spectral reflectance of the other three spectral transformation forms has increased significantly, especially the number of significantly correlated bands of the spectrum after the envelope removal treatment. From the maximum correlation coefficient and its corresponding band, the maximum correlation coefficient of the average particle size of loess with the first-order differential, second-order differential, envelope removal and reciprocal logarithm spectrum has increased compared with the original spectrum, among which the maximum correlation coefficient of the second-order differential spectrum with the average particle size is 0.74, which is higher than the other transformed spectra, and the corresponding band is 762nm.

[0059] The correlation analysis results of the original spectrum after profile smoothing, the first-order differential, the second-order differential, the envelope-removed spectrum and the reciprocal logarithm of the five spectral transformation data and TOC show (Table 1) that the original spectrum and the spectrum after the reciprocal logarithm transformation have poor correlation with the TOC content, and the number of related bands is also small. From the maximum correlation coefficient and its corresponding band, the maximum correlation coefficient of the total organic carbon content with the first-order, second-order differential and envelope-removed spectrum is significantly improved compared with the original spectrum, among which the first-order differential spectrum and the envelope-removed spectrum have the highest correlation with the TOC content, and the correlation coefficients of both are -0.72, and the corresponding bands are 761nm and 952nm respectively.

[0060] After mathematical transformation, the original spectrum of soil can obviously highlight the spectral reflectance characteristics hidden in the soil and effectively enhance the correlation. Therefore, the relevant bands with a significance level of P=0.05 under four spectral transformation forms are selected as characteristic bands, and the confidence interval is selected as 95%. The significantly correlated bands of soil spectra are selected as independent variables to construct the hyperspectral inversion model of average particle size and TOC. That is, in the modeling of average particle size and soil spectral data, in addition to selecting 5 characteristic bands for modeling of original spectrum and average particle size, the 10 bands with the largest correlation are selected for the modeling characteristic bands of the four spectral transformations of the smoothed spectrum, namely, the first-order differential, second-order differential, de-envelope and inverse logarithm; in the modeling of TOC and soil spectral data, the 10 bands with the largest correlation are selected as the characteristic bands for modeling.

[0061] Table 1 Maximum values ​​and corresponding bands of correlation coefficients between different transformed spectral reflectances and average loess particle size and TOC

[0062]

[0063] The correlation analysis between constant elements and reflectance spectra is as follows:

[0064] The correlation analysis was performed between the original spectrum after smoothing, the first-order differential (FD), the second-order differential (SD), the de-envelope curve (CR) and the reciprocal logarithm (Log(1 / R)) and the constant elements Fe2O3, CaO, MgO and CaO / MgO in the loess profile (Table 2). From the maximum correlation coefficient and the corresponding band, compared with the original spectrum, the significant correlation between the loess spectral reflectance and the parameters of the constant elements of loess increased to varying degrees after the spectrum transformation. Among them, the second-order differential spectrum had the highest correlation with Fe2O3 and MgO, with the maximum correlation coefficients of 0.67 and 0.55, respectively, and the corresponding bands were 766nm and 904nm, respectively; the first-order differential spectrum had the highest correlation with CaO and CaO / MgO, and the two were similar, with the maximum correlation coefficients R of -0.75 and -0.74, respectively, and the corresponding bands were 661nm. After mathematical transformation, the original soil spectrum can obviously highlight the spectral reflectance characteristics hidden in the soil and effectively enhance the correlation. This embodiment uses a confidence interval of 95%, that is, a significance test at the P=0.05 level, and selects significantly correlated bands of the soil spectrum as independent variables for constructing an inversion model.

[0065] Table 2 Maximum values ​​and corresponding bands of correlation coefficients between different transformed spectral reflectances and profile constant elements

[0066]

[0067]

[0068] S103, using the selected spectral characteristic band as an independent variable and the measured value of the loess profile environmental change index as a dependent variable, constructing a loess profile parameter inversion model based on partial least squares method and performing training and performance evaluation;

[0069] Specifically, the partial least square regression (PLSR) method is a perfect combination of multiple linear regression, canonical correlation analysis and principal component analysis. PLSR can not only perform regression modeling under the conditions that there are multiple correlations in the independent variables and the number of sample points is less than the number of variables, but also effectively extract several comprehensive variables with the strongest system capabilities and exclude noise without explanatory effect, making the constructed model convincing. In the construction of the inversion model of average grain size and TOC, the distribution of the modeling set and prediction set of each stratigraphic unit is shown in Table 3, where the samples used for modeling are selected according to about 2 / 3 of the total number of samples in the stratigraphic unit, and the number of samples used for prediction is selected according to about 1 / 3 of the total number of samples in the stratigraphic unit. There are 32 samples for average grain size modeling and 12 samples for prediction; there are 15 samples for TOC modeling and 9 samples for prediction. The modeling samples are randomly selected in each stratigraphic unit according to the divided number, and the remaining number of samples is used for model prediction.

[0070] In this embodiment, the spectral characteristic bands of the environmental parameters of the average particle size and TOC after 5 spectral transformations are selected as independent variables, and the measured values ​​of the environmental change indicators of the loess profile are selected as dependent variables to construct an inversion model of the constant elements of the loess profile based on the partial least squares method. The independent variable is the soil spectral data, and there are 2001 bands from 400 to 2400 nm after removing the influence of the head and tail noise. In order to simplify the model, shorten the running time of the model and improve the accuracy of the model, and considering the collinearity problem of the spectral reflectance data, it is decided to select the spectral data with correlation for the characteristic band modeling and verification, and evaluate its accuracy.

[0071] Table 3 Statistical results of average particle size and TOC modeling samples

[0072]

[0073] In this embodiment, the spectral characteristic bands of environmental parameters such as Fe2O3, CaO, MgO and CaO / MgO after five spectral transformations are selected as independent variables, and the measured values ​​of loess profile environmental change indicators are selected as dependent variables to construct an inversion model of constant elements in loess profile based on partial least squares method.

[0074] Model accuracy is used to measure the comprehensive estimation ability of a model. The better the modeling accuracy, the better the stability of the model; the higher the verification accuracy, the stronger the prediction ability of the established model. 2 The accuracy and predictive ability of the established model are determined by combining three model evaluation indicators: root mean square error (RMSE) and relative percent deviation (RPD). The calculation formula of the determination coefficient is:

[0075]

[0076] The formula for calculating the root mean square error is:

[0077]

[0078] The calculation formula of relative analysis error is:

[0079] RPD=STDEV(Y i ) / RMSE

[0080] Where n is the number of samples; Y i is the measured value of sample i; X i is the predicted value of sample i; is the average value of Xi; is the mean value of Yi; STDEV is the standard deviation.

[0081] According to the calculation results, the model accuracy is divided into 4 levels: A, B, C, and D (as shown in Table 4).

[0082] Table 4 Model overall accuracy evaluation level standard

[0083]

[0084] Specifically, the spectral characteristic bands of the smoothed original spectrum, FD spectrum transformation, SD spectrum transformation, CR spectrum transformation, and LOG(1 / R) spectrum transformation were used as independent variables to construct the average particle size and TOC inversion model of loess profile based on partial least squares method, and the inversion model was constructed according to R 2 The three model evaluation indicators of RMSE and RPD were combined to determine the accuracy of the constructed model and its prediction ability (Table 5, Table 6).

[0085] Table 5 Modeling and verification of the partial least squares model of average particle size

[0086]

[0087] From the modeling results in Table 5, it can be seen that the modeling effect of the average particle size of the loess profile based on the partial least squares method is relatively ideal. Except for the original spectrum and the spectrum model after LOG(1 / R) spectrum transformation, which have poor accuracy and cannot achieve the prediction effect, the other three can achieve a rough prediction of the average particle size of the loess profile. 2 (0.73) and verification R 2 (0.78) are the maximum values ​​of the five spectral transformations, the modeling RMSE (0.26) and the verification RMSE (0.31) are both the lowest values ​​of the five spectral transformations, the relative analysis error PRD is 1.60, and the model accuracy level is B, which has a relatively accurate prediction ability.

[0088] Comparing the best fitting curves of measured and predicted values ​​with the straight line y=x, it can be seen that most samples are concentrated near the straight line y=x, and the correlation coefficient R between the measured and predicted values ​​of the validation set samples can reach the significance test at the P=0.05 level. Through analysis, it can be seen that the PLSR model with the spectrum after FD transformation as the independent variable is the best model for inverting the average particle size of the loess profile.

[0089] Table 6 Modeling and verification of the partial least squares model of TOC

[0090]

[0091] The modeling effect of TOC in loess profile based on partial least squares method is good overall (Table 6). Except that the original spectrum and the spectrum model transformed by CR and LOG (1 / R) have poor accuracy and cannot achieve the TOC prediction effect of loess profile, the prediction effects of the other two transformed spectrum models are relatively ideal. 2 (0.71) and verification R 2 (0.60), modeling RMSE (0.08) and verification RMSE (0.60), the relative analysis error PRD is 1.55, and the spectral modeling R after SD transformation 2 (0.84) and verification R 2 (0.71), modeling RMSE (0.06) and verification RMSE (0.09), the relative analysis error PRD is 1.59, and the accuracy level of the two transformed spectral models is B, which has a good prediction ability for loess profile TOC. Overall, the modeling effect of loess profile TOC based on partial least squares method is ideal, and it can accurately invert the content of loess profile TOC.

[0092] Comparing the best fitting curves of measured and predicted values ​​with the straight line y=x, it can be seen that most samples are concentrated near the straight line y=x, and the correlation coefficient R between the measured and predicted values ​​of the validation set samples has passed the significance test at the P=0.05 level. Through comparative analysis, it can be seen that among the five transformations, the fitting curves of the measured and predicted values ​​of the spectrum after FD and SD transformation are closest to the straight line y=x, and the fitting effects of the other three are poor, and none of them has the ability to invert. Therefore, the PLSR model with SD spectral transformation as the independent variable is selected as the best model for inverting TOC of loess profiles.

[0093] In the model construction of the constant elements in the loess profile, the spectral characteristic bands of environmental parameters such as Fe2O3, CaO, MgO and CaO / MgO after five spectral transformations were selected as independent variables to construct the inversion model of the constant elements in the loess profile based on the partial least squares method. Among them, the modeling samples of the four environmental parameters are 32 and the verification samples are 15.

[0094] Model inversion of Fe2O3:

[0095] According to R 2The accuracy and prediction ability of the Fe2O3 model were determined by using three model evaluation indicators: RMSE and RPD (Table 7). The modeling effect of Fe2O3 in loess profile based on partial least squares method is relatively general. Except for the spectral model transformed by FD, which has good accuracy and can achieve a relatively accurate prediction effect, the accuracy level of the spectral models of the other four transformations is D, and none of them can achieve the ability to predict Fe2O3 in loess profile, indicating that the correlation between Fe2O3 and loess reflectance spectrum is poor. Among them, the spectral modeling R transformed by FD has a good prediction effect. 2 (0.79) and verification R 2 (0.68) are the maximum values ​​of the five spectral transformations, the modeling RMSE (0.19) and the verification RMSE (0.27) are the lowest values ​​of the five spectral transformations, the relative analysis error PRD is 1.62, and the model accuracy level is B, which has a good prediction ability for Fe2O3 in loess profiles. Overall, the modeling effect of Fe2O3 in loess profiles based on partial least squares method is poor. After spectral transformation, the FD model has a good prediction ability. The other transformations cannot meet the prediction requirements and do not have the inversion ability of Fe2O3 in loess profiles.

[0096] Table 7 Modeling and verification of the partial least squares model of Fe2O3

[0097]

[0098] The best fitting curves of the measured and predicted values ​​of Fe2O3 are compared with the straight line y=x. It can be seen that most samples are concentrated near the straight line y=x, and the correlation coefficient R between the measured and predicted values ​​of the validation set samples has passed the significance test at the P=0.05 level. Through comparative analysis, it can be seen that the fitting curves of the measured and predicted values ​​of the spectrum after FD transformation are closest to the straight line y=x among the five transformations, and the fitting effects of the other four are poor, and none of them has the ability to invert. Therefore, the PLSR model with FD spectrum transformation as the independent variable is the best model for inverting Fe2O3 in loess profiles.

[0099] Model inversion of CaO, MgO and CaO / MgO

[0100] According to R 2 The three model evaluation indicators of , RMSE and RPD were combined to determine the accuracy and prediction ability of the models built for CaO, MgO and CaO / MgO. The modeling accuracy and results are shown in Table 8.

[0101] Table 8 Modeling and verification of partial least squares model of CaO, MgO and CaO / MgO

[0102]

[0103] The modeling effect of CaO and MgO in the loess profile based on partial least squares method is good. The spectral modeling of CaO after FD transformation is R 2 (0.68) and verification R 2 (0.68) are the maximum values ​​after spectral transformation, the modeling RMSE (0.75) and the verification RMSE (0.84) are the lowest values ​​after spectral transformation, the relative analysis error PRD is 1.61, the model accuracy level is B, and it has good CaO prediction ability in loess profiles; the spectral modeling R of MgO after CR transformation is 2 (0.67) and verification R 2 (0.63) are the maximum values ​​after spectral transformation, the modeling RMSE (0.06) and the verification RMSE (0.07) are the lowest values ​​after spectral transformation, the relative analysis error PRD is 1.80, the model accuracy level is B, and it has good prediction ability of MgO in loess profile; the spectral modeling R of CaO / MgO after FD transformation is 2 (0.66) and verification R 2 (0.57), modeling RMSE (0.33) and verification RMSE (0.84), the relative analysis error PRD is 1.41, the model accuracy level is C, and the prediction ability of CaO / MgO in loess profiles basically meets the requirements; the spectral model accuracy level of CaO, MgO and CaO / MgO after other transformations is D, which cannot achieve the ability to predict CaO, MgO and CaO / MgO in loess profiles, that is, it does not have the inversion ability of CaO, MgO and CaO / MgO in loess profiles.

[0104] Comparing the best fitting curves of measured and predicted values ​​with the straight line y=x, it can be seen that most samples are concentrated near the straight line y=x, and the correlation coefficient R between the measured and predicted values ​​of the validation set samples has passed the significance test at the P=0.05 level. Through comparative analysis, it can be seen that the fitting curves of the measured and predicted values ​​of the spectrum of CaO after FD transformation are closest to the straight line y=x; the fitting curves of the measured and predicted values ​​of the spectrum of MgO after CR transformation are closest to the straight line y=x; the fitting effects of the other several spectral transformations are poor, and none of them have the inversion ability. The correlation between CaO / MgO and the loess reflectance spectrum is poor. After the five spectral transformations, only the model accuracy level of the spectrum after FD transformation reaches C, and the others are poor. Therefore, the PLSR model with FD spectral transformation, CR spectral transformation and FD spectral transformation as independent variables is the best model for inverting CaO, MgO and CaO / MgO in the loess profile.

[0105] S104. Using the loess profile parameter inversion model based on partial least squares method after training and evaluation, a hyperspectral inversion analysis is performed on the unknown loess soil samples to predict the particle size, total organic carbon and major element content of the unknown loess soil samples.

[0106] According to another embodiment of the present invention, a loess profile hyperspectral inversion system based on partial least squares method is provided, and the loess profile hyperspectral inversion system based on partial least squares method includes:

[0107] The sample collection and measurement module is used to collect loess soil samples at different levels of the loess profile using the interval sampling method, and to measure the particle size, total organic carbon and major element content of loess soil samples at different levels;

[0108] The module for determining the spectral characteristic bands is used to collect soil hyperspectral data of loess soil samples at different levels based on the ground feature spectrum collection specifications, and to perform spectral transformation on the soil hyperspectral data, analyze the correlation between the spectral data and the particle size, total organic carbon and major element content of the loess soil, and determine the spectral characteristic bands;

[0109] The inversion model building module is used to build a loess profile parameter inversion model based on partial least squares method, with the selected spectral characteristic bands as independent variables and the measured values ​​of loess profile environmental change indicators as dependent variables, and to conduct training and performance evaluation;

[0110] The hyperspectral inversion analysis module is used to perform hyperspectral inversion analysis on unknown loess soil samples using the loess profile parameter inversion model based on partial least squares after training and evaluation, and to predict the particle size, total organic carbon and major element content of the unknown loess soil samples.

[0111] Wherein, the spectral characteristic band determination module includes a spectral transformation processing module and a correlation analysis module;

[0112] The spectral transformation processing module is used to collect soil hyperspectral data of loess soil samples at different levels in a laboratory equivalent to a darkroom using a portable ground object spectrometer based on ground object spectrum collection specifications, remove noise data in the soil hyperspectral data, and perform spectral transformation of the first-order differential, second-order differential, reciprocal logarithm and envelope removal on the soil hyperspectral data; it is also used to draw a spectral curve based on the soil hyperspectral data, remove jump points on the spectral curve, smooth the spectral curve, and perform arithmetic averaging on a number of spectral curves of each loess soil sample;

[0113] The correlation analysis module is used to perform correlation analysis on the smoothed original soil hyperspectral data, first-order differential spectral transformation data, second-order differential spectral transformation data, de-envelope spectral transformation data and reciprocal logarithmic spectral transformation data with the average particle size, total organic carbon and constant element content to determine the spectral characteristic bands.

[0114] To sum up, with the help of the above-mentioned technical scheme of the present invention, the present invention takes loess profile as the research object, combines hyperspectral technology, and conducts correlation analysis on the original spectrum after smoothing, the first-order differential, the second-order differential, the envelope-removed line and the reciprocal logarithm with the particle size, total organic carbon and the number of constant elements of the loess profile. The correlation between the spectral data and the particle size, total organic carbon and the number of constant elements of the loess profile is obtained, so that the loess profile parameter inversion model based on the partial least squares method can be established according to the selected band with a larger correlation coefficient as the characteristic band, which provides new ideas and technical means for future remote sensing technology in the study of loess profile environmental evolution.

[0115] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A loess profile hyperspectral inversion method based on partial least squares method, characterized in that: The loess profile hyperspectral inversion method based on partial least squares method includes the following steps: The loess soil samples at different levels of the loess profile were collected by interval sampling method, and the particle size, total organic carbon and major element content of loess soil samples at different levels were determined; Based on the ground feature spectrum collection specification, soil hyperspectral data of loess soil samples at different levels were collected, and spectral transformation was performed on the soil hyperspectral data. The correlation between the spectral data and the particle size, total organic carbon and major element content of the loess soil was analyzed to determine the spectral characteristic bands. With the selected spectral characteristic bands as independent variables and the measured values ​​of loess profile environmental change indicators as dependent variables, a loess profile parameter inversion model based on partial least squares method was constructed and trained and its performance evaluated. The loess profile parameter inversion model based on partial least squares method after training and evaluation was used to perform hyperspectral inversion analysis on unknown loess soil samples to predict the particle size, total organic carbon and major element contents of the unknown loess soil samples.

2. The method for hyperspectral inversion of loess profile based on partial least squares method according to claim 1 is characterized in that: The method of collecting loess soil samples at different levels of the loess profile by using the interval sampling method includes: During sampling, the top disturbance layer of preset thickness and the weathered surface of the loess profile are removed, loess soil samples are collected in the topsoil layer at a first preset interval, loess soil samples are collected in the loess layer and paleosol layer at a second preset interval, and loess soil samples are collected in the transition layer at a third preset interval.

3. The method for hyperspectral inversion of loess profile based on partial least squares method according to claim 1 is characterized in that: The particle size of loess soil samples was measured by a laser particle size analyzer, the total organic carbon was measured by the potassium dichromate method, and the content of major elements was measured by atomic absorption spectrophotometry, where the major elements included Fe2O3, CaO and MgO.

4. The method for hyperspectral inversion of loess profile based on partial least squares method according to claim 1 is characterized in that: The method of collecting soil hyperspectral data of loess soil samples at different levels based on the ground object spectrum collection specification and performing spectral transformation processing on the soil hyperspectral data includes: The soil hyperspectral data of loess soil samples at different levels were collected using a portable ground feature spectrometer based on the ground feature spectrum collection specification, and the noise data in the soil hyperspectral data was removed. At the same time, the soil hyperspectral data was transformed by the first-order differential, second-order differential, reciprocal logarithm and envelope removal. The spectral curve is drawn according to the soil hyperspectral data, and the jump points of the spectral curve are removed and the spectral curve is smoothed. The arithmetic average is performed on several spectral curves of each loess soil sample.

5. The method for hyperspectral inversion of loess profile based on partial least squares method according to claim 1 is characterized in that: The soil hyperspectral data of loess soil samples at different levels collected based on the ground object spectrum collection specification also includes: Before measurement, a reference white plate is used for calibration optimization to obtain the absolute reflectivity.

6. The method for hyperspectral inversion of loess profile based on partial least squares method according to claim 1, characterized in that: The correlation between the analyzed spectral data and the particle size, total organic carbon and major element content of loess soil is determined to include: The smoothed original soil hyperspectral data, first-order differential spectral transformation data, second-order differential spectral transformation data, de-envelope spectral transformation data and reciprocal logarithmic spectral transformation data were correlated with the average particle size, total organic carbon and major element content to determine the spectral characteristic bands.

7. The method for hyperspectral inversion of loess profile based on partial least squares method according to claim 1 is characterized in that: The method of using the selected spectral characteristic band as an independent variable and the measured value of the loess profile environmental change index as a dependent variable to construct a loess profile parameter inversion model based on the partial least squares method and to perform training and performance evaluation includes: The spectral characteristic bands of the original soil hyperspectral after smoothing, the first-order differential spectral transformation, the second-order differential spectral transformation, the envelope-removed spectral transformation and the reciprocal logarithmic spectral transformation were used as independent variables, and the measured values ​​of the loess profile environmental change indicators were used as dependent variables. The loess profile parameter inversion model based on the partial least squares method was constructed and trained. The accuracy and prediction ability of the trained loess profile parameter inversion model based on partial least squares method were evaluated by combining the evaluation indicators of determination coefficient, root mean square error and relative analysis error.

8. The method for hyperspectral inversion of loess profile based on partial least squares method according to claim 7 is characterized in that: The calculation formula of the determination coefficient is: The formula for calculating the root mean square error is: The calculation formula of relative analysis error is: RPD=STDEV(Y i ) / RMSE Where n is the number of samples; Y i is the measured value of sample i; X i is the predicted value of sample i; is the average value of Xi; Y i The mean value of ; STDEV is the standard deviation.

9. A loess profile hyperspectral inversion system based on partial least squares method, used to implement the steps of the loess profile hyperspectral inversion method based on partial least squares method described in any one of claims 1 to 8, characterized in that: The loess profile hyperspectral inversion system based on partial least squares method includes: The sample collection and measurement module is used to collect loess soil samples at different levels of the loess profile using the interval sampling method, and to measure the particle size, total organic carbon and major element content of loess soil samples at different levels; The module for determining the spectral characteristic bands is used to collect soil hyperspectral data of loess soil samples at different levels based on the ground feature spectrum collection specifications, and to perform spectral transformation on the soil hyperspectral data, analyze the correlation between the spectral data and the particle size, total organic carbon and major element content of the loess soil, and determine the spectral characteristic bands; The inversion model building module is used to build a loess profile parameter inversion model based on partial least squares method, with the selected spectral characteristic bands as independent variables and the measured values ​​of loess profile environmental change indicators as dependent variables, and to conduct training and performance evaluation; The hyperspectral inversion analysis module is used to perform hyperspectral inversion analysis on unknown loess soil samples using the loess profile parameter inversion model based on partial least squares after training and evaluation, and to predict the particle size, total organic carbon and major element content of the unknown loess soil samples.

10. The loess profile hyperspectral inversion system based on partial least squares method according to claim 9, characterized in that: The spectral feature band determination module includes a spectral transformation processing module and a correlation analysis module; The spectral transformation processing module is used to collect soil hyperspectral data of loess soil samples at different levels in a laboratory equivalent to a darkroom using a portable ground object spectrometer based on ground object spectrum collection specifications, and remove noise data in the soil hyperspectral data, and perform spectral transformation of the first-order differential, second-order differential, reciprocal logarithm and envelope removal on the soil hyperspectral data; it is also used to draw a spectral curve based on the soil hyperspectral data, remove jump points of the spectral curve, smooth the spectral curve, and perform arithmetic averaging on several spectral curves of each loess soil sample; The correlation analysis module is used to perform correlation analysis on the smoothed original soil hyperspectral data, first-order differential spectral transformation data, second-order differential spectral transformation data, de-envelope spectral transformation data and reciprocal logarithmic spectral transformation data with the average particle size, total organic carbon and constant element content to determine the spectral characteristic bands.

Images