Inversion method of SPAD value and nitrogen concentration of leaves in whole growth period of rice based on hyperspectral technology

Abstract

The invention discloses a hyperspectral technology-based inversion method for SPAD values and nitrogen concentrations of leaves in the whole growth period of rice. The hyperspectral data of the leaves of the rice are obtained by means of an ASD Field Spec4 portable ground object spectrometer in the initial differentiation period (PI), the heading period (HD), the middle filling period (MF) and the mature period (MA) of young ears; measuring the SPAD value of the rice as the chlorophyll content by using an SPAD-502 chlorophyll meter; determining the nitrogen concentration of the rice leaves by adopting a Kjeldahl method; performing correlation analysis on the hyperspectral data of the leaves, the nitrogen concentration and SPAD, and screening sensitive wavebands. According to the method, the hyperspectral characteristic wave band of the rice leaf is extracted through a correlation analysis method, so that the redundancy of full-spectrum data modeling is reduced, data are optimized, an inversion model of the nitrogen concentration and SPAD value of the rice leaf is constructed based on linear regression (LR), random forest (RF), support vector regression (SVR) and gradient boosting (GBRT), and the nitrogen concentration and SPAD value of the rice leaf are obtained. The optimal machine learning algorithm with high precision and good stability is obtained, and a scientific basis can be provided for non-destructive real-time leaf color monitoring and precise fertilization control of indica rice in the South China.

Description

Technical Field

[0001] The present invention belongs to the technical field of precision agriculture, and in particular relates to an inversion method for SPAD values ​​and nitrogen concentrations of rice leaves during the entire growth period based on hyperspectral technology. Background Art

[0002] Rice is one of the food crops that humans depend on for survival. More than 50% of people in my country rely on rice as their staple food. Therefore, rice occupies an important position in the food system. Nitrogen is the main element for synthesizing chlorophyll, and chlorophyll is an important pigment for converting light energy into chemical energy. Therefore, there is an inevitable connection between leaf nitrogen concentration and chlorophyll content. A large number of studies have shown that leaf SPAD value can indirectly reflect the chlorophyll content of crops. However, in actual production, leaf SPAD value is usually used to characterize the chlorophyll content or nitrogen content of plants, and increases with the increase of nitrogen application. Although the change trends of leaf SPAD value and leaf nitrogen concentration are similar, there are few reports on the quantitative difference between leaf SPAD value and leaf nitrogen concentration based on spectral technology. Hyperspectral technology has the advantages of real-time, rapid, convenient and non-destructive, and provides new technical means and methods for monitoring crop leaf color and nitrogen nutrient monitoring. Since Curran et al. revealed the physical mechanism between spectral signals and leaf nitrogen content, more and more scholars have carried out a large number of studies on spectral nitrogen diagnosis, involving different crops such as winter wheat, cotton, corn, soybean, peanut, and rice.

[0003] These studies mostly focus on data sets consisting of a single or full growth period of crops, and most of the research objects are northeast japonica rice, the algorithms used are relatively simple, and the spectral inversion model is mainly for leaf nitrogen concentration or SPAD value, lacking comparative analysis of the two. Both leaf nitrogen concentration and SPAD value can be used as relevant basis for adjusting the amount of nitrogen applied to rice, but there are differences in the spectral response characteristics between the two. Therefore, conducting a study on the spectral difference between SPAD value and nitrogen concentration of indica rice leaves is of great significance for accurate leaf color diagnosis and guiding precise fertilization of rice. The present invention uses Meixiangzhan No. 2, a high-quality indica conventional rice variety with the largest planting area in South China, as the test material, and uses the ASD portable ground feature spectrometer to collect spectral data of four key growth periods. Through sensitive band screening and different learning algorithms, the rice leaf SPAD value and nitrogen concentration inversion model are constructed to explore the spectral differences and mutual connections between the rice leaf SPAD value and nitrogen concentration. The research results will provide theoretical and scientific guidance for the monitoring and diagnosis of leaf color and nitrogen of high-quality indica rice varieties in South China.

[0004] To this end, we proposed an inversion method for SPAD values ​​and nitrogen concentrations of rice leaves throughout the growth period based on hyperspectral technology to solve the above problems. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide an inversion method for SPAD value and nitrogen concentration of rice leaves during the whole growth period based on hyperspectral technology.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] An inversion method for SPAD value and nitrogen concentration of rice leaves during the whole growth period based on hyperspectral technology, which uses correlation analysis to extract spectral characteristic bands of SPAD value and nitrogen concentration of rice leaves, builds inversion models of nitrogen concentration and SPAD value of rice leaves based on linear regression, random forest, support vector regression and gradient boosting, and compares the differences in characteristic bands and inversion models of leaf nitrogen concentration and SPAD value, specifically including the following steps:

[0008] (1) Obtain hyperspectral data of rice leaves at different growth stages, measure the SPAD value of rice leaves with a SPAD-502 chlorophyll meter, and measure the nitrogen concentration of rice leaves with the Kjeldahl method;

[0009] (2) Correlation analysis was performed between leaf hyperspectral data and nitrogen concentration and SPAD value to select sensitive bands;

[0010] (3) Randomly construct the normalized difference spectral index (NDSI) in pairs within the sensitive band range;

[0011] (4) The sensitive band spectral reflectance and NDSI were used as input to construct the inversion models of rice leaf nitrogen concentration and SPAD value, and the differences in the inversion models of leaf nitrogen concentration and SPAD value were compared;

[0012] (5) The inversion model is verified by dividing the training set into a test set ratio of 7:3.

[0013] The above-mentioned method for inverting SPAD value and nitrogen concentration of rice leaves during the whole growth period based on hyperspectral technology, in step (1), select the American ASD Field Spec4 portable ground feature spectrometer to collect rice leaf hyperspectral data; select clear and cloudless weather at different growth periods of rice, the measurement time is 10:00-13:00, with the help of the American ASD FieldSpec4 ground feature spectrometer, use its built-in handheld leaf clip leaf spectrum detector and built-in quartz halide lamp light source for measurement, select the first leaf of the rice top that is fully unfolded, and measure the upper, middle and lower parts of the leaf. During the measurement, the leaf is placed in the leaf chamber of the leaf clip to ensure that the leaf is horizontal and the detected area is the same. Three spectral curves are collected for each sample leaf, and the average value of 10 spectral curves is taken as the final reflectance information of the sample for each sample point, and the data of the collected spectral reflectance is recorded using View Spec Pro software. White board correction is performed before each repeated data collection. The different growth periods are the initial stage of panicle differentiation (PI), heading stage (HD), mid-filling stage (MF) and maturity stage (MA).

[0014] In the above-mentioned inversion method of SPAD value and nitrogen concentration of rice leaves during the whole growth period based on hyperspectral technology, the portable ground feature spectrometer needs to be calibrated with a white board before collecting data, and the calibration result requires that the spectral reflectance within the band range is 1. White board calibration is performed before each repeated data collection, and 10 data points are repeatedly measured for each data point, and the average value is the leaf spectrum data.

[0015] In the above-mentioned method for inverting SPAD value and nitrogen concentration of rice leaves during the whole growth period based on hyperspectral technology, in step (2), correlation analysis is used to select characteristic spectra of hyperspectral data; the best sensitive bands of leaf nitrogen concentration obtained by correlation analysis are 541nm, 542nm and 712nm. The best sensitive bands of leaf SPAD value are 603nm, 702nm and 703nm. 497-730nm is the sensitive band of leaf nitrogen concentration and SPAD value during the whole growth period.

[0016] In the above-mentioned method for inverting the SPAD value and nitrogen concentration of rice leaves during the whole growth period based on hyperspectral technology, in step (3), according to the range of the sensitive band, the reflectance of each band is randomly combined in pairs to construct the normalized difference spectral index (NDSI), and the calculation formula is as follows. NDSI (R497, R654) has the strongest correlation with the leaf SPAD value, R = 0.97, and NDSI (R729, R730) has the strongest correlation with the leaf nitrogen concentration, R = -0.90.

[0017] NDSI=(R1-R2) / (R1+R2) (1)

[0018] In formula (1), R1 and R2 are the spectral reflectances of any band in the range of 497 to 730 nm.

[0019] In the above-mentioned method for inverting SPAD values ​​and nitrogen concentrations of rice leaves during the entire growth period based on hyperspectral technology, in step (4), the reflectivity corresponding to the characteristic wavelength selected in step (2) and the NDSI constructed in step (3) are used as independent variables of the inversion model, and the nitrogen concentration and SPAD value of rice leaves are used as target variables to construct inversion models of four algorithms, namely, linear regression (LR), random forest (RF), support vector machine (SVR) and gradient boosting (GBRT), respectively.

[0020] The above-mentioned inversion method of SPAD value and nitrogen concentration of rice leaves during the whole growth period based on hyperspectral technology, LR is a classic modeling method, which can explain the linear relationship between single factor variables and multiple independent variables, the model result is unique, and can well reflect the correlation between each independent variable and the dependent variable. The linear regression (LR) calculation formula can be expressed by the following formula:

[0021] The univariate linear regression (SLR) model function expression is as follows:

[0022] Y=β0+β1X (2)

[0023] In formula (2), Y is the estimated value of the model output, β0 is the intercept, and β1 is the prediction coefficient, which can be calculated using the least squares method.

[0024] Multiple linear regression (MLR) uses multiple correlated independent variables to solve the regression problem. Its model function expression is:

[0025] Y=β0+β1X1+β2X2+…+βmXm (3)

[0026] In formula (3), similar to the univariate linear regression model function, m is the total number of independent variables X used, β0 is the intercept, and β1…βm are the prediction coefficients corresponding to each independent variable.

[0027] The above-mentioned inversion method for SPAD value and nitrogen concentration of rice leaves during the whole growth period based on hyperspectral technology is characterized in that RF is a tree structure classifier proposed by Breiman. The principle is based on classification regression tree, and there is no need to screen variables, and the importance of features can be output. It uses the bootsrap resampling method to extract multiple samples from the original sample, and makes a decision tree model for each bootsrap sample, and then combines multiple decision trees to obtain the final prediction result through voting. The RF algorithm is insensitive to noisy data and has strong robustness. It can be used to solve nonlinear regression problems and has the characteristics of being interpretable, requiring few samples, not overfitting, and having high precision. The calculation process of random forest (RF) is as follows:

[0028] Extract training data set: Randomly extract a certain amount of samples from the original data set as training samples, which will be used to generate decision trees. When building the model, randomly divide it according to the ratio of training set: test set = 7:3.

[0029] Feature selection: For each decision tree node, when a feature needs to be tested at that node, only the best features are selected from a subset of all features. In this experiment, all tested features include the input best band spectral reflectance.

[0030] Decision tree generation: Generate multiple decision trees based on feature selection methods, and optimize the performance of decision trees by adjusting the maximum depth of the tree, the number of trees, etc.

[0031] Decision tree integration: All generated decision trees are combined by weighting and other methods to obtain the final regression result. Through the integration of decision trees, a prediction model is finally generated to predict the leaf nitrogen concentration and SPAD value of rice.

[0032] The above-mentioned inversion method of SPAD value and nitrogen concentration of rice leaves during the whole growth period based on hyperspectral technology is characterized in that support vector regression SVR is an application model of support vector machine SVM in regression problem, and its core idea is to find a hyperplane so that the total deviation of all sample points from the hyperplane is minimized; the support vector regression SVR model formula is as follows:

[0033]

[0034] In formula (4), i represents the i-th sample point, a i is the Lagrange coefficient; represents the mean of the random variable, the estimated value; M represents the sequence of data vectors; b is the threshold, representing the kernel function of the SVM model, and T represents the state.

[0035] The above-mentioned inversion method of SPAD value and nitrogen concentration of rice leaves during the whole growth period based on hyperspectral technology is characterized by a gradient boosting (GBRT) ensemble learning algorithm for supervised learning tasks. Gradient boosting trains multiple weak learners, and uses an iterative method to train each weak learner on the residual of the previous learner, and gradually adjusts its weight to minimize the loss function, and finally generates a strong learner. The gradient boosting (GBRT) calculation formula can be expressed by the following formula:

[0036]

[0037] In formula (5), T(x, θ m ) represents a decision tree; θ m is the parameter of the decision tree; M is the number of trees. Gradient boosting tree regression can flexibly process various types of data, capture complex nonlinear relationships in the data, handle interactions, and make the model more robust with less parameter adjustment.

[0038] The above-mentioned method for inverting SPAD value and nitrogen concentration of rice leaves during the whole growth period using hyperspectral technology is characterized in that the determination coefficient R 2 , root mean square error RMSE and normalized root mean square error NRMSE are used as evaluation indicators of the inversion model to evaluate the performance of the inversion model.

[0039] The calculation formulas for each evaluation index are as follows, where n is the number of samples. is the predicted value, is the mean value, y i is the actual value

[0040]

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] The present invention extracts the spectral characteristic bands of rice leaves through correlation analysis method to reduce the redundancy and optimize the data of full-spectrum data modeling, and constructs the inversion model of rice leaf nitrogen concentration and SPAD value based on linear regression (LR), random forest (RF), support vector regression (SVR) and gradient boosting (GBRT). By comparing the accuracies of different models, the best model suitable for estimating rice leaf nitrogen concentration and SPAD value is determined, in order to provide theoretical basis and technical support for rapid monitoring of nitrogen and precise nutrient management of field rice in the double-season rice area in South China. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Correlation coefficients between spectral reflectance and randomly constructed NDSI and leaf SPAD values ​​and leaf nitrogen concentration at four key growth stages

[0044] Figure 2 This is the prediction diagram of leaf SPAD value based on the model of reflectance and NDSI (R497, R654) during the whole growth period;

[0045] Figure 3 This is the prediction chart of leaf nitrogen concentration based on the model of reflectance and NDSI (R729, R730) during the entire growth period. DETAILED DESCRIPTION

[0046] The following will describe the implementation methods of the present application in detail with the help of accompanying drawings and examples, so that the implementation process of how the present application uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0047] 1. Experimental part

[0048] 1.1 Test materials and soil physical and chemical properties

[0049] The test variety is Meixiangzhan No. 2, a high-quality indica conventional rice variety with the largest planting area in South China. The test soil comes from long-term rice soil, which is dried, rolled, sieved, weighed, and potted for subsequent tests. Among them, the soil pH is 6.26, the organic matter content is 23.25g / kg, the total nitrogen is 1.055g / kg, the total phosphorus is 1.068g / kg, the total potassium is 19.39g / kg, the alkaline nitrogen is 99.54mg / kg, the available phosphorus is 69.80mg / kg, and the available potassium is 50.00mg / kg.

[0050] 1.2 Hyperspectral data acquisition

[0051] For rice PI, HD, MF and MA, clear and cloudless weather was selected, and the measurement time was 10:00-13:00. With the help of the American ASDField Spec4 ground feature spectrometer, the built-in handheld leaf clip leaf spectrum detector and built-in quartz halide lamp light source were used for measurement. The band range of the spectrometer is 350-2500nm, among which the sampling interval in the range of 350-1000nm is 1.4nm, and the sampling interval in the range of 1000-2500nm is 1.1nm. The first leaf that is fully expanded at the top of the rice was selected to measure the upper, middle and lower parts of the leaf. During the measurement, the leaf was placed in the leaf chamber of the leaf clip to ensure that the leaf was horizontal and the detected area was the same. Whiteboard correction was performed before each repeated data collection, and 10 data points were measured repeatedly for each data point, and the average value was taken as the leaf spectrum data of the point.

[0052] 1.3 Leaf SPAD value determination

[0053] The SPAD value of rice was measured using a SPAD-502 chlorophyll meter. The measurement site was consistent with the spectrum collection site, avoiding leaf veins. The upper, middle and lower parts of the leaf were measured and the average value was calculated to obtain the leaf SPAD value.

[0054] 1.4 Analysis of leaf nitrogen concentration

[0055] The rice leaves of each replicate were collected, placed in an oven at 105°C for 30 min, and then dried at a constant temperature of 75°C until constant. After grinding, the nitrogen concentration of the rice leaves was determined by the Kjeldahl method.

[0056] 1.5 Spectral index construction

[0057] The spectral data were processed using the View Spec Pro software that comes with the spectrometer. According to the range of the sensitive bands, the reflectance of each band was randomly combined in pairs to construct the normalized difference spectral index (NDSI).

[0058] NDSI=(R1-R2) / (R1+R2) (1)

[0059] In formula (1), R1 and R2 are the spectral reflectances of different bands.

[0060] 1.6 Leaf spectral sensitivity band screening

[0061] In this study, the correlation analysis method was used to select sensitive bands. When the absolute value of the correlation coefficient was greater than 0.6, it was considered to have a strong correlation. The optimal sensitive bands for leaf nitrogen concentration were 541nm, 542nm and 712nm. The optimal sensitive bands for leaf SPAD value were 603nm, 702nm and 703nm. 497-730nm was the sensitive band for leaf nitrogen concentration and SPAD value during the whole growth period.

[0062] 1.7 Basic Principles of the Algorithm

[0063] 1.7.1 Linear Regression (LR)

[0064] LR is a classic modeling method that can explain the linear relationship between a single factor variable and multiple independent variables. The model result is unique and can well reflect the degree of correlation between each independent variable and the dependent variable. The linear regression (LR) calculation formula can be expressed by the following formula:

[0065] The univariate linear regression (SLR) model function expression is as follows:

[0066] Y=β0+β1X (2)

[0067] In formula (2), Y is the estimated value of the model output, β0 is the intercept, and β1 is the prediction coefficient, which can be calculated using the least squares method.

[0068] Multiple linear regression (MLR) uses multiple correlated independent variables to solve the regression problem. Its model function expression is:

[0069] Y=β0+β1X1+β2X2+…+βmXm (3)

[0070] In formula (3), similar to the univariate linear regression model function, m is the total number of independent variables X used, β0 is the intercept, and β1…βm are the prediction coefficients corresponding to each independent variable.

[0071] 1.7.2 Random Forest (RF)

[0072] RF is a tree structure classifier proposed by Breiman. Its principle is based on classification and regression trees. It does not need to screen variables and can output the importance of features. It uses the bootsrap resampling method to extract multiple samples from the original sample, builds a decision tree model for each bootsrap sample, and then combines multiple decision trees to obtain the final prediction result through voting. The RF algorithm is insensitive to noisy data and has strong robustness. It can be used to solve nonlinear regression problems and has the characteristics of being interpretable, requiring fewer samples, not overfitting, and having high accuracy. The calculation process of random forest (RF) is as follows:

[0073] Extract training data set: Randomly extract a certain amount of samples from the original data set as training samples, which will be used to generate decision trees. When building the model, randomly divide it according to the ratio of training set: test set = 7:3.

[0074] Feature selection: For each decision tree node, when a feature needs to be tested at that node, only the best features are selected from a subset of all features. In this experiment, all tested features include the input best band spectral reflectance.

[0075] Decision tree generation: Generate multiple decision trees based on feature selection methods, and optimize the performance of decision trees by adjusting the maximum depth of the tree, the number of trees, etc.

[0076] Decision tree integration: All generated decision trees are combined by weighting and other methods to obtain the final classification result. Through the integration of decision trees, a classification prediction model is finally generated to predict the leaf nitrogen concentration and SPAD value of rice.

[0077] 1.7.3 Support Vector Machine Regression (SVR)

[0078] Support vector regression (SVR) is an application model of support vector machine (SVM) in regression problems. Its core idea is to find a hyperplane so that the total deviation of all sample points from the hyperplane is minimized. The support vector regression (SVR) model formula is as follows:

[0079]

[0080] In formula (4), i represents the i-th sample point, ai is the Lagrange coefficient; represents the mean of the random variable, the estimated value; M represents the sequence of data vectors; b is the threshold, representing the kernel function of the SVM model, and T represents the state.

[0081] 1.7.4 Gradient Boosting (GBRT)

[0082] Gradient boosting (GBRT) ensemble learning algorithm is used for supervised learning tasks. Gradient boosting trains multiple weak learners, using an iterative method to train each weak learner on the residual of the previous learner and gradually adjust its weights to minimize the loss function, eventually generating a strong learner. The gradient boosting (GBRT) calculation formula can be expressed by the following formula:

[0083]

[0084] In formula (5), T(x, θ m ) represents a decision tree; θ m is the parameter of the decision tree; M is the number of trees. Gradient boosting tree regression can flexibly process various types of data, capture complex nonlinear relationships in the data, handle interactions, and make the model more robust with less parameter adjustment.

[0085] Table 1 Model parameters

[0086]

[0087]

[0088] 1. SVR uses grid search to determine the parameters "c" and "gamma". The search range of "c" is 0.01, 0.1, 1, 10, 100, 110; the search range of "gamma" is "scale, auto, 0.01, 0.1, 1, 10, 100"; the kernel function is rbf.

[0089] 2. GBRT uses grid search to determine the parameters "n_estimators", "learning_rate", "max_depth", and "subsample". The search range for "n_estimators" is 50, 100, 150; the search range for "learning_rate" is 0.1, 0.2, 0.15, 1; the search range for "max_depth" is 1, 2, 3, 4, 5, 7; and the search range for "subsample" is 0.5, 0.7, 0.8, 1.

[0090] 1.8 Model Evaluation

[0091] The determination coefficient (R 2), root mean square error (RMSE) and normalized root mean square error (NRMSE) to test the accuracy of the model. 2 To evaluate the model fit, R 2 The larger the value, the better the model fit; RMSE and NRMSE are used to evaluate model stability. The smaller the RMSE, the more stable the model. NRMSE less than 10% indicates very good stability, 10% to 20% indicates good stability, 20% to 30% indicates average stability, and greater than 30% indicates poor stability.

[0092] The calculation formulas for each evaluation index are as follows, where n is the number of samples. is the predicted value, is the mean value, y i is the actual value

[0093]

[0094]

[0095] 2. Results

[0096] 2.1 Sensitive band selection and normalized difference spectral index construction

[0097] This study used the correlation analysis method to select sensitive bands, and when the absolute value of the correlation coefficient was greater than 0.6, it was considered to have a strong correlation. The optimal sensitive bands for leaf nitrogen concentration were 541nm, 542nm and 712nm. The optimal sensitive bands for leaf SPAD values ​​were 603nm, 702nm and 703nm. 497-730nm is the sensitive band for leaf nitrogen concentration and SPAD value during the entire growth period. The reflectance of each band was randomly combined in pairs to construct the normalized difference spectral index (NDSI), and the calculation formula is as follows. NDSI (R497, R654) has the strongest correlation with leaf SPAD value, R=0.97, and NDSI (R729, R730) has the strongest correlation with leaf nitrogen concentration, R=-0.90, such as Figure 1 shown.

[0098] Figure 1 In the figure, A and C are the correlation analyses between the spectral reflectance during the whole growth period and the leaf SPAD value and leaf nitrogen concentration; B and D are the correlation analyses between the randomly constructed NDSI and the leaf SPAD value and leaf nitrogen concentration, respectively.

[0099] 2.2 Model construction

[0100] The model constructed with RF has the best effect when using the best sensitive band and NDSI as input. The inverted leaf SPAD value R 2 are all 0.99, and the NRMSE are 2.99% and 3.55% respectively; the inversion of leaf nitrogen concentration R 2The RF algorithm has a great advantage in inverting SPAD value and leaf nitrogen concentration.

[0101] Table 2 Linear regression analysis of spectral reflectance and NDSI with SPAD value and leaf nitrogen concentration

[0102]

[0103] Table 3 Construction of leaf SPAD value and nitrogen concentration model

[0104]

[0105] 2.3 Leaf SPAD value verification

[0106] 70% of the data set was randomly selected as the training set and 30% as the test set, and the spectral reflectance and NDSI of the whole growth period were used as input for model verification. The four algorithms all had good inversion effects on the SPAD value. 2 All of them are above 0.90, and NRMSE is less than 10%. Taking the reflectance of the best sensitive band in the whole growth period as input value, the random forest algorithm has the best effect, R 2 , RMSE and NRMSE were 0.98, 1.95 and 4.88% respectively. With NDSI (R497, R654) as the input value, the support vector machine algorithm has the best effect. 2 , RMSE and NRMSE are 0.97, 2.53 and 6.28% respectively. Figure 2 shown.

[0107] Figure 2 In the figure, AD is the prediction of leaf SPAD value based on the full growth period model; EH is the prediction of leaf SPAD value based on the NDSI model. A and E are MLR models; B and F are RF models; C and G are SCR models; D and H are GBRT models.

[0108] 2.4 Verification of leaf nitrogen concentration

[0109] The four algorithms all have good inversion effects on leaf nitrogen concentration. 2 The RF model has the best performance when the reflectance of the best sensitive band in the whole growth period is used as the input value. 2 , RMSE and NRMSE are 0.83, 4.05 and 12.16% respectively. Using NDSI (R729, R730) as input value, GBRT algorithm has the best effect. 2 , RMSE and NRMSE were 0.90, 3.09 and 9.36% respectively ( Figure 3 ).

[0110] Figure 3 In the figure, AD is the prediction of leaf nitrogen concentration based on the model of full growth period reflectance; EH is the prediction of leaf nitrogen concentration based on the model of NDSI. A and E are MLR models; B and F are RF models; C and G are SCR models; D and H are GBRT models.

[0111] The results showed that leaf SPAD value and nitrogen concentration were significantly negatively correlated with spectral reflectance at 497nm-730nm and 503nm-727nm, respectively, and the absolute values ​​of the correlation coefficients were greater than 0.6. On this basis, the normalized difference spectral index (NDSI) was randomly constructed pairwise and the correlation analysis was performed with leaf SPAD value and nitrogen concentration. It was found that NDSI (R497, R654) had the strongest correlation with leaf SPAD value, R=0.97, and NDSI (R729, R730) had the strongest correlation with leaf nitrogen concentration, R=-0.90. The model constructed by RF and GBRT was significantly better than MLR and SVR. When inverting leaf SPAD value and leaf nitrogen concentration, the spectral reflectance of the whole growth period was used as the input value. The model constructed by random forest algorithm had the best effect, R 2 The results are 0.99 and 0.98, RMSE is 1.18 and 1.46, and NRMSE is 2.99% and 4.61%. The results can provide a scientific basis for non-destructive real-time monitoring of leaf color and precise fertilization management of indica rice in South China.

[0112] In this study, the spectral characteristic bands of rice leaves were extracted by correlation analysis to reduce the redundancy and optimize the data for full-spectrum data modeling. An inversion model of nitrogen concentration and SPAD value of rice leaves was constructed based on linear regression (LR), random forest (RF), support vector regression (SVR) and gradient boosting (GBRT). The accuracies of different models were compared to determine the best model suitable for estimating nitrogen concentration and SPAD value of rice leaves, in order to provide a theoretical basis and technical support for rapid monitoring of nitrogen and precise nutrient management of field rice in the double-season rice area in South China.

[0113] In order to quickly and accurately detect the nitrogen concentration and SPAD value of rice leaves in the double-season rice area in South China, analyze the growth of rice, and provide a basis for precise fertilization and scientific management of rice fields, Meixiangzhan No. 2 was taken as the research object, and the ASDField Spec4 portable ground feature spectrometer was used to obtain the hyperspectral reflectance data of rice leaves, the SPAD value of rice leaves was measured by SPAD-502 chlorophyll meter, and the leaf nitrogen concentration was measured by Kjeldahl method. Correlation analysis method was used for feature selection.

[0114] The sensitive band of leaf nitrogen concentration screened by correlation analysis was 503nm-727nm, of which there were three characteristic bands, namely 541nm, 542nm and 712nm; the sensitive band of leaf SPAD value screened was 497nm-730nm, of which there were three characteristic bands, namely 602nm, 603nm and 703nm. The normalized difference spectral vegetation index (NDSI) was constructed by randomly combining two bands in the sensitive band range (497nm-730nm), and the NDSI with the strongest correlation with leaf nitrogen concentration (R729, R730) and the NDSI with the strongest correlation with leaf SPAD value (R497, R654) were obtained.

[0115] The reflectance and NDSI of the characteristic bands were used as input to construct the LR, RF, SVR, and GBRT inversion models of rice leaf nitrogen concentration and SPAD value, respectively.

[0116] The results showed that leaf nitrogen concentration and SPAD value were significantly negatively correlated with spectral reflectance at 497-730nm, and the absolute values ​​of the correlation coefficients were greater than 0.6. On this basis, by randomly constructing the normalized difference spectral index (NDSI) pairwise, it was found that NDSI (R497, R654) had the strongest correlation with leaf SPAD value, R=0.97, and NDSI (R729, R730) had the strongest correlation with leaf nitrogen concentration, R=-0.90. The model constructed using RF and GBRT was significantly better than MLR and SVR. When inverting leaf SPAD value and leaf nitrogen concentration, the spectral reflectance of the whole growth period was used as the input value. The model constructed using the random forest algorithm had the best effect, R 2 The SPAD values ​​and the characteristic spectra of leaf nitrogen concentration of indica rice varieties are significantly different, and the model constructed by RF algorithm has the highest accuracy and the best stability. The research results can provide a scientific basis for non-destructive real-time monitoring of leaf color and precise fertilization management of indica rice in South China.

[0117] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the present invention in any form. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention fall within the protection scope of the present invention.

[0118] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are within the scope of the technical solution of the present invention.

Claims

1. A method for inverting SPAD values ​​and nitrogen concentrations of rice leaves during the whole growth period based on hyperspectral technology, characterized in that: The spectral characteristic bands of SPAD value and nitrogen concentration of rice leaves were extracted by correlation analysis method. The inversion model of nitrogen concentration and SPAD value of rice leaves was constructed based on linear regression, random forest, support vector regression and gradient boosting. The characteristic bands and inversion models of leaf nitrogen concentration and SPAD value were compared. The specific steps include the following: S1. Obtain hyperspectral data of rice leaves at different growth stages, measure the SPAD value of rice leaves at the same time, and then measure the nitrogen concentration of rice leaves through destructive sampling; S2, perform correlation analysis on leaf hyperspectral data, leaf nitrogen concentration and SPAD value to select sensitive bands; S3, randomly constructing normalized difference spectral index (NDSI) in pairs within the sensitive band range, and performing correlation analysis between the constructed NDSI and leaf nitrogen concentration and SPAD value to obtain the spectral index with the strongest correlation with leaf nitrogen concentration and SPAD value; S4. Use sensitive bands and NDSI as input to construct rice leaf nitrogen concentration and SPAD value inversion models with different algorithms, and compare the differences in the inversion models of leaf nitrogen concentration and SPAD value; S5. Divide the training set and the test set in proportion to verify the inversion model.

2. The method for inverting SPAD values ​​and nitrogen concentrations of rice leaves during the whole growth period based on hyperspectral technology according to claim 1, characterized in that: In the step S1, clear and cloudless weather is selected at different growth stages of rice, the measurement time is 10:00-13:00, and the handheld leaf-clip spectrum detector and quartz halide light source provided by the spectrometer are used for measurement. The first leaf that is fully unfolded at the top of the rice is selected, and the upper, middle and lower parts of the leaf are measured. During the measurement, the leaf is placed in the leaf chamber of the leaf clip, and 10 spectral curves are collected at each leaf sample point. The average value is the leaf spectrum data of the point, and the ViewSpec Pro software is used to record the data of the collected spectral reflectance. Whiteboard correction is performed before each repeated data collection, and the correction result requires that the spectral reflectance within the band range is 1. The different growth stages are the initial stage of panicle differentiation (PI), heading stage (HD), mid-filling stage (MF) and maturity stage (MA).

3. The inversion method of SPAD value and nitrogen concentration of rice leaves during the whole growth period based on hyperspectral technology according to claim 1 is characterized by: In step S2, correlation analysis is used to select characteristic spectra for the hyperspectral data, and the sensitive band of leaf nitrogen concentration and SPAD value in the whole growth period is 497nm-730nm.

4. The inversion method for SPAD value and nitrogen concentration of rice leaves during the whole growth period based on hyperspectral technology according to any one of claims 1 or 3, characterized in that: In step S2, the optimal sensitive wavelengths for leaf nitrogen concentration are 541 nm, 542 nm and 712 nm, and the optimal sensitive wavelengths for leaf SPAD value are 603 nm, 702 nm and 703 nm.

5. The method for inverting SPAD values ​​and nitrogen concentrations of rice leaves during the whole growth period based on hyperspectral technology according to claim 1, characterized in that: In step S3, according to the range of the sensitive band, the reflectance of each band in the sensitive band is randomly combined in pairs to construct a normalized difference spectral index (NDSI), and the calculation formula is as follows: NDSI=(R1-R2) / (R1+R2) (1) In formula (1), R1 and R2 are the spectral reflectances of any band within the optimal sensitive band of leaf nitrogen concentration. NDSI (R497, R654) has the strongest correlation with leaf SPAD value, and NDSI (R729, R730) has the strongest correlation with leaf nitrogen concentration.

6. The method for inverting SPAD values ​​and nitrogen concentrations of rice leaves during the whole growth period based on hyperspectral technology according to claim 1, characterized in that: In step S4, the reflectivity corresponding to the characteristic wavelength selected in step S2 and the NDSI constructed in step S2 are used as independent variables of the inversion model, and the nitrogen concentration of rice leaves and the SPAD value are used as target variables to construct inversion models of four algorithms, namely linear regression (LR), random forest (RF), support vector machine (SVR) and gradient boosting (GBRT).

7. The method for inverting SPAD values ​​and nitrogen concentrations of rice leaves during the whole growth period based on hyperspectral technology according to claim 1, characterized in that: In step S5, the train_test_split function in the sklearn.model_selection module in Python is used to divide the training set and the test set to verify the model.

8. The method for inverting SPAD values ​​and nitrogen concentrations of rice leaves during the whole growth period based on hyperspectral technology according to claim 6, characterized in that: The linear regression (LR) calculation formula is as follows: The univariate linear regression (SLR) model function expression is as follows: Y=β0+β1X(2) In formula (2), Y is the estimated value of the model output, while β0 is the intercept and β1 is the prediction coefficient, which can be calculated using the least squares method; Multiple linear regression (MLR) uses multiple correlated independent variables to solve the regression problem. Its model function expression is: Y=β0+β1X1+β2X2+…+βmXm (3) In formula (3), similar to the univariate linear regression model function, m is the total number of independent variables X used, β0 is the intercept, and β1…βm are the prediction coefficients corresponding to each independent variable; The random forest (RF) calculation process is as follows: Extract training data set: randomly extract a certain amount of samples from the original data set as training samples. These samples will be used to generate decision trees. When building the model, they are randomly divided according to the ratio between the training set and the test set. Feature selection: For each decision tree node, when a feature needs to be tested at that node, only the best feature is selected from a subset of all features; Decision tree generation: Generate multiple decision trees based on feature selection methods, and optimize the performance of decision trees by adjusting the maximum depth of the tree, the number of trees, etc. Decision tree integration: All generated decision trees are combined by weighting and other methods to obtain the final regression results. Through the integration of decision trees, a prediction model is finally generated to predict the leaf nitrogen concentration and SPAD value of rice; When the full growth period spectral reflectance is used as input to invert SPAD values, the Random_state for dividing the training set and the test set is set to 10, and n_estimators is set to 20; when the NDSI is used as input to invert SPAD values, the Random_state for dividing the training set and the test set is set to 42, and n_estimators is set to 100; when the full growth period spectral reflectance is used as input to invert leaf nitrogen concentration, the Random_state for dividing the training set and the test set is set to 25, and n_estimators is set to 80; when the NDSI is used as input to invert SPAD values, the Random_state for dividing the training set and the test set is set to 25, and n_estimators is set to 100; The support vector regression SVR model formula is as follows: In formula (4), i represents the i-th sample point, a i is the Lagrange coefficient; represents the mean of random variables, the estimated value; M represents the sequence of data vectors; b is the threshold, representing the kernel function of the SVM model, and T represents the state; When using the full growth period spectral reflectance and NDSI as input to invert the SPAD value, the Random_state for dividing the training set and the test set was set to 10; when inverting the leaf nitrogen concentration, the Random_state for dividing the training set and the test set was set to 80; grid search was used to determine the parameters "c" and "gamma", and the search range of "c" was 0.01, 0.1, 1, 10, 100, 110; the search range of "gamma" was "scale, auto, 0.01, 0.1, 1, 10, 100"; the kernel function was rbf. The gradient boosting (GBRT) calculation formula is as follows: In formula (5), T(x,θ m ) represents the decision tree; θm is the parameter of the decision tree; M is the number of trees. When using the full growth period spectral reflectance and NDSI as input to invert SPAD values, the Random_state for dividing the training set and the test set was set to 42; when inverting leaf nitrogen concentration, the Random_state for dividing the training set and the test set was set to 45; grid search was used to determine the parameters "n_estimators", "learning_rate", "max_depth", and "subsample". The search range for "n_estimators" is 50, 100, and 150; the search range for "learning_rate" is 0.1, 0.2, 0.15, and 1; the search range for "max_depth" is 1, 2, 3, 4, 5, and 7; and the search range for "subsample" is 0.5, 0.7, 0.8, and 1.

9. The method for inverting SPAD values ​​and nitrogen concentrations of rice leaves during the whole growth period based on hyperspectral technology according to claim 6, characterized in that: The evaluation index of the inversion model adopts the determination coefficient R 2 , root mean square error RMSE and normalized root mean square error NRMSE, the calculation formulas of each evaluation index are as follows, where n is the number of samples, is the predicted value, is the mean value, y i is the actual value 10. The method for inverting SPAD values ​​and nitrogen concentrations of rice leaves during the whole growth period based on hyperspectral technology according to any one of claims 1 or 7, characterized in that: The ratio of the training set to the test set is 7:3.

Images