Nondestructive monitoring method for calcium content of karst plant leaves

Abstract

A nondestructive monitoring method for calcium content of karst plant leaves, including the following steps: acquiring spectral reflectance data and calcium content data of sample plant leaves; preprocessing the spectral reflectance data of the sample plant leaves to obtain preprocessed data; performing Pearson's correlation test on the preprocessed data and the calcium content data to obtain the eliminated spectral bands; using a minimum absolute shrinkage and selection operator to obtain the compressed spectral bands; using a principal component analysis method to obtain the transformed spectral bands; inputting the transformed spectral bands and the calcium content data into an initial network for training, optimizing the parameters of the initial network by using a preset optimization algorithm to obtain a final leaf calcium content monitoring model; inputting the spectral reflectance data of the plant leaves to be detected into the leaf calcium content monitoring model to obtain the final calcium content data.

Description

TECHNICAL FIELD The invention relates to the technical field of ecological remote sensing monitoring, in particular to a . BACKGROUND Karst landforms are widely distributed in China, and karst vegetation is rich and diverse. However, the karst ecological environment is very fragile, and the restoration and conservation of vegetation is particularly important. Calcium is one of the indispensable nutrient elements in the process of plant growth and development, which plays an important role in stabilizing the structure of plant biofilm, maintaining the integrity of cells and enhancing the stability and strength of cell walls. The traditional calcium content is obtained by chemical extraction in the laboratory. Although this method has high accuracy, it is timeconsuming and laborious, and it needs to destroy the living plants, so it is difficult to meet the demand for rapid diagnosis of calcium content in large area of vegetation region. In recent years, the development of hyperspectral remote sensing technology provides a new idea and method for rapid extraction of calcium content in plant leaves. However, the existing research mainly focuses on the rapid extraction of specific species or relatively single vegetation types. The existing methods and technologies are not applicable for the vast karst areas. For mixed plants, the accuracy of extraction has always been a key problem in extracting calcium content from plant leaves by hyperspectral remote sensing technology. The rapid estimation of calcium content in plant leaves in karst area is realized by simple linear regression methods such as spectral index method, stepwise regression method and partial least square method, which is difficult to meet the requirements in accuracy. The relationship between spectral reectance and calcium content in plant leaves can be fitted by nonlinear models with high fitting ability such as neural network, and the accuracy can be satisfied. However, due to the high dimensional and complex characteristics of spectral data, the established model usually has overfitting phenomenon. Therefore, how to solve the problem of overfitting has become the key to the rapid extraction of calcium content in karst plant leaves. On the other hand, spectral differentiation technology has been proved to be helpful in improving the extraction accuracy of calcium content in plant leaves. However, integerorder differential (usually referred to as firstorder differential and secondorder differential) will make the spectrum jump in geometric shape, so it is impossible to fully tap the potential of spectral data for rapid extraction of calcium content in plant leaves. Fractional differential makes up for this deficiency of integer differential. The disadvantage of the prior art is that the chemical extraction method needs various steps such as dry, grinding, digestion and the like, which is timeconsuming and laborious. The existing methods of extracting calcium content from plant leaves by hyperspectral technology are mainly aimed at a single plant, which is difficult to adapt to complex and diverse vegetation types in karst areas, easy to overfit and has low prediction accuracy. SUMMARY In order to overcome the shortcomings of the prior art, the objective of the invention is to provide a , which has important practical significance and broad application value for vegetation growth monitoring and ecological environment protection in karst areas. In order to achieve the above objectives, the present invention provides the following scheme. The invention relates to a , including the following steps: acquiring spectral reectance data and calcium content data of sample plant leaves; preprocessing the spectral reectance data of the sample plant leaves by using fractional differential technology to obtain preprocessed data; performing Pearson's correlation test on the preprocessed data and the calcium content data to eliminate spectral bands with insignificant correlation between the spectrum and the calcium content of leaves to obtain the eliminated spectral bands; using a minimum absolute shrinkage and selection operator to compress the eliminated spectral bands to obtain the compressed spectral bands; using a principal component analysis method to perform principal component transform on the compressed spectral bands to compress the spectral bands and obtain the transformed spectral bands; inputting the transformed spectral bands and the calcium content data into an initial network for training, and optimizing the parameters of the initial network by using a preset optimization algorithm to obtain a final leaf calcium content monitoring model; inputting the spectral reectance data of the plant leaves to be detected into the leaf calcium content monitoring model to obtain the final calcium content data. Preferably, obtaining the spectral reectance data and the calcium content data of the sample plant leaves comprises: collecting the spectral reectance data of the sample plant leaves in the field by using the ground object spectrometer; extracting the calcium content data of the sample plant leaves by chemical extraction. Preferably, preprocessing the spectral reectance data of the sample plant leaves by using fractional differential technology to obtain the preprocessed data, comprising: performing fractional differential processing on the spectral reectance data of the sample plant leaves from 0 to 3, with an interval of 0.1; the calculation formula of the fractional differential processing is: dv  v v+l Fv+l %:f(Â)+(V)f(Â1)+L2)f(Â2)+L +mmm wherein  is the wavelength of the spectral reectance data, n is the wavelength interval difference of differential, v is the fractional differential order, f (g) is a one dimensional spectrum, and F(g) is the Gamma function, and the calculation formula of the Gamma function is F() z-ÍO e [r ldt =(ß_1)!; wherein ß is an arbitrary variable and t is a differential variable. Preferably, the spectral bands with insignificant correlation are the spectral bands with P value greater than 0.05 by Pearson correlation significance test. Preferably, the calculation formula of the minimum absolute shrinkage and selection operator is: 2 6% = arga min {É [yi Ë xaj] + êÉ lûd} i=1 j=1 j=1 wherein x is the eliminated spectral band, y is the calcium content data, y i is the i th observation value, x is the jth feature of the ith observation value, a is the regression coefficient, 8 is the regularization parameter, m is the number of observation values, and n is the number of features. Preferably, the initial network is a generalized regression neural network. Preferably, the parameters of the generalized regression neural network comprise spread parameters. Preferably, the optimization algorithm is a particle swarm optimization algorithm. According to the specific embodiment provided by the invention, the invention discloses the following technical effects. The invention provides a , which comprises the following steps: acquiring spectral reectance data and calcium content data of sample plant leaves; preprocessing the spectral reectance data of the sample plant leaves by using fractional differential technology to obtain preprocessed data; performing Pearson's correlation test on the preprocessed data and the calcium content data to eliminate spectral bands with insignificant correlation between the spectrum and the calcium content of leaves to obtain the eliminated spectral bands; using a minimum absolute shrinkage and selection operator to compress the eliminated spectral bands to obtain the compressed spectral bands; using a principal component analysis method to perform principal component transform on the compressed spectral bands to compress the spectral bands and obtain the transformed spectral bands; inputting the transformed spectral bands and the calcium content data into an initial network for training, and optimizing the parameters of the initial network by using a preset optimization algorithm to obtain a final leaf calcium content monitoring model; inputting the spectral reectance data of the plant leaves to be detected into the leaf calcium content monitoring model to obtain the final calcium content data. According to the invention, hyperspectral remote sensing technology and AI technology are organically integrated, so that the calcium content of plant leaves can be quickly and accurately estimated only by collecting the spectral reectance of the plant leaves. The invention has important practical significance and broad application value for vegetation growth monitoring and ecological environment protection in karst areas. BRIEF DESCRIPTION OF THE FIGURES In order to explain the embodiments of the present invention or the technical scheme in the prior art more clearly, the drawings needed in the embodiments will be briey introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For ordinary people in the field, other drawings can be obtained according to these drawings without paying creative labor. Fig. 1 is a owchart of a method provided by an embodiment of the present invention; Fig. 2 is a technical roadmap provided by an embodiment of the present invention; Fig. 3 is a schematic diagram of case sample distribution provided by an embodiment of the present invention; Fig. 4 is a schematic diagram of the band distribution of different fractional differential spectral signals passing Pearson correlation significance test provided by the embodiment of the present invention; Fig. 5 is a schematic diagram of the prediction accuracy of the method (LASSO+PCA+PSO+GRNN) provided by the embodiment of the present invention for calcium content in plant leaves under different fractional differential conditions; Fig. 6 is a schematic diagram of the prediction accuracy of PLSR model for calcium content in plant leaves under different fractional differential conditions; Fig. 7 is a schematic diagram of the prediction accuracy of GRNN model for calcium content in plant leaves under different fractional differential conditions; Fig. 8 is a schematic diagram of the prediction accuracy of the LASSO+PCA+GRNN model for calcium content in plant leaves under different fractional differential conditions. DESCRIPTION OF THE INVENTION In the following, the technical scheme in the embodiment of the invention will be clearly and completely described with reference to the attached drawings. Obviously, the described embodiment is only a part of the embodiment of the invention, but not the whole embodiment. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in the field without creative labor belong to the scope of protection of the present invention. The objective of the invention is to provide a , which has important practical significance and broad application value for vegetation growth monitoring and ecological environment protection in karst areas. In order to make the above objects, features and advantages of the present invention more obvious and easy to understand, the present invention will be further described in detail with the attached drawings and specific embodiments. Fig. 1 is a owchart of a method provided by an embodiment of the present invention. As shown in Fig. l, the present invention provides a , including: S 100: acquiring spectral reectance data and calcium content data of sample plant leaves; S 200: preprocessing the spectral reectance data of the sample plant leaves by using fractional differential technology to obtain preprocessed data; S 300: performing Pearson's correlation test on the preprocessed data and the calcium content data to eliminate spectral bands with insignificant correlation between the spectrum and the calcium content of leaves to obtain the eliminated spectral bands; S 400: using a minimum absolute shrinkage and selection operator to compress the eliminated spectral bands to obtain the compressed spectral bands; S 500: using a principal component analysis method to perform principal component transform on the compressed spectral bands to compress the spectral bands and obtain the transformed spectral bands; S 600: inputting the transformed spectral bands and the calcium content data into an initial network for training, and optimizing the parameters of the initial network by using a preset optimization algorithm to obtain a final leaf calcium content monitoring model; S 700: inputting the spectral reectance data of the plant leaves to be detected into the leaf calcium content monitoring model to obtain the final calcium content data. Specifically, the technical route of this embodiment is shown in Fig. 2, which specifically includes: Process 1: the spectrum of plant leaves is collected in the field by using a ground object spectrometer (such as ASD FieldSpec 4 ground object spectrometer). For tall trees, branches are cut off with averruncator and then measurement is carried out, but the interval should not be too long. It is best to complete the spectral reectance measurement within 10 minutes after the branches are cut off, so as to avoid the physiological and biochemical characteristics inside the leaves from changing too much after the branches are detached, which will affect the measurement accuracy of the spectral reectance of the leaves. After the spectra of leaves are measured, they are collected and brought back to the laboratory, and the calcium content of plant leaves is extracted by traditional chemical extraction methods (such as atomic absorption spectrometry). Process 2: preprocessing is carried out on the spectral reectance data obtained in Process 1 by using fractional differential technology; fractional differentiation (such as GriinwaldLetnikov algorithm) is performed on the spectral reectance of leaves from 0 to 3, with an interval of 0.1. Through preprocessing, on the one hand, the data noise is eliminated, on the other hand, the prediction potential of spectrum for calcium content in plant leaves is improved. The calculation formula of GriinwaldLetnikov algorithm is as follows: dv  v v+l Fv+l %zf(Â)+(V)f(Â1)+L2)f(Â2)+L +mmm (1) In formula (1),  is the wavelength of spectral reectance, n is the wavelength interval difference of differential, and v represents the fractional differential order (0<n<v<n+1). F(g) is Gamma function as follows: nm 210 e ft 1alt=(1)! @) In formula (2), ß can be any variable, and here refers to the fractional differential order. Theoretically, fractional differential can be infinite, but highorder differential is very sensitive to highfrequency signals, which may lead to distortion of spectral signals, so the differential is selected from 0 to 3. In addition, in order to fully tap the characteristic signal in the spectral reectance, the interval is positioned at 0.1. Process 3: Pearson's correlation test is performed on the preprocessed data in the Process 2 and the calcium content data in the plant leaves in the Process 1, and spectral bands with insignificant correlation between the spectrum and the calcium content in the leaves are removed; the spectral bands with insignificant correlation are the spectral bands with Pearson correlation significance test P value greater than 0.05. Process 4: the spectral band selected in Process 3 is further compressed by using LASSO (Least Absolute Shrinkage and Selection Operator) method. The number of spectral bands that can pass the Pearson correlation significance test varies from 800 to 1800, so it is easy to overfit the model if it is directly used to establish the neural network model. LASSO makes some regression coefficients zero through penalty term (regularization), thus identifying and optimizing characteristic variables and realizing further compression of spectral bands. The basic idea of LASSO is to introduce Ll regularization term on the basis of linear regression, and to achieve feature selection and model compression by adding a constant constraint to the sum of absolute values of regression coefficients. The mathematical expression of LASSO is: m n 2 n â : arg a min {2 [yi Z xl. / aj} + 52 ad} (3) i=1 j=1 j=1 In formula (3), x is the independent variable, i.e. the spectral band, y is the dependent variable, i.e. the calcium content; y [ is the ith observation value, xii is the j th feature of the ith observation value, a is the regression coefficient, 6 is the regularization parameter, which controls the regularization intensity, and m and n respectively represent the number of observed values and the number of characteristics. In n 2 Z [y, _ Z xijaj] There is no significant difference between i:] H in the first half of 52 af formula (3) and ordinary linear regression, and the introduction of «le regularization in the second half pushes some regression coefficients to zero, thus realizing feature selection. Process 5: Principal Component Analysis (PCA) is used to transform the spectral band selected in Process 4, and the spectral data is compressed again. The number of bands filtered by LASSO still varies from 20 to 129, and it is still easy to cause over fitting of the model when it is directly used to establish the neural network model. PCA can reduce the dimension and simplify the data by projecting the original data set into a lower dimensional space, while retaining the main changing trends and features in the data set. Through PCA transform, the data can be further compressed while retaining the spectral information to the maximum. Process 6: the principal component with cumulative contribution rate of 75% after the Process 5 is taken as the independent variable, and the calcium content of plant leaves obtained in Process 1 is taken as the dependent variable. GRNN (General regression neural network) is optimized by PSO (Particle Swarm Optimization) algorithm, and a combined neural network model (LASSO+PCA+PSO+GRNN) is established. If the accuracy of the established model is high (the decisive coefficient R2>0.6) and the operation is stable (the difference between the training R2 and the verification R2 is within 0.15), the modeling is completed; if the accuracy cannot meet the requirements, the modeling samples are added and the Process 1 to Process 4 are repeated until the model is stable. GRNN has good nonlinear mapping ability, strong robustness and fault tolerance, but the spread parameters in GRNN have great inuence on the model and can only be adjusted by experience, and GRNN may lead the model to fall into a local optimal solution instead of a global optimal solution when the initial solution is not properly selected. PSO is a global optimization algorithm. Optimizing the spread parameters of GRNN model by PSO can improve the overall performance of the model. Process 6, the spectral reectance data obtained by measuring the spectra of plant leaves is input into the combined neural network model established in Process 5, so as to reversely deduce the calcium content data of the corresponding plant leaves. The invention has the following beneficial effects. (1) Compared with the method of chemical extraction of calcium from plant leaves in the laboratory, this method is not only fast, but also harmless to the leaves themselves. (2) The invention realizes the monitoring of complex plant samples, and overcomes the defect of poor transferability of the model established by a single plant sample. For example, if the model based on wheat monitoring is applied to rice, the prediction performance may be greatly reduced. However, this method is based on 301 samples (including 37 families, 59 genera and 68 species, and the distribution of calcium content in samples is shown in Fig. 3) collected in karst areas of Guangxi, and the model runs stably and has good application effect. From Process 1 and Process 5. (3) The invention adopts fractional differential to transform the spectral reectance, which improves the richness of spectral information and also improves the correlation between spectral signals and calcium content of plant leaves. In process 2, the correlation coefficient distribution between the fractional spectral signal processed by GrünwaldLetnikov algorithm and calcium content is shown in Fig. 4. Among them, the band distribution of different fractional differential spectral signals passed the Pearson correlation significance test (p<0.05), the blank part indicated that it failed the significance test (p>0.05), and FD in the figure indicated fractional differential, for example, FD(0.5) indicated that the original spectral reectance was transformed by 0.5 order. (4) GRNN model has a good prediction ability, but due to the characteristics of large number of bands and high dimension of spectral data, GRNN model is prone to overfitting in prediction. Pearson's correlation test and LASSO model are used to select bands, then PCA is used to reduce the dimension of the selected bands, and finally PSO model is used to optimize the parameters of GRNN model, and a combined neural network model (LASSO+PCA+PSO+GRNN) is established. This model can not only ensure the prediction accuracy (compared with traditional regression models, such as PLSR (partial least squares regression) model, the prediction accuracy is improved by about 17%; compared with GRNN model alone, the prediction accuracy is improved by about 26%, and the overfitting phenomenon is well controlled), and the occurrence of overfitting can be reduced. As shown in steps 4, 5 and 6 from part 2, Fig. 5, Fig. 6, Fig. 7 and Fig. 8 show the accuracy comparison of the corresponding models. In Fig. 5, R2 is the decisive coefficient, and the value is between 0 and 1. The larger the value, the higher the accuracy. GRNN model will uctuate after each operation, and the vertical short vertical line is R2 standard deviation of 10 cycles, the same below. The vertical long vertical line indicates the best fractional differential position, that is, 2.6 order. Compared with this method in Fig. 6, the verification set R2 of this method is 0.76, which is 0.17 higher than PLSR under the same optimal fractional differential condition. Compared with this method in Fig. 7, the verification set R2 of this method is 0.76, which is 0.26 higher than GRNN under the same optimal fractional differential condition. The training set R2 of GRNN model is 1.00, but the verification set R2 is only 0.50, which is quite different, and the overfitting phenomenon is obvious. Compared with GRNN model in Fig. 8, LASSO+PCA+GRNN model has a lower training set R2, and the overfitting phenomenon is well controlled, which shows that the band screening by LASSO and principal component transformation by PCA are helpful to improve the overfitting problem of GRNN model. Compared with LASSO+PCA+GRNN model, the validation set R2 of this method is 0.76, which is 0.16 higher than LASSO+PCA+GRNN model, and the accuracy has been greatly improved, which shows that PSO optimizing GRNN parameters is helpful to improve the performance of the model. l 1 Each embodiment in this specification is described in a progressive way, and each embodiment focuses on the differences from other embodiments, so it is only necessary refer to the same and similar parts between each embodiment. In this paper, specific examples are used to explain the principle and implementation of the invention, and the description of the above embodiments is only used to help understand the method and its core idea of the invention; at the same time, ordinary technicians in this field, according to the idea of the invention, there will be changes in the specific implementation and application scope. In summary, the contents this specification should not be construed as limiting the present invention.

Claims

1. Method for non-destructive monitoring of calcium content in leaves of karst plants, characterised in that the method comprises: obtaining spectral response data and calcium content data from sample leaves of the plant, preprocessing the spectral response data of the sample leaves of a plant using fractional differential technology to pre-process to obtain data, performing a Pearson correlation test on the preprocessed data and calcium content data to eliminate the spectral band in which the correlation between spectrum and calcium content of leaf is not significant, to to obtain eliminated spectral band, compressing the minimum eliminated spectral band using absolute shrinkage and selection algorithm to compressed spectral band was obtained, transforming the compressed spectral band using the principal component analysis to compress the spectral band and the to obtain transformed spectral band, entering the transformed spectral band and calcium content data in an initial network for training, and optimizing parameters of the initial network using a preset optimization algorithm to final model for monitoring and measuring leaf calcium content to acquire, entering the spectral response data of the leaves to be measured plant in the model for monitoring the calcium content of the leaves to to obtain definitive data on calcium content.

2. Method for non-destructive monitoring of calcium content in leaves of karst plants according to claim 1, characterized in that obtaining the spectral response data and calcium content data from sample leaves of the plant includes: the field collection of the spectral response data from the sample leaves of the plant using the spectrometer for ground objects, extracting the calcium content data from the plant leaf samples by chemical extraction, 3. Method for non-destructive monitoring of calcium content in leaves of karst plants according to claim 1, characterized in that the pre-treatment of the spectral response data from the sample leaves of the plant using fractional differential technology to obtain preprocessed data includes: processing the spectral response data of the sample leaves of the plant by fractional differential processing from 0 to 3 order with an interval of 0.1, where a calculation formula of the fractional differential accounting is as follows: d"f( / ì) (v)(v + 1) F(v + 1) Wx f(Â)+(v)f(Âl)+Îf(Â2)+L +mfn) where, Ä is a g length of the spectral response data, n is the difference between length intervals of the differential, v the order of the fractional is differential, f(g) is a one-dimensional spectrum, F(g) is a gamma function, and a formula for calculating the gamma function is: T() : I: {ttßldt = (ß 1) !; where, ß is a random variable and t is a differential variable.

4. Method for non-destructive monitoring of calcium content in leaves of karst plants according to claim 1, characterized in that the spectral band without significant correlation the spectral band is with a p-value from the Pearson test correlation and significance greater than 0.

05.

5. Method for non-destructive monitoring of calcium content in leaves of karst plants according to claim 1, characterized in that a calculation formula of minimum absolute shrinkage and selection algorithm is: mn 2 n àla = arg min {î} [yi Z; XI. / al.] + 5; af} l= FF where, x is an eliminated spectral band, y is the calcium content data, y, is an observation, Xi]- is a jth characteristic of an observation, a is a regression coefficient is, 6 is a regularization parameter, m is the number of observations, and n is the number of features.

6. Method for non-destructive monitoring of calcium content in leaves of karst plants according to claim 1, characterised in that the initial network is a generalized neural regression network.

7. Method for non-destructive monitoring of calcium content in leaves of karst plants according to claim 1, characterised in that the parameters of the generalized neural regression network includes a distribution parameter.

8. Method for non-destructive monitoring of calcium content in leaves 5 of karst plants according to claim 1, characterized in that the optimization algorithm is a particle swarm algorithm.