A method for estimating nitrogen content in grassland vegetation leaves

Abstract

The invention discloses a method for estimating nitrogen content in grassland vegetation leaves, including collecting hyperspectral data of vegetation leaves for preprocessing to obtain hyperspectral reflectance data, and measuring nitrogen content in vegetation leaves; performing dimensionality reduction transformation on hyperspectral reflectance data, including constructing vegetation index, first-order derivative transformation, continuous wavelet transformation and / or continuum removal transformation; screening spectral characteristic parameters of hyperspectral reflectance data after dimensionality reduction transformation; constructing a grassland vegetation leaf nitrogen content estimation model based on various spectral characteristic parameters after screening, and evaluating the model accuracy in combination with the measured vegetation leaf nitrogen content, selecting the optimal model as the target grassland vegetation leaf nitrogen content estimation model to estimate nitrogen content in the vegetation leaf. The method is applicable to the estimation of nitrogen content in leaves of various vegetations, while reducing the monitoring cost, improving the objectivity of the monitoring results, and can quickly, non-destructively and accurately obtain vegetation nitrogen content.

Description

Technical Field

[0001] The invention relates to the technical field of leaf nitrogen content estimation, and in particular to a method for estimating the nitrogen content of grassland vegetation leaves. Background Art

[0002] In the quality evaluation system of grassland and forage, the quality of high-quality grassland and forage will directly affect the quality of livestock products. Nitrogen is generally considered to be a key nutrient element to support grassland growth and development. In order to maintain high-quality grassland and obtain higher-quality forage, it is usually necessary to apply nitrogen fertilizer to the grassland regularly to optimize grassland productivity. However, excessive use of nitrogen fertilizer is not only costly, but also causes serious environmental damage. Therefore, accurately estimating the nitrogen content of grassland vegetation is important for optimizing fertilization and improving forage quality.

[0003] The traditional method of detecting nitrogen content in vegetation is to conduct destructive sampling and take the vegetation samples back to the laboratory for chemical analysis. Although this can achieve high-precision results, the process not only consumes a lot of manpower, material resources and time, but also exhibits highly destructive characteristics. If there is a method that can quickly, non-destructively and accurately obtain vegetation nitrogen content detection, it will be of great significance for optimizing fertilization and improving forage quality.

[0004] Since the nitrogen-containing chemical bonds in the leaves of vegetation will vibrate under a certain intensity of spectral radiation and induce absorption and reflection of specific spectral bands, the intensity of leaf spectral absorption is closely related to the content of nitrogen-containing chemical bonds, and chlorophyll composed of nitrogen elements will also strongly absorb specific wavelength spectra, resulting in changes in the reflectivity of vegetation leaves to specific bands. Based on this, the present invention collects hyperspectral data of grassland vegetation leaves through a ground object spectrometer (abbreviated as: ASD) for research, and proposes a method for estimating the nitrogen content of grassland vegetation leaves. Summary of the invention

[0005] The main technical problem solved by the present invention is to provide a method for estimating the nitrogen content of typical grassland vegetation leaves based on ASD hyperspectral data, which can avoid the continuous destructive sampling and the consumption of a large amount of manpower, material resources and time in the traditional measurement of vegetation nitrogen content, and obtain the vegetation nitrogen content quickly, non-destructively and accurately.

[0006] In order to solve the above technical problems, the present invention provides a method for estimating nitrogen content in grassland vegetation leaves, comprising:

[0007] Collect hyperspectral data of vegetation leaves, pre-process them to obtain hyperspectral reflectance data, and measure the nitrogen content of vegetation leaves;

[0008] Performing a dimensionality reduction transformation on the hyperspectral reflectance data, wherein the dimensionality reduction transformation method includes constructing a vegetation index, a first-order derivative transformation, a continuous wavelet transformation, and / or a continuum removal transformation;

[0009] Performing spectral characteristic parameter screening on the hyperspectral reflectance data after dimensionality reduction transformation, wherein the spectral characteristic parameter screening includes single variable screening and multivariate screening;

[0010] Based on the screened spectral characteristic parameters, nitrogen content estimation models were constructed respectively, and the accuracy of several nitrogen content estimation models was evaluated in combination with the measured nitrogen content of vegetation leaves. The optimal model was selected as the target grassland vegetation leaf nitrogen content estimation model to estimate the nitrogen content of the vegetation leaves.

[0011] Furthermore, the collecting of vegetation leaf hyperspectral data and preprocessing to obtain hyperspectral reflectance data includes:

[0012] Obtain hyperspectral data of vegetation leaves through ground object spectrometer;

[0013] Performing reflectance conversion on the hyperspectral data and calculating the average reflectance of each group of samples to obtain hyperspectral reflectance data;

[0014] The hyperspectral reflectance data is filtered.

[0015] Furthermore, the vegetation index is a linear or nonlinear combination of two or more hyperspectral reflectance data bands, and is used to estimate the nitrogen content of vegetation leaves.

[0016] Furthermore, the first-order derivative transformation includes:

[0017] Performing a first-order derivative transformation on the hyperspectral reflectance data,

[0018]

[0019] In the formula, FDR (λi) is the first-order differential reflectance of the wavelength values ​​of bands i and i+1; R (λi) and R (λi+1) are the hyperspectral reflectances at bands i and i+1, respectively; Δλ is the wavelength value from band i to i+1;

[0020] According to the bending point of the first-order derivative spectral curve and the wavelength positions of the maximum and minimum reflectance, the “three-sided parameters” are extracted to analyze the spectral difference characteristics.

[0021] Furthermore, the continuous wavelet transform includes:

[0022] The hyperspectral reflectance data is decomposed at several scales to obtain the wavelet coefficients decomposed at different positions and scales. The calculation formula is as follows:

[0023]

[0024] Where ψ(λ) is the mother wavelet; ψ a,b (λ) is a series of continuous wavelets generated by translation and scaling of the mother wavelet; λ represents the spectral wavelength in the range of 350-2500nm; a and b are positive real numbers, representing the scaling factor and translation factor, respectively; the original hyperspectral reflectance data is convolved with the continuous wavelet to obtain the wavelet coefficient W f (a,b), also known as wavelet feature WF b,a :

[0025]

[0026] in <f,ψ a,b > represents the inner product of the original hyperspectral reflectance data and the continuous wavelet, and f(λ) is the original hyperspectral reflectance data.

[0027] Furthermore, the continuum removal transformation includes calculating the envelope of the hyperspectral reflectance data, removing the continuity information in the original hyperspectral reflectance data, highlighting the absorption characteristics, and normalizing the hyperspectral reflectance data.

[0028] Furthermore, the method for univariate screening of spectral characteristic parameters is a Pearson correlation analysis method, and the method for multivariate screening of spectral characteristic parameters is a LASSO minimum shrinkage operator screening method.

[0029] Furthermore, the nitrogen content estimation models are constructed based on the screened spectral characteristic parameters, including:

[0030] Based on the spectral characteristic parameters after univariate screening, it was taken as the independent variable, and the actual nitrogen content of vegetation leaves was taken as the dependent variable. The linear fitting method was used to construct a grassland vegetation leaf nitrogen content estimation model based on univariate spectral characteristic parameters.

[0031] Furthermore, the nitrogen content estimation models are constructed based on the screened spectral characteristic parameters, including:

[0032] Based on the spectral characteristic parameters after multivariate screening, they are input into the basic machine learning model to construct a grassland vegetation nitrogen content estimation model based on multivariate spectral characteristic parameters. The basic machine learning model includes the extreme gradient boosting model eXtreme Gradient Boosting, XGboost, the support vector machine model Support Vector Machines, SVM, the artificial neural network model Artificial Neural Network, ANN or the K-nearest neighbor algorithm model K-NearestNeighbor, KNN.

[0033] Furthermore, the accuracy evaluation of the model in combination with the actual nitrogen content of vegetation includes:

[0034] Select the coefficient of determination R 2 The leaf nitrogen content estimation model was evaluated by using the accuracy evaluation indicators of root mean square error (RMSE) and mean absolute error (MAE). The specific calculation formula is as follows:

[0035]

[0036]

[0037]

[0038] Where M i is the measured value of leaf nitrogen content, is the average value of the measured nitrogen content in leaves, P i is the predicted value based on the leaf nitrogen content estimation model, is the average value of the predicted value of leaf nitrogen content, and n is the total number of samples.

[0039] The beneficial effects of the present invention are:

[0040] The present invention uses an ASD ground object spectrometer to measure and preprocess hyperspectral data of grassland vegetation at the leaf scale, obtain vegetation hyperspectral reflectance data, and perform dimensionality reduction transformation on the hyperspectral reflectance data. The dimensionality reduction transformation method includes constructing a vegetation index, a first-order derivative transformation, a spectral continuum removal transformation and a continuous wavelet transformation, and uses a Pearson correlation analysis method or a minimum shrinkage operator LASSO method to screen sensitive hyperspectral feature variables. The data after dimensionality reduction transformation and feature screening are used as model input parameters to construct a typical grassland vegetation leaf nitrogen content estimation model based on spectral feature parameters under different dimensionality reduction transformations at the leaf scale, and the accuracy of the model is evaluated in combination with the measured actual vegetation leaf nitrogen content, and the optimal model is selected as the final target model, and the final target model can quickly, accurately and non-destructively estimate the grassland vegetation leaf nitrogen content.

[0041] This method is suitable for estimating the leaf nitrogen content of various vegetation. While reducing the monitoring cost, it improves the objectivity and accuracy of the monitoring results. It can quickly and accurately estimate the nitrogen content of vegetation leaves without permanently destructively collecting vegetation leaves, avoiding the continuous destructive sampling and the large amount of manpower, material resources and time consumed in traditional determination of vegetation nitrogen content. It is of great significance for subsequent optimization of fertilization and improvement of forage quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a main flow chart of the method for estimating nitrogen content in grassland vegetation leaves provided by an embodiment of the present invention;

[0043] Figure 2 It is the original hyperspectral reflectance data curve of Leymus chinensis and Cleistogenes scabra and the reflectance curves under different transformations;

[0044] Figure 3 This is the relationship between the nitrogen content of the leaves of Leymus chinensis and different spectral characteristic variables;

[0045] Figure 4 is the relationship between the nitrogen content of Cleistogenes scabra leaves and different spectral characteristic variables;

[0046] Figure 5 It is a scatter plot of measured and predicted values ​​of the univariate inversion model constructed based on the six spectral characteristic variables of Leymus chinensis;

[0047] Figure 6 This is a scatter plot of measured and predicted values ​​of a univariate inversion model constructed based on the six spectral characteristic variables of Cleistogenes scabra.

[0048] Figure 7 It is a scatter plot of measured and predicted values ​​of the multivariate inversion model constructed based on the four groups of spectral characteristic variables of Leymus chinensis;

[0049] Figure 8 This is a scatter plot of measured and predicted values ​​based on the construction of a multivariate inversion model based on four groups of spectral characteristic variables of Cleistogenes scabra. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described in conjunction with the accompanying drawings below, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0051] See also Figure 1 The embodiment of the present invention provides a method for estimating the nitrogen content of typical grassland vegetation leaves based on ASD hyperspectral data. The method can quickly, non-destructively and accurately obtain the nitrogen content of vegetation leaves, comprising the following steps:

[0052] S1, collecting hyperspectral data of vegetation leaves, preprocessing them to obtain hyperspectral reflectance data, and measuring the nitrogen content of vegetation leaves;

[0053] S2, performing a dimensionality reduction transformation on the hyperspectral reflectance data, wherein the dimensionality reduction transformation method includes vegetation index construction, first-order derivative transformation, continuum removal transformation and continuous wavelet transformation;

[0054] S3, screening spectral characteristic parameters based on the hyperspectral reflectance data after dimensionality reduction transformation, wherein the spectral characteristic parameter screening method includes Pearson correlation analysis or LASSO minimum shrinkage operator screening method;

[0055] S4. Based on the original hyperspectral reflectance data and the screened spectral characteristic parameters, the grassland vegetation leaf nitrogen content estimation models were constructed respectively, and the accuracy of the models was evaluated in combination with the measured vegetation leaf nitrogen content. The optimal model was selected as the target leaf nitrogen content estimation model to estimate the nitrogen content of the vegetation leaves.

[0056] In this embodiment, Leymus chinensis and Cleistogenes scabra are used as samples for illustration, as follows:

[0057] In step S1, hyperspectral data of grassland vegetation leaves are obtained and preprocessed to obtain hyperspectral reflectance data of vegetation leaves, including:

[0058] The hyperspectral data were acquired in the laboratory using the ASD Field Spec 4 spectrometer (ASD) equipped with a handheld leaf clamp. Specifically, the illuminator provided a light source simulating sunlight to support indoor measurement experiments. Before measurement, a standard white board was used for calibration, and the surface of the leaf to be measured was wiped clean with absorbent cotton. The leaf was evenly covered on the acquisition board to avoid gaps, and then fixed with a clamp to complete the acquisition of the leaf hyperspectral data.

[0059] The preprocessing of the hyperspectral data includes first converting the hyperspectral data into reflectance using View SpecPro software, and calculating the average reflectance of each group of samples respectively, and the average reflectance data is used as the hyperspectral reflectance data of vegetation leaves. View SpecPro is a software specifically used for processing ASD hyperspectral data. The preprocessing also includes performing Savitzky-Golay filtering and smoothing processing on the hyperspectral reflectance data using matlab software to reduce random signal noise in the hyperspectral reflectance data, wherein the Savitzky-Golay filter is a digital filter that can improve the accuracy of the data without changing the signal trend and width.

[0060] The nitrogen content of vegetation leaves was determined using the traditional vegetation nitrogen content determination method, specifically: after the hyperspectral data was collected, the vegetation samples were placed in a 120°C oven for 2 hours and dried at 65°C for 48 hours until the weight no longer changed. The completely dried leaves were crushed and ground using a Latsch (M400) ball mill (particle size of about 5 μm) to obtain dry grass-like powder. 0.02 g of the sample was weighed, and the sample was wrapped in tin foil and placed in the elemental analyzer (Vario MACROCube) in order to measure the vegetation nitrogen content.

[0061] See also Figure 2 In step S2, the hyperspectral reflectance data obtained in step S1 is used as the original spectral data for dimensionality reduction transformation, and the dimensionality reduction transformation method includes vegetation index construction, spectral first-order derivative transformation, spectral continuum removal transformation and spectral continuous wavelet transformation. Specifically,

[0062] (1) Construction of vegetation index:

[0063] Vegetation Index (VI) mainly reflects the difference between vegetation reflectance in visible light, near-infrared bands and soil background. It is a linear or nonlinear combination of two or more spectral bands. It can reduce or eliminate the impact of background noise on grassland vegetation spectral information, thereby enhancing the estimation of vegetation structure.

[0064] Based on the research experience of vegetation index in the prior art, this embodiment evaluates 22 vegetation indices, which are widely used to estimate vegetation biochemical parameters. These vegetation indices can be used to establish subsequent estimation models to find the best method for monitoring the nitrogen content of vegetation leaves.

[0065] Specifically, the 22 vegetation indices include 15 two-band vegetation indices and 7 three-band vegetation indices. The corresponding peaks or troughs are screened out and related calculations are performed to highlight the spectral reflectance information of vegetation leaves. For specific calculations, see Table 1 below.

[0066] Table 1. Vegetation index calculation formula

[0067]

[0068]

[0069] Note: R in the calculation formula (λ) is the spectral reflectance at band λ.

[0070] (2) First-order derivative transformation:

[0071] The original hyperspectral reflectance data is transformed by the first-order derivative to eliminate soil background noise. At the same time, the “trilateral parameters” are extracted according to the bending point of the first-order derivative spectral curve, the wavelength position at the maximum and minimum reflectance, and other characteristics to analyze the spectral difference characteristics. The specific calculation formula is as follows:

[0072]

[0073] In the formula, FDR (λi) is the first-order differential reflectance of the wavelength values ​​of bands i and i+1; R (λi) and R (λi+1) are the spectral reflectances at band i and band i+1 respectively; Δλ is the wavelength value from band i to band i+1.

[0074] The "three-edge parameters" are related variables based on the position characteristics of the first-order derivative spectrum. The "three edges" include red edge, blue edge and yellow edge, and their corresponding parameters include the "three-edge" position, "three-edge" amplitude and "three-edge" area. The "three-edge" parameters are closely related to the physical and chemical parameters of green vegetation, especially the "red valley" formed by strong absorption in the red light region, and the high reflection platform formed in the near-infrared band region. The transition area in the middle forms the typical spectral feature of green vegetation that is different from other landforms - the "red edge". The "red edge" is a diagnostic feature of green vegetation, which can better reflect the spectral characteristics of green vegetation in the spectral curve, especially the red edge parameter, which is sensitive to the content of vegetation nutrients such as N and chlorophyll.

[0075] In this embodiment, 19 characteristic parameters based on different spectral positions and spectral areas are summarized, including peak value, peak area, wavelength of peak, peak area ratio, etc., see Table 2 for details, which are used for subsequent estimation model establishment to find the best method for monitoring vegetation nitrogen content.

[0076] Table 2. Hyperspectral feature parameters and extraction methods

[0077]

[0078]

[0079] (3) Continuous wavelet transform:

[0080] Continuous Wavelet Transformation (CWT) decomposes hyperspectral reflectance data at multiple scales, and can obtain wavelet coefficients decomposed at different locations and scales, and better extract some sensitive spectral feature information for quantitative remote sensing research of vegetation.

[0081] In this embodiment, continuous wavelet transform is used to decompose the original spectra of the leaves of Leymus chinensis and Cleistogenes scabra. Each wavelet coefficient of the continuous wavelet transform is directly comparable to the input reflectance band. The result is easy to interpret. The calculation formula is as follows:

[0082]

[0083] Where ψ(λ) is the mother wavelet; ψ a,b (λ) is a series of continuous wavelets generated by translation and scaling of the mother wavelet; λ represents the spectral wavelength in the range of 350-2500nm; a and b are positive real numbers, representing the scaling factor and translation factor, respectively. The original spectrum is convolved with these continuous wavelets to obtain the wavelet coefficients W f (a,b), also known as wavelet feature WF b,a .

[0084]

[0085] in <f,ψ a,b > represents the inner product of the original spectrum and the continuous wavelet; f(λ) is the original reflection spectrum.

[0086] This embodiment uses the second-order derivative of the Gaussian function as the wavelet basis function of the continuous wavelet transform, that is, the mexh wavelet as the mother wavelet basis, because it is similar to the characteristics of the vegetation spectrum curve. The entire transformation process is implemented in matlab software, and a total of 9 scaling scales are set, namely 2 1 , 2 2 ……2 9 , one scale corresponds to 350-2500nm, each 1nm corresponds to a value, a total of 2151, 9 scales totaling 9*2151=19359 values, and finally the wavelet coefficient data sets of different scales after the original spectrum transformation are obtained for subsequent analysis and modeling.

[0087] (4) Continuum removal:

[0088] Continuum Removal (CR), also known as envelope removal, is a spectrum analysis method that effectively enhances absorption features. By calculating the envelope of the spectrum, the continuity information in the original hyperspectral reflectance spectrum is removed, thereby highlighting the absorption features in the spectrum.

[0089] After the continuum removal transformation, the absorption characteristics of the hyperspectral reflectance data are more obvious, and the reflectance is normalized to between 0 and 1, which can suppress the background noise and effectively separate the spectral absorption characteristics. The calculation formula is:

[0090]

[0091] Where: R'(i) is the continuum-removed spectral reflectance, R(i) is the original spectral reflectance, and RC(i) is the continuum line reflectance.

[0092] The original spectrum has a wavelength range of 350-2500. After being transformed into a continuum to remove the spectrum, there are 3 obvious absorption peaks. In this embodiment, 15 CR transformation characteristic parameters are extracted for subsequent feature screening and model establishment. See Table 3 for details.

[0093] Table 3. Continuum removal parameter definitions and extraction methods

[0094]

[0095] In step S3, in order to avoid the problems of complex inversion model, increased model construction time, reduced model stability, etc., Pearson correlation analysis and LASSO are used to screen the spectral characteristic parameters involved in step S2, and the spectral characteristic parameters that are significantly correlated with the nitrogen content of vegetation leaves are screened for model construction. Figure 3 and Figure 4 , are the relationship diagrams between nitrogen content of leaves of Leymus chinensis and Cleistogenes scabra and different spectral characteristic variables;

[0096] (1) Correlation analysis feature screening:

[0097] In order to screen the best hyperspectral variables for constructing a univariate vegetation nitrogen content inversion model, the Pearson correlation method was used to analyze the variables under different spectral transformations of the original spectra of Leymus chinensis and Cleistogenes scabra. The formula is as follows:

[0098]

[0099] Where: r i is the single correlation coefficient between vegetation N content (Nitrogen) and hyperspectral variable (R), i is the band number, R ni is the spectral value of the ith spectral reflectance of the nth sample or its transformed row, is the average value of the spectral reflectance of the nth sample in the i-th band or its transformed spectrum, is the average N content.

[0100] According to the size of the single correlation coefficient between the vegetation nitrogen content and the hyperspectral variable, the original spectrum and the spectra under different dimensionality reduction transformations were selected to have a greater impact on the nitrogen content of the leaves of Leymus chinensis and Cleistogenes striata. The screening results and the size of the correlation coefficient are shown in Table 4.

[0101] In addition, the determination coefficients R between wavelet coefficients of different scales and nitrogen content of two vegetation were calculated. 2 , through the coefficient of determination R 2The linear correlation between the nitrogen content of vegetation leaves and the wavelet coefficient was evaluated. Specifically, the wavelet coefficient WF2043.5, R 2 =0.848, the wavelet coefficient WF2126.5, R 2 =0.396.

[0102] Table 4 Correlation screening results between nitrogen content of two vegetation leaves and original spectra and different spectral transformations

[0103]

[0104] Note: ** indicates a significant level of p<0.01.

[0105] Among them, ρ 716 Indicates the corresponding value of reflectance at 716nm of the original spectrum; It represents the value corresponding to the reflectance at 640nm of the first-order derivative spectrum.

[0106] (2) Minimum shrinkage operator LASSO method feature screening:

[0107] In order to screen the best hyperspectral variables for constructing a multivariate vegetation nitrogen content inversion model, the minimum shrinkage operator LASSO method was used to screen out the important variables that have a greater impact on the nitrogen content of Leymus chinensis leaves and Cleistogenes striata leaves from VI variables, hyperspectral feature variables, CR variables, and wavelet power variables. The screening results are shown in Tables 5 and 6. As can be seen from Table 5, after the LASSO method feature screening, the number of variables is significantly reduced compared with the previous Tables 1, 2, and 3. The screened feature variables all contain variables with good correlation with the nitrogen content of the two vegetation leaves, and in the LAASO feature screening process, zero variance, near-zero variance, and highly correlated variables are eliminated to avoid errors caused by redundant variables in the subsequent modeling process.

[0108] Table 5 LASSO screening results of leaf nitrogen content and vegetation index, hyperspectral characteristic variables and continuum removal variables of two types of vegetation

[0109]

[0110] Table 6 LASSO screening results of nitrogen content and wavelet power of two vegetation leaves

[0111]

[0112]

[0113] Step S4. Construct a grassland vegetation leaf nitrogen content estimation model based on the screened spectral characteristic parameters, and evaluate the accuracy of the model in combination with the actual nitrogen content of the vegetation.

[0114] In this embodiment, a vegetation leaf nitrogen content inversion model is constructed based on the screened univariate spectral characteristic parameters and multivariate spectral characteristic parameters through different modeling methods, wherein a vegetation nitrogen content inversion model is constructed based on the univariate spectral characteristic parameters by a linear fitting method; and based on the multivariate spectral characteristic parameters, a machine learning algorithm is used to input different types of multivariate characteristic parameters after feature screening into the extreme gradient boosting (eXtreme GradientBoosting, XGboost), support vector machine (SVM), artificial neural network (ANN), and K-nearest neighbor algorithm (KNN) basic models to establish a typical grassland vegetation nitrogen content estimation model based on characteristic parameters under different spectral transformations.

[0115] The accuracy of the inversion model was evaluated by combining the measured true nitrogen content of vegetation leaves, among which the determination coefficient (R 2 ), root mean square error (RMSE) and mean absolute error (MAE) were used as accuracy evaluation indicators to comprehensively evaluate the leaf nitrogen content estimation model. The determination coefficient (R 2 ) represents the correlation between the independent variable and the dependent variable. The closer the value is to 1, the higher the accuracy of the model. The root mean square error (RMSE) is used to measure the error between the predicted value and the true value. The smaller the RMSE value, the better the prediction result of the model. The mean absolute error (MAE) represents the average absolute error between the predicted value and the true value. The smaller the MAE, the better the model. The specific calculation formula is as follows:

[0116]

[0117]

[0118]

[0119] Where M i is the measured value of leaf nitrogen content, is the average value of the measured nitrogen content in leaves, P i is the predicted value based on the leaf nitrogen content estimation model, is the average value of the predicted value of leaf nitrogen content, and n is the total number of samples.

[0120] The specific construction of the grassland vegetation leaf nitrogen content estimation model is as follows:

[0121] (1) Constructing a model for estimating nitrogen content in Leymus chinensis leaves based on correlation analysis:

[0122] According to the results of correlation analysis, the original hyperspectral reflectance data of Leymus chinensis and the spectral characteristic variables obtained after various dimensionality reduction transformation methods, a total of 6 spectral variables, were used as independent variables, and the nitrogen content of vegetation leaves was used as the dependent variable. The linear fitting method was used to construct an inversion model for the nitrogen content of the leaves of the main vegetation in a typical grassland. The differences between different models were compared to determine the optimal nitrogen content inversion model based on single-variable spectral characteristic variables.

[0123] Table 7 shows the inversion results of the estimation model for nitrogen content in leaves of Leymus chinensis. The results show that among the six nitrogen content inversion models, the first-order transformation spectral sensitive band The nitrogen content inversion model of the hyperspectral characteristic parameter SDb and the vegetation index TCARI has better performance, which can explain 76%-78% of the change in nitrogen content. The wavelet coefficient WF2043.5 variable has the highest accuracy, and its determination coefficient R 2 is 0.84. But the original spectral sensitive band ρ 716 The nitrogen content inversion model with the continuum removal parameter S1 performed poorly and could only explain 28%-48% of the nitrogen content variation. From the model validation results, the model stability of the wavelet coefficient WF2043.5 variable was better, and the determination coefficient R of the validation data set was 2 The value is 0.81, so this model is the most stable and reliable.

[0124] Table 7 Establishment of linear inversion model of nitrogen content in leaves of Leymus chinensis and spectral variables

[0125]

[0126] Among them, T is the test set, n=63; V is the validation set, n=13.

[0127] The results of the nitrogen content inversion model constructed using six spectral variables of Leymus chinensis are as follows: Figure 5 shown. Figure 5 The scatter plot shows the predicted value and measured value of leaf nitrogen content inversion model, where the solid line represents the fitting line of the measured value and the predicted value, and the dotted line is the 1:1 straight line y=x, indicating that the model predicted value and the measured value are equal. The farther away from the dotted line, the greater the gap between the predicted value and the measured value. The larger the angle between the fitting line and the 1:1 line, the poorer the overall prediction effect of the sample. The smaller the angle, the better the overall prediction effect. It can be clearly seen from the figure that the continuum removal parameter S1 and the original spectrum ρ 716The accuracy of the constructed nitrogen content inversion model is poor, especially when the nitrogen content is high, the model cannot make a good prediction. However, the estimated values ​​and measured values ​​of the estimation model constructed based on the wavelet coefficient WF2043.5 are compact and evenly distributed, and the inversion accuracy is good.

[0128] (2) Construction of an inversion model for nitrogen content in leaves of Leymus chinensis based on LASSO screening:

[0129] According to the LASSO feature screening results, four groups of variables were obtained for Leymus chinensis: LASSO-vegetation index variables, LASSO-hyperspectral feature variables, LASSO-continuum removal parameters, and LASSO-wavelet coefficients. The four groups of variables were used as model input variables, and the nitrogen content of Leymus chinensis leaves was used as the dependent variable. The nitrogen content inversion models of Leymus chinensis using the XGBoost, SVM, ANN, and KNN methods were established, and the differences between different algorithms were compared to determine the best model. The modeling results of the four machine learning algorithms are shown in Table 7.

[0130] Table 7 shows the model inversion results of nitrogen content and spectral variables of Leymus chinensis leaves. The test accuracy of each inversion model is high. The four model validation sets R 2 The four model validation sets of LASSO-hyperspectral feature variables are between 0.65-0.74, RMSE is between 0.26-0.30, and R 2 The four models with LASSO-continuum parameters removed have a validation set R between 0.71 and 0.77, RMSE between 0.24 and 0.28, and 2 The four model validation sets R based on LASSO-wavelet coefficients are between 0.55 and 0.77. 2 The value ranged from 0.85 to 0.92, which improved the estimation accuracy by about 10.4% compared with the other three methods. The multivariate nitrogen content model constructed by the SVM algorithm had the highest accuracy, which could explain 92% of the nitrogen content variation. The difference between the modeling accuracy and the verification accuracy was about 6%, and the model was relatively stable.

[0131] Table 7 Results of the inversion model of nitrogen content in Leymus chinensis leaves and four groups of spectral variables

[0132]

[0133]

[0134] Among them, T is the test set, n=63; V is the validation set, n=13.

[0135] Figure 7 The results of constructing a multivariate inversion model for the four groups of spectral variables of Leymus chinensis show that among the four modeling methods based on vegetation index variables, the model constructed by the SVM method has the highest accuracy, R 2=0.86, the other three models R 2 The model accuracy is between 0.78 and 0.83, which is good. However, the model verification accuracy is reduced. 2 The SVM model was between 0.65 and 0.74, and the SVM model dropped by 25%, indicating that the model had some overfitting. In contrast, the ANN model was more stable, with only a 5% difference in verification accuracy, and the RMSE and MAE differences were also smaller. The predicted points and measured points were more evenly distributed on both sides of the fitting line, and the fitting effect was better. From the four nitrogen content inversion models constructed based on hyperspectral feature variables and continuum removal parameters, it can be seen that the SVM model has a higher model accuracy (R 2 0.90-0.93), but the verification accuracy is poor (R 2 0.55-0.70). In comparison, the ANN and KNN models performed more stably and had higher verification accuracy. The inversion model based on wavelet coefficient variables showed the best fitting effect. The nitrogen content inversion models constructed by the four machine learning methods had higher accuracy, R 2 Between 0.95 and 0.98, the verification accuracy is good, and the verification R 2 The model is relatively stable, ranging from 0.85 to 0.92. Among the nitrogen inversion models based on wavelet coefficient variable groups, the SVM model has the best performance, with the training set and validation set explaining 98% and 92% of the nitrogen content changes, respectively. The validation set RMSE = 0.18, MAE = 0.13, and has a better fitting effect than other models.

[0136] (3) Constructing an inversion model for nitrogen content in Cleistogenes scabra leaves based on correlation analysis;

[0137] According to the results of correlation analysis, the original hyperspectral reflectance data of Cleistogenes striata and the spectral characteristic variables obtained after various dimensionality reduction transformation methods, a total of 6 spectral variables, were used as independent variables, and the nitrogen content of vegetation leaves was used as the dependent variable. The linear fitting method was used to construct an inversion model for the nitrogen content of the leaves of the main vegetation in a typical grassland. The differences between different inversion models were compared to determine the optimal single variable nitrogen inversion model.

[0138] The model inversion results of nitrogen content in leaves of Cleistogenes scabra and six spectral variables are shown in the table. 633 , The nitrogen estimation model constructed with A1 as independent variables has poor accuracy and the determination coefficient R 2 The nitrogen content estimation model constructed with mND705, VI5 and WF2126.5 as independent variables has better accuracy, and the determination coefficient R2 is between 0.51 and 0.62. Among them, the model constructed with the wavelet coefficient WF2126.5 has the best performance, and the validation set R 2 =0.57.

[0139] Table 8 Establishment of linear inversion model between nitrogen content of Cleistogenes scabra leaves and spectral variables

[0140]

[0141] Among them, T is the test set, n=63; V is the validation set, n=13.

[0142] The results of the nitrogen content inversion model constructed using six spectral variables of Cleistogenes scabra are shown in Figure 2. Figure 6 As shown in the figure, it can be clearly seen that A1, ρ 633 The accuracy of the nitrogen content inversion model constructed by mND705 is poor, which is reflected in the fact that as the nitrogen content increases, the values ​​predicted by the model are mostly smaller than the measured values. However, the predicted values ​​of the wavelet coefficient WF2126.5 inversion model are closer to the measured values, and the model accuracy is better.

[0143] (4) Establishment of an inversion model for nitrogen content in Cleistogenes scabra leaves based on LASSO screening:

[0144] According to the LASSO feature screening results, four groups of variables for Cleistogenes serrata were obtained: LASSO-vegetation index variables, LASSO-hyperspectral feature variables, LASSO-continuum removal parameters and LASSO-wavelet coefficients. The four groups of variables were used as model input variables, and the nitrogen content of Cleistogenes serrata leaves was used as the dependent variable. The nitrogen content inversion models of Cleistogenes serrata using XGBoost, SVM, ANN and KNN methods were established respectively, and the differences between different algorithms were compared to determine the optimal model.

[0145] Table 9 shows the model inversion results of nitrogen content and spectral variables of Cleistogenes scabra. From the results, it can be seen that the four model training sets R based on LASSO-VI variables 2 Between 0.45-0.61, validation set R 2 Between 0.36-0.72, LASSO-hyperspectral feature variables and LASSO-continuum removal variables four inversion model training sets R 2 The validation set R 2 The model fitting effect is generally average, the gap between the training set and the validation set is large, and the model is not stable enough. However, among the four inversion models based on LASSO-wavelet coefficients, the ANN model has the highest accuracy, and the training set R 2 =0.98, validation set R 2 =0.72. Compared with the modeling accuracy, the verification accuracy of the ANN model is slightly reduced, but compared with other models, the ANN model based on LASSO-wavelet coefficients has the best stability and reliability.

[0146] Table 9 Results of the inversion model of nitrogen content in Cleistogenes scabra leaves and four groups of spectral variables

[0147]

[0148] Among them, T is the training set, n=63; V is the validation set, n=13.

[0149] Figure 8 The results of constructing a multivariate inversion model for four groups of spectral variables of Cleistogenes scabra. It can be seen from the figure that among the four modeling methods based on vegetation index variables, the model constructed by the SVM method has the highest accuracy, R 2 =0.61, the other three models R 2 The SVM model validation accuracy is between 0.45 and 0.57, but the R 2 Only 0.36, compared with XGBoost model validation accuracy is the highest, R 2 =0.72, the ANN model performs basically the same in the training set and the validation set, and the determination coefficient R 2 The results of the inversion models constructed by the other three groups of spectral variables show that the XGBoost model has a poor fitting effect. As the nitrogen content increases, the points composed of the predicted values ​​and the measured values ​​are distributed on the right side of the 1:1 line, and the model prediction value is less than the true value. The SVM model and the KNN model have a good fitting effect in the training set, but from the point distribution of the validation set, the model validation accuracy decreases significantly, and there are some overfitting phenomena. The validation set R 2 In contrast, the ANN model is more stable, with smaller differences in accuracy between the training set and the validation set. The ANN inversion model based on LASSO-wavelet coefficients has the best performance, with a modeling R 2 =0.98, verification R 2 =0.72, and the predicted points and measured points are more evenly distributed on both sides of the 1:1 line, and the fitting effect is better.

[0150] In summary, the variables highly correlated with the nitrogen content of Leymus chinensis leaves include the original spectral sensitive band ρ 716 , first-order transform spectral sensitive band The narrow-band vegetation index TCARI, the hyperspectral characteristic parameter SDb, the continuum removal parameter S1, and the wavelet coefficient WF2043.5, all of the six spectral variables were significantly correlated with the nitrogen content of the leaves of Leymus chinensis (p < 0.01), with the wavelet coefficient WF2043.5 having the highest correlation. The variables highly correlated with the nitrogen content of the leaves of Cleistogenes scabra included the original spectral sensitive band ρ 633 , first-order transform spectral sensitive band The narrow-band vegetation index mND705, hyperspectral characteristic parameter VI5, continuum removal parameter A1, and wavelet coefficient WF2126.5 were all significantly correlated with the leaf nitrogen content of Cleistogenes scabra (p<0.01), with the hyperspectral characteristic parameter VI5 having the highest correlation. After establishing the nitrogen estimation model of six spectral parameters of Leymus chinensis and Cleistogenes scabra, it was found that among the six single-variable models of Leymus chinensis, ρ 716 and S1 modeling R 2 The R values ​​of the other four univariate models were 0.48 and 0.28, respectively, indicating poor model performance. 2 The model performance was good, and the univariate nitrogen estimation model constructed with wavelet coefficient WF2043.5 was the best, which could explain 81% of the variation of nitrogen content in leaves of Leymus chinensis. 633 , Model R with A1 as independent variable 2 The R value of the model with mND705, VI5 and WF2126.5 as independent variables is between 0.14 and 0.38, indicating poor estimation accuracy. 2 The value of the single variable nitrogen estimation model is between 0.51 and 0.62, among which the single variable nitrogen estimation model constructed with the wavelet coefficient WF2126.5 has the best performance, and the modeling accuracy and verification accuracy can explain 62% and 57% of the nitrogen content changes in the leaves of Cleistogenes scabra. In contrast, it was found that the single variable nitrogen inversion model with wavelet coefficients as independent variables for both vegetation had the best accuracy, indicating that the wavelet features extracted by continuous wavelet transform have good performance in estimating the nitrogen content of the two vegetations, and the wavelet coefficients of Leymus chinensis and Cleistogenes scabra after correlation screening are WF2043.5 and WF2126.5, respectively. The positions of the optimal wavelet features corresponding to the nitrogen content of the leaves of the two vegetations are both located in the shortwave infrared region, and are all within the fifth scale after transformation, which indicates that the shortwave infrared region is the sensitive spectral region for the nitrogen content of the leaves of the two vegetations, and the fifth scale wavelet features can better reflect the changes in the leaf structure of the original spectrum and the spectral absorption characteristics of nitrogen.

[0151] The LASSO method was used to screen the characteristic variables of vegetation index, hyperspectral characteristic parameters, continuum removal parameters and wavelet coefficients. Finally, four groups of variables were obtained for each of the two vegetation types as model input variables. Four machine learning algorithms, XGBoost, SVM, ANN and KNN, were used to construct multivariate nitrogen inversion models for Leymus chinensis and Cleistogenes striata. The results showed that among the 16 multivariate nitrogen content inversion models constructed with four groups of spectral variables of Leymus chinensis, the SVM model based on wavelet coefficients performed best, and the modeling R 2 The value of R 2The value of the ANN model based on wavelet coefficients was 0.92, RMSE and MAE were 0.18 and 0.13 respectively, and there was a 6% difference in accuracy between the modeling accuracy and the verification accuracy, indicating that the model was stable and reliable. 2 =0.98, RMSE and MAE are 0.03 and 0.02 respectively, verifying R 2 =0.72, RMSE and MAE are 0.18 and 0.14 respectively.

[0152] The results showed that, regardless of the single variable or multivariate inversion model, for the leaf nitrogen content of two typical grassland vegetation types, Leymus chinensis and Cleistogenes scabra, in Inner Mongolia, the estimation effect of nitrogen inversion using wavelet transform sensitive parameters as independent variables was the best, providing a reference for the selection of spectral transformation methods and regression models in future grassland vegetation nitrogen content monitoring.

[0153] In this embodiment, by constructing single variable and multivariate nitrogen content inversion models and performing accuracy evaluation, the best model among the estimation models for the nitrogen content of the leaves of two vegetations, Leymus chinensis and Cleistogenes scabra, typical grasslands in Inner Mongolia, is obtained as the final model to estimate the nitrogen content of the leaves of vegetation. This method provides a reference for the selection of spectral transformation methods and regression models in nitrogen content monitoring of other vegetation in grasslands in the future. The estimation model constructed according to the above process can batch process hundreds of spectral data and screen the characteristic parameters of the spectrum, with high efficiency and accuracy.

[0154] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein. These modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for estimating nitrogen content in grassland vegetation leaves, characterized in that: include: Collect hyperspectral data of vegetation leaves, pre-process them to obtain hyperspectral reflectance data, and measure the nitrogen content of vegetation leaves; Performing a dimensionality reduction transformation on the hyperspectral reflectance data, wherein the dimensionality reduction transformation method includes constructing a vegetation index, a first-order derivative transformation, a continuous wavelet transformation, and a continuum removal transformation; The continuous wavelet transform includes decomposing the hyperspectral reflectance data at several scales to obtain the wavelet coefficients decomposed at different positions and scales. The calculation formula is as follows: Where ψ(λ) is the mother wavelet; ψ a,b (λ) is a series of continuous wavelets generated by translation and scaling of the mother wavelet; λ represents the spectral wavelength in the range of 350-2500nm; a and b are positive real numbers, representing the scaling factor and translation factor, respectively; the original hyperspectral reflectance data is convolved with the continuous wavelet to obtain the wavelet coefficient W f (a,b), also known as wavelet feature WF b,a : in <f,ψ a,b > represents the inner product of the original hyperspectral reflectance data and the continuous wavelet, and f(λ) is the original hyperspectral reflectance data; Multivariate screening is performed on the hyperspectral reflectance data after dimensionality reduction transformation; the method for multivariate screening of spectral characteristic parameters is LASSO minimum shrinkage operator screening method; Based on the spectral characteristic parameters after multivariate screening, the spectral characteristic parameters are input into a machine learning basic model to construct a nitrogen content estimation model based on the multivariate spectral characteristic parameters, wherein the machine learning basic model includes an extreme gradient boosting model, a support vector machine model, an artificial neural network model or a K-nearest neighbor algorithm model; The accuracy of several nitrogen content estimation models is evaluated in combination with the measured nitrogen content of vegetation leaves, and the optimal model is selected as the nitrogen content estimation model for the target grassland vegetation leaves to estimate the nitrogen content of the vegetation leaves.

2. The method for estimating nitrogen content in grassland vegetation leaves according to claim 1, characterized in that: The collecting of vegetation leaf hyperspectral data and preprocessing to obtain hyperspectral reflectance data includes: Obtain hyperspectral data of vegetation leaves through ground object spectrometer; Performing reflectance conversion on the hyperspectral data and calculating the average reflectance of each group of samples to obtain hyperspectral reflectance data; The hyperspectral reflectance data is filtered.

3. The method for estimating nitrogen content in grassland vegetation leaves according to claim 1, characterized in that: The vegetation index is a linear or nonlinear combination of two or more hyperspectral reflectance data bands and is used to estimate the nitrogen content of vegetation leaves.

4. The method for estimating nitrogen content in grassland vegetation leaves according to claim 1, characterized in that: The first-order derivative transformation includes, Performing a first-order derivative transformation on the hyperspectral reflectance data, are the hyperspectral reflectances at bands i and i+1, respectively; Δλ is the wavelength value from band i to i+1; According to the bending point of the first-order derivative spectrum curve and the wavelength positions of the maximum and minimum reflectance, the "three-sided parameters" are extracted to analyze the spectral difference characteristics.

5. The method for estimating nitrogen content in grassland vegetation leaves according to claim 1, characterized in that: The continuum removal transformation includes calculating the envelope of the hyperspectral reflectance data, removing the continuity information in the original hyperspectral reflectance data, highlighting the absorption characteristics, and normalizing the hyperspectral reflectance data.

6. The method for estimating nitrogen content in grassland vegetation leaves according to claim 1, characterized in that: The accuracy evaluation of several nitrogen content estimation models in combination with the measured nitrogen content of vegetation leaves includes: Select the coefficient of determination R 2 The leaf nitrogen content estimation model was evaluated by using the accuracy evaluation indicators of root mean square error (RMSE) and mean absolute error (MAE). The specific calculation formula is as follows: Where M i is the measured value of leaf nitrogen content, is the average value of the measured nitrogen content in leaves, P i is the predicted value based on the leaf nitrogen content estimation model, is the average value of the predicted leaf nitrogen content, and n is the total number of samples.

Images