Method and device for constructing wheat stripe rust monitoring model based on multi-source feature cooperation

Abstract

The invention provides a method and a device for constructing a wheat stripe rust monitoring model based on multi-source feature cooperation. The method comprises the following steps: acquiring a hyperspectral image and a thermal image of a target wheat field and disease index values of investigation sample points; constructing a vegetation index set and a texture feature set based on the hyperspectral image; extracting hyperspectral data of each survey sample point from the hyperspectral image and inputting the hyperspectral data into the non-measured character parameter inversion model and the empirical character parameter inversion model to obtain a non-measured character parameter value and an empirical character parameter value; extracting a canopy temperature value of each survey sample point from the thermal imaging; based on each non-measured character parameter value, each empirical character parameter value and each canopy temperature value, constructing a vegetation function character set; and constructing a wheat stripe rust monitoring model based on the vegetation index set, the texture feature set, the vegetation function character set and each disease index value. According to the scheme, modeling is carried out through the combined features of the three types of features, and the purpose of improving the precision of the stripe rust monitoring model is achieved in a multi-feature cooperation mode.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method and device for constructing a wheat stripe rust monitoring model with multi-source feature coordination. Background Art

[0002] The sensitive features widely used in vegetation disease research mainly include spectral features and phenotypic features. According to the difference in the processing method of the original spectrum, the spectral features can be subdivided into reflectance bands, vegetation indices (VIs) and derived features. The above features have been proven to be effective in disease stress monitoring. For example, the proportional vegetation index, anthocyanin reflectance index and triangular vegetation index were optimized based on the minimum redundancy maximum correlation algorithm, and a detection model for apple fire blight with an accuracy of 94.0% was constructed in combination with random forests. In a large number of studies, phenotypic features with texture, color and size as the main indicators have significantly improved the detection ability and recognition accuracy of crop diseases through collaborative modeling with traditional spectral features. For example, a study explored the effect of combining vegetation index, color index and texture feature (TFs) in monitoring the severity of cotton Verticillium wilt. The results showed that the fusion effect of the three data sources is better than any single or combination of two data sources.

[0003] Current wheat stripe rust monitoring is the same as traditional disease monitoring based on hyperspectral, both of which rely on vegetation index sets and texture feature sets for modeling. However, they transmit vegetation status information indirectly and are greatly affected by external environmental factors, thus affecting the accuracy of the wheat stripe rust monitoring model. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides a method and device for constructing a wheat stripe rust monitoring model with multi-source feature collaboration, so as to achieve the purpose of improving the accuracy of the stripe rust monitoring model.

[0005] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0006] A first aspect of an embodiment of the present invention discloses a method for constructing a wheat stripe rust monitoring model with multi-source feature collaboration, the method comprising:

[0007] Acquire a hyperspectral image, a thermal image and a disease index value corresponding to each survey sample point of a target wheat field; the survey sample point is preset in the target wheat field; the disease index value is used to characterize the severity of the disease at the survey sample point;

[0008] Based on the hyperspectral image, construct a vegetation index set and a texture feature set;

[0009] Extracting the hyperspectral data corresponding to each of the survey sample points from the hyperspectral image;

[0010] Inputting each of the hyperspectral data into a pre-constructed non-measured trait parameter inversion model to obtain a plurality of non-measured trait parameter values ​​through inversion, and inputting each of the hyperspectral data into a pre-constructed empirical trait parameter inversion model to obtain a plurality of empirical trait parameter values ​​through inversion;

[0011] Extracting the canopy temperature value corresponding to each of the survey sample points from the thermal image;

[0012] Based on each of the non-measured trait parameter values, each of the empirical trait parameter values ​​and each of the canopy temperature values, a vegetation functional trait set is constructed;

[0013] The vegetation index set, the texture feature set and the vegetation functional trait set are combined to obtain a combined feature set, and a wheat stripe rust monitoring model is constructed based on the combined feature set and each disease index value.

[0014] Preferably, constructing a vegetation index set and a texture feature set based on the hyperspectral image includes:

[0015] Based on the hyperspectral image and calculation formulas corresponding to the multiple vegetation indices, multiple vegetation index images are calculated;

[0016] Based on the plurality of vegetation index images, the vegetation index value of each vegetation index at each survey sample point is extracted, and each vegetation index is screened, and a vegetation index set is constructed based on the vegetation index values ​​corresponding to the screened vegetation indexes;

[0017] Performing dimensionality reduction processing on the hyperspectral image to obtain a plurality of principal component graphs, and respectively extracting texture feature images corresponding to a plurality of texture features from each of the principal component graphs using a gray level co-occurrence matrix;

[0018] Based on the plurality of texture feature images, the texture feature value of each texture feature at each survey sample point is extracted, and each texture feature is screened, and a texture feature set is constructed based on the texture feature values ​​corresponding to the screened texture features.

[0019] Preferably, the step of extracting the vegetation index value of each vegetation index at each survey sample point based on the plurality of vegetation index images, screening each vegetation index, and constructing a vegetation index set based on the vegetation index values ​​corresponding to the screened vegetation indexes comprises:

[0020] For each of the vegetation indices, using the corresponding vegetation index image and Python programming method, the vegetation index value corresponding to the vegetation index at each of the survey sample points is obtained;

[0021] For each of the vegetation indices, the corresponding vegetation index values ​​and the disease index values ​​corresponding to the survey sample points are regressed to obtain a Pearson correlation coefficient, and the vegetation indices whose Pearson correlation coefficients are less than a correlation coefficient threshold are eliminated;

[0022] For each of the remaining vegetation indices, a variance inflation factor analysis is performed based on the corresponding vegetation index values ​​and the disease index values ​​corresponding to the survey sample points to obtain a VIF value;

[0023] If there is a VIF value greater than or equal to the VIF threshold, the vegetation index corresponding to the largest VIF value is eliminated, and the step of performing variance inflation factor analysis on each of the remaining vegetation indices based on the corresponding vegetation index values ​​and the disease index values ​​corresponding to each survey sample point to obtain the VIF value is returned to the step, until each VIF value is less than the VIF threshold, and then a vegetation index set is constructed based on the vegetation index values ​​of the remaining vegetation indices at each survey sample point.

[0024] Preferably, the extracting of the texture feature value of each texture feature at each survey sample point based on the plurality of texture feature images, screening each texture feature, and constructing a texture feature set based on the texture feature values ​​corresponding to the screened texture features includes:

[0025] For each of the texture features, using the corresponding texture feature image and Python programming method, the texture feature value corresponding to the texture feature at each survey sample point is obtained;

[0026] For each of the texture features, the corresponding texture feature values ​​and the disease index values ​​corresponding to the survey sample points are regressed to obtain a Pearson correlation coefficient, and the texture features whose Pearson correlation coefficient is less than a correlation coefficient threshold are eliminated;

[0027] For each of the remaining texture features, a variance inflation factor analysis is performed based on the corresponding texture feature values ​​and the disease index values ​​corresponding to the survey sample points to obtain a VIF value;

[0028] If there is a VIF value greater than or equal to the VIF threshold, the texture feature corresponding to the largest VIF value is eliminated, and the process returns to the step of performing variance inflation factor analysis on each of the remaining texture features based on the corresponding texture feature values ​​and the disease index values ​​corresponding to each of the survey sample points to obtain the VIF value, until each of the VIF values ​​is less than the VIF threshold, and then a texture feature set is constructed based on the texture feature values ​​of the remaining texture features at each of the survey sample points.

[0029] Preferably, the process of constructing the non-measured property parameter inversion model includes:

[0030] Generate a plurality of simulated spectra and non-measured property simulation values ​​corresponding to each of the simulated spectra using the PROSAIL model;

[0031] Taking each of the simulated spectra and the corresponding simulated value of the non-measured property as sample data, constructing a sample data set, and dividing the sample data set into a training set and a test set;

[0032] Inputting the training set into a Gaussian process regression algorithm for training to obtain an inversion model of non-measured trait parameters to be evaluated;

[0033] The non-measured trait parameter inversion model to be evaluated is evaluated using the test set. If the evaluation passes, a trained non-measured trait parameter inversion model is obtained.

[0034] Preferably, the process of constructing the empirical property parameter inversion model includes:

[0035] Detecting and obtaining the actual measured value of the empirical trait corresponding to each of the survey sample points;

[0036] Generate a plurality of simulated spectra and empirical property simulation values ​​corresponding to each of the simulated spectra using the PROSAIL model;

[0037] Constructing a sample data set from each of the simulated spectra and the corresponding simulated value of the empirical property, and dividing the sample data set into a plurality of subsets of different sizes;

[0038] For each of the subsets, the subset is input into the empirical trait parameter inversion model to be trained that integrates active learning, so as to obtain the empirical trait parameter inversion model to be selected corresponding to the subset;

[0039] For each of the candidate empirical trait parameter inversion models, the hyperspectral data corresponding to each of the survey sample points is input to obtain multiple empirical trait estimation values, and performance evaluation is performed based on each of the empirical trait estimation values ​​and each of the empirical trait measured values ​​to determine the empirical trait parameter inversion model with the best performance.

[0040] Preferably, the wheat stripe rust monitoring model constructed based on the combined feature set and each of the disease index values ​​comprises:

[0041] The combined feature set and each of the disease index values ​​are input into the LASSO algorithm for training to obtain a wheat stripe rust monitoring model.

[0042] A second aspect of an embodiment of the present invention discloses a device for constructing a wheat stripe rust monitoring model with multi-source feature collaboration, the device comprising:

[0043] An acquisition unit, used for acquiring a hyperspectral image, a thermal image and a disease index value corresponding to each survey sample point of a target wheat field; the survey sample point is preset in the target wheat field; the disease index value is used for characterizing the severity of the disease at the survey sample point;

[0044] A first construction unit is used to construct a vegetation index set and a texture feature set based on the hyperspectral image;

[0045] A first extraction unit, used for extracting the hyperspectral data corresponding to each of the survey sample points from the hyperspectral image;

[0046] An inversion unit, used for inputting each of the hyperspectral data into a pre-constructed non-measured property parameter inversion model to obtain a plurality of non-measured property parameter values ​​through inversion, and inputting each of the hyperspectral data into a pre-constructed empirical property parameter inversion model to obtain a plurality of empirical property parameter values ​​through inversion;

[0047] A second extraction unit is used to extract the canopy temperature value corresponding to each of the survey sample points from the thermal image;

[0048] A second construction unit is used to construct a vegetation functional trait set based on each of the non-measured trait parameter values, each of the empirical trait parameter values ​​and each of the canopy temperature values;

[0049] The model building unit is used to combine the vegetation index set, the texture feature set and the vegetation functional trait set to obtain a combined feature set, and to build a wheat stripe rust monitoring model based on the combined feature set and each disease index value.

[0050] Preferably, the first building block comprises:

[0051] A first image generation subunit, configured to calculate and obtain a plurality of vegetation index images based on the hyperspectral image and calculation formulas corresponding to a plurality of vegetation indices;

[0052] The first set construction subunit is used to extract the vegetation index value of each vegetation index at each survey sample point based on the multiple vegetation index images, and screen each vegetation index, and construct a vegetation index set based on the vegetation index values ​​corresponding to the screened vegetation index;

[0053] The second image generation subunit is used to perform dimensionality reduction processing on the hyperspectral image to obtain a plurality of principal component graphs, and respectively extract texture feature images corresponding to a plurality of texture features from each of the principal component graphs using a gray level co-occurrence matrix;

[0054] The second set construction subunit is used to extract the texture feature value of each texture feature at each survey sample point based on the multiple texture feature images, screen each texture feature, and construct a texture feature set based on the texture feature values ​​corresponding to the screened texture features.

[0055] Preferably, the first set construction subunit is specifically used for:

[0056] For each of the vegetation indices, using the corresponding vegetation index image and Python programming method, the vegetation index value corresponding to the vegetation index at each of the survey sample points is obtained;

[0057] For each of the vegetation indices, the corresponding vegetation index values ​​and the disease index values ​​corresponding to the survey sample points are regressed to obtain a Pearson correlation coefficient, and the vegetation indices whose Pearson correlation coefficients are less than a correlation coefficient threshold are eliminated;

[0058] For each of the remaining vegetation indices, a variance inflation factor analysis is performed based on the corresponding vegetation index values ​​and the disease index values ​​corresponding to the survey sample points to obtain a VIF value;

[0059] If there is a VIF value greater than or equal to the VIF threshold, the vegetation index corresponding to the largest VIF value is eliminated, and the step of performing variance inflation factor analysis on each of the remaining vegetation indices based on the corresponding vegetation index values ​​and the disease index values ​​corresponding to each survey sample point to obtain the VIF value is returned to the step, until each VIF value is less than the VIF threshold, and then a vegetation index set is constructed based on the vegetation index values ​​of the remaining vegetation indices at each survey sample point.

[0060] Based on the method and device for constructing a wheat stripe rust monitoring model with multi-source feature coordination provided by the above-mentioned embodiment of the present invention, a hyperspectral image, thermal image and disease index value corresponding to each survey sample point of a target wheat field are obtained; the survey sample point is preset in the target wheat field; the disease index value is used to characterize the severity of the disease at the survey sample point; based on the hyperspectral image, a vegetation index set and a texture feature set are constructed; hyperspectral data corresponding to each of the survey sample points are extracted from the hyperspectral image; each of the hyperspectral data is input into a pre-constructed non-measured trait parameter inversion model, and the inversion is obtained. Multiple non-measured trait parameter values, each of the hyperspectral data is input into a pre-constructed empirical trait parameter inversion model to invert multiple empirical trait parameter values; the canopy temperature value corresponding to each of the survey sample points is extracted from the thermal imaging; based on each of the non-measured trait parameter values, each of the empirical trait parameter values ​​and each of the canopy temperature values, a vegetation functional trait set is constructed; the vegetation index set, the texture feature set and the vegetation functional trait set are combined to obtain a combined feature set, and a wheat stripe rust monitoring model is constructed based on the combined feature set and each of the disease index values. In this solution, a combined feature model of three types of features, namely, the vegetation functional trait set, the vegetation index set and the texture feature set, is used to achieve the purpose of improving the accuracy of the stripe rust monitoring model by using a multi-feature collaborative approach. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0062] Figure 1 A flowchart of a method for constructing a wheat stripe rust monitoring model with multi-source feature collaboration disclosed in an embodiment of the present invention;

[0063] Figure 2 This is a structural diagram of a device for constructing a wheat stripe rust monitoring model with multi-source feature collaboration disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0064] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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 creative work are within the scope of protection of the present invention.

[0065] In this application, the terms "comprises", "comprising" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or device comprising the element.

[0066] From the background technology, we can know that the current wheat stripe rust monitoring is the same as the traditional disease monitoring based on hyperspectral, and both rely on vegetation index sets (VIs) and texture feature sets (TFs) for modeling. However, they transmit vegetation status information indirectly and are greatly affected by external environmental factors, thus affecting the accuracy of the wheat stripe rust monitoring model.

[0067] Therefore, an embodiment of the present invention discloses a method and device for constructing a wheat stripe rust monitoring model with multi-source feature collaboration. In this scheme, a combined feature model of three types of features, namely a vegetation functional trait set, a vegetation index set and a texture feature set, is used to achieve the purpose of improving the accuracy of the stripe rust monitoring model by utilizing a multi-feature collaborative approach.

[0068] It should be noted that the key to the embodiment of the present invention is to establish a wheat stripe rust monitoring model with three types of features as input: plant functional traits (PTs), vegetation index set, and texture feature set. Therefore, the embodiment of the present invention will focus on how to extract these three types of features, the screening principles of each type of features, and how to use these three types of features to build a model.

[0069] like Figure 1 FIG. 1 is a flow chart of a method for constructing a wheat stripe rust monitoring model with multi-source feature coordination disclosed in an embodiment of the present invention. The method mainly includes the following steps:

[0070] Step S101: Obtain the hyperspectral image, thermal image and disease index value corresponding to each survey sample point of the target wheat field.

[0071] Among them, the target wheat field is the wheat field where wheat stripe rust is infected; the survey sample points are obtained by specifying multiple geographic coordinates in the target wheat field by researchers; the disease index value (DI) corresponding to each survey sample point is used to characterize the severity of the disease at the survey sample point. It is a number between 0 and 1. The larger the value, the more serious the disease level. The disease index value is obtained by agronomic experts through an assessment of the health status of wheat at these 90 survey sample points.

[0072] Preferably, the hyperspectral image and thermal image of the target wheat field can be collected by drones.

[0073] Step S102: constructing a vegetation index set and a texture feature set based on the hyperspectral image.

[0074] In one embodiment, vegetation indices related to crop pigment, water content, and stress quantification may be obtained as candidate features for constructing a vegetation index set.

[0075] In the specific implementation process of constructing the vegetation index set, multiple vegetation index images are calculated based on the calculation formulas corresponding to the hyperspectral image and multiple vegetation indices. Then, based on the multiple vegetation index images, the vegetation index value of each vegetation index at each survey sample point is extracted, and each vegetation index is screened. The vegetation index set is constructed based on the vegetation index values ​​corresponding to the screened vegetation indices.

[0076] In the embodiment of the present invention, 48 vegetation index calculation formulas related to crop pigment, water content and stress quantification obtained from the literature are used, as shown in the following table:

[0077]

[0078] It should be noted that R in the calculation formula represents reflectivity, and the subscript of R represents a specific wavelength, such as R 700 Refers to the reflectivity at a wavelength of 700 nanometers.

[0079] Among them, the vegetation index image is in TIFF (Tagged Image File Format) format.

[0080] The more specific process of vegetation index value extraction and screening includes the following steps:

[0081] Step S201: for each vegetation index, using the corresponding vegetation index image and Python programming method, obtain the vegetation index value corresponding to each survey sample point.

[0082] In step S201, the mean value of the vegetation index of all pixels within a range of 0.6m*0.6m around each survey sample point is extracted by Python programming to obtain the vegetation index value corresponding to the survey sample point.

[0083] Step S202: for each vegetation index, the corresponding vegetation index values ​​and the disease index values ​​corresponding to each survey sample point are regressed to obtain the Pearson correlation coefficient, and the vegetation index with a Pearson correlation coefficient less than the correlation coefficient threshold is eliminated.

[0084] In step S202, for each of the 48 vegetation indices, a regression calculation is performed between the vegetation index value at each survey sample point and the disease index value at each survey sample point to obtain the Pearson correlation coefficient (Correlation Coefficient, CC), and the correlation coefficient threshold is set to 0.4, that is, all vegetation indices with a correlation coefficient with the disease index value less than 0.4 are first eliminated.

[0085] The Pearson correlation coefficients of various vegetation indices and disease index values ​​are shown in the following table:

[0086]

[0087] Step S203: for each remaining vegetation index, variance inflation factor analysis is performed based on the corresponding vegetation index values ​​and the disease index values ​​corresponding to each survey sample point to obtain a VIF value.

[0088] In step S203, a variance inflation factor (VIF) analysis is performed on the remaining vegetation index and disease index values, and each vegetation index will obtain a corresponding VIF value.

[0089] Step S204: If there is a VIF value greater than or equal to the VIF threshold, the vegetation index corresponding to the largest VIF value is eliminated, and the process returns to step S203 until all VIF values ​​are less than the VIF threshold. Based on the vegetation index values ​​of the remaining vegetation indices at each survey sample point, a vegetation index set is constructed.

[0090] It should be noted that the larger the VIF value, the higher the multicollinearity between the current vegetation indices, which may have an adverse effect on modeling. Therefore, the features with the maximum VIF value are deleted one by one, and the variance inflation factor (VIF) analysis is repeated. This process is iterated until all the retained variables meet VIF < 10.

[0091] In the embodiment of the present invention, a total of 6 vegetation indices are finally retained, namely Macc01, TCARI, CRI550, PRI m1 , WI and HI_2014, the correlation coefficients of these vegetation indices with the disease index values ​​were -0.549, -0.466, -0.662, 0.492, -0.643 and -0.637, and the VIFs were 6.753, 7.389, 6.667, 2.899, 6.395 and 9.711, respectively. The variance inflation factor (VIF) analysis results of the 6 vegetation indices finally retained are shown in the following table:

[0092]

[0093] In the specific implementation process of constructing the texture feature set, the hyperspectral image is subjected to dimensionality reduction processing to obtain multiple principal component maps, and the gray-level co-occurrence matrix is ​​used to extract the texture feature images corresponding to the multiple texture features from each principal component map. Then, based on the multiple texture feature images, the texture feature values ​​of each texture feature at each survey sample point are extracted, and each texture feature is screened, and the texture feature set is constructed based on the texture feature values ​​corresponding to the screened texture features.

[0094] It should be noted that as the disease develops, winter wheat plants also show certain differences in canopy structure. Since texture features can capture grayscale changes between adjacent pixels, they help enhance the ability to interpret samples at different disease levels. The gray-level co-occurrence matrix (GLCM) can be applied to texture feature extraction because of its simplified calculation and ability to provide multi-scale features in different directions.

[0095] In the embodiment of the present invention, eight commonly used texture features are selected, including mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment and correlation.

[0096] It should be noted that a hyperspectral image usually has more than 100 bands, and the hyperspectral image in the embodiment of the present invention has 150 bands, with highly overlapping features and a large amount of redundant information. Therefore, the steps of texture feature image extraction are generally as follows:

[0097] Firstly, the built-in principal component analysis (PCA) calculation module of the software ENVI 5.6 was used to reduce the dimension of the hyperspectral data, and the first three principal component maps (PC1, PC2 and PC3, with a cumulative variance of 94.5%) were obtained.

[0098] Then, the GLCM tool in the software ENVI was used to extract the texture mean at multiple relative orientations (0°, 45°, 90°, and 135°) with a sliding window of size 17×17, and a total of 3*8=24 texture feature images in TIFF format were created for the first three principal component images.

[0099] The more specific process of texture feature value extraction and screening includes the following steps:

[0100] Step S301: For each texture feature, using the corresponding texture feature image and Python programming method, obtain the texture feature value corresponding to the texture feature at each survey sample point.

[0101] Step S302: for each texture feature, the corresponding texture feature values ​​and the disease index values ​​corresponding to each survey sample point are regressed to obtain the Pearson correlation coefficient, and texture features whose Pearson correlation coefficient is less than the correlation coefficient threshold are eliminated.

[0102] Step S303: for each remaining texture feature, variance inflation factor analysis is performed based on the corresponding texture feature values ​​and the disease index values ​​corresponding to each survey sample point to obtain a VIF value.

[0103] Step S304: If there is a VIF value greater than or equal to the VIF threshold, the texture feature corresponding to the largest VIF value is eliminated, and the process returns to step S303 until all VIF values ​​are less than the VIF threshold. Then, a texture feature set is constructed based on the texture feature values ​​of the remaining texture features at each survey sample point.

[0104] It should be noted that the texture feature value extraction and screening process is consistent with the extraction and screening of the above-mentioned vegetation index value, with the only difference being the extraction object and the screening object. The explanation parts can be referred to each other.

[0105] There are a total of 9 texture features finally retained in the embodiment of the present invention, namely Mea1, Sem1, Mea2, Var2, Con2, Sem2, Mea3, Var3 and Con3. Their correlation coefficients with the disease index values ​​are 0.615, 0.406, 0.713, -0.463, -0.560, 0.585, 0.425, -0.431 and -0.577, respectively, and their VIFs are 6.113, 3.715, 4.174, 4.699, 7.010, 7.883, 2.910, 4.870 and 4.751, respectively.

[0106] To simplify the description, texture features are expressed using feature abbreviations combined with principal component numbers. For example, Mea1 represents the Mean feature corresponding to the PC1 component.

[0107] The variance inflation factor (VIF) analysis results of the 9 texture features finally retained are shown in the following table:

[0108]

[0109] Step S103: extracting the hyperspectral data corresponding to each survey sample point from the hyperspectral image.

[0110] In one embodiment, when the hybrid inversion model is used later, in order to eliminate band collinearity and spectral information redundancy, after obtaining the hyperspectral data, it is also necessary to use the principal component analysis method to compress the 150 bands of the hyperspectral data into 5 principal components.

[0111] Step S104: input each hyperspectral data into a pre-constructed non-measured trait parameter inversion model to obtain multiple non-measured trait parameter values ​​through inversion, and input each hyperspectral data into a pre-constructed empirical trait parameter inversion model to obtain multiple empirical trait parameter values ​​through inversion.

[0112] It should be noted that steps S103 to S105 are a process of extracting vegetation functional trait parameters, and the vegetation functional trait parameters include: non-measured trait parameters, empirical trait parameters and canopy temperature.

[0113] Plant functional traits (PTs), such as canopy temperature, dry matter content, biochemical composition, vegetation structure, water content, etc., are closely related to the health of the vegetation itself and its stress response. As stress intensifies, parameters such as leaf pigment content and canopy structure will be affected. Quantifying these changes is crucial for developing effective and scalable stress detection methods. Compared with healthy vegetation, declining vegetation usually shows an increase in canopy temperature and dry matter content, as well as a decrease in some functional traits (such as pigment content, leaf area index, water and fluorescence).

[0114] Inversion models are the key to effectively quantifying PTs. Traditional methods rely on the empirical statistical relationship between PTs and spectra to construct simple linear, polynomial, exponential or logarithmic models. Although this method is simple and easy to use, the empirical relationship between the two is easily disturbed by external factors (such as sensors and vegetation types). Mechanistic models with solid theoretical foundations and physical significance, such as the Radiative Transfer Model (RTM), quantitatively explain the relationship between canopy structural parameters and reflectance, and are not affected by vegetation types. At present, PTs inversion based on RTM is mainly divided into two methods: (i) Lookup Table (LUT), which associates simulated spectra with measured parameters through a cost function, but this method has high storage and computational overhead, and the optimal cost function is difficult to determine; (ii) Hybrid Inversion Method (HIM), which effectively combines the simplicity of empirical statistical methods with the mechanical and generalization of physical models. In essence, it generates a database of simulated spectra and corresponding PTs based on RTM, and uses machine learning algorithms to determine the relationship between the two.

[0115] The hybrid inversion model adopted in the embodiment of the present invention is essentially based on forward RTM (the RTM used in the embodiment of the present invention is the PROSAIL model) to generate simulated spectra (150 bands of hyperspectral data) and simulated values ​​of vegetation functional traits, and uses a machine learning regression algorithm (Machine Learning Regression Algorithm, MLRA) to establish a mapping relationship between the simulated spectra and the simulated values ​​of vegetation functional traits. Finally, a reliable inversion model (established by the machine learning regression algorithm MLRA, the algorithm used in the embodiment of the present invention is a Gaussian process regression algorithm) is applied to the hyperspectral data of each survey sample point extracted from the hyperspectral image to obtain the vegetation functional trait parameters.

[0116] Among them, Gaussian Process Regression (GPR) is a non-parametric probability model based on Bayesian theory. It can capture complex nonlinear relationships through kernel functions and has excellent modeling capabilities under small sample conditions.

[0117] Preferably, the vegetation functional traits (PTs) to be extracted include leaf carotenoid content (Car), leaf anthocyanin content (Anth), carbon-based components (CBC), leaf area index (LAI), canopy chlorophyll content (CCC) and canopy temperature (T). C ). Among them, there are three non-measured trait parameters, namely carotenoids, anthocyanins and carbon-based components, while leaf area index and canopy chlorophyll content are empirical trait parameters.

[0118] Based on the above explanation of the hybrid inversion model, the construction process of the non-measured property parameter inversion model includes the following steps:

[0119] Step S401: using the PROSAIL model to generate multiple simulated spectra and non-measured property simulation values ​​corresponding to each simulated spectrum.

[0120] In step S401 , the PROSAIL model is composed of the leaf-level PROSPECT-PRO radiation transfer model and the canopy-level 4SAIL model.

[0121] Among them, the PROSPECT-PRO model includes 8 leaf parameters: structure, chlorophyll, carotenoids, anthocyanins, brown pigments, equivalent water thickness, protein content and carbon-based composition, and provides leaf directional-hemispherical reflectance and transmittance in the range of 400-2500nm (spectral resolution of 1nm). The 4SAIL model is used to simulate the optical properties of vegetation canopies, and its parameters include LAI, leaf angle distribution, diffuse / direct radiation ratio, hot spot parameters, sun-target-sensor geometry, etc.

[0122] The input parameters of the PROSAIL model are set according to the measured values ​​and relevant literature references. The specific input parameter settings are shown in the following table:

[0123]

[0124] Preferably, the PROSAIL model with the above input parameter settings randomly creates 10,000 simulated spectra and their corresponding non-measured property simulation values ​​based on the probability density function.

[0125] Step S402: taking each simulated spectrum and the corresponding simulated value of the non-measured property as sample data, constructing a sample data set, and dividing the sample data set into a training set and a test set.

[0126] Preferably, the 10,000 simulated spectra and their corresponding simulated values ​​of non-measured traits are randomly divided into a training set (80%) and a test set (20%) to construct and evaluate the performance of the non-measured trait parameter inversion model.

[0127] In one embodiment, since the PROSAIL model directly generates a spectrum in the range of 400-2500nm (spectral resolution of 1nm) after the input parameters are set, before step S402, the 10,000 simulated spectra need to be resampled according to the spectral bandwidth and spectral response function of the hyperspectral sensor on the drone (the hyperspectral sensor used in the present invention includes 150 bands, the spectral resolution is 5nm, and the wavelength range is 400-1000nm) to match the measured hyperspectral data of each survey sample point.

[0128] In one embodiment, when using a hybrid inversion model, in order to eliminate band collinearity and spectral information redundancy, it is also necessary to use a principal component analysis method to compress 150 bands into 5 principal components. This operation must be performed after obtaining the hyperspectral data in the simulated spectrum and the drone's measured hyperspectral image.

[0129] Step S403: input the training set into the Gaussian process regression algorithm for training to obtain an inversion model of the non-measured trait parameters to be evaluated.

[0130] Step S404: using the test set to evaluate the non-measured trait parameter inversion model to be evaluated, if the evaluation passes, a trained non-measured trait parameter inversion model is obtained.

[0131] It can be understood that after the inversion model of non-measured trait parameters was constructed, the model was used to estimate the values ​​of three types of non-measured trait parameters, namely carotenoids, anthocyanins, and carbon-based components, from the hyperspectral data of 90 survey sample points.

[0132] Based on the above explanation of the hybrid inversion model, the construction process of the empirical property parameter inversion model includes the following steps:

[0133] Step S501: Detect and obtain the actual measured value of the empirical trait corresponding to each survey sample point.

[0134] It should be noted that the empirical trait parameters are the Leaf Area Index (LAI) corresponding to each survey sample point measured by the equipment. mea ) and Leaf Chlorophyll Content (LCC) mea ), and the canopy chlorophyll content (Canopy Chlorophyll Content, CCC) based on Equation 1 mea ).

[0135] (1)

[0136] Formula 1 realizes the conversion of chlorophyll content from leaf level to canopy level, CCC mea Used to simulate the aboveground chlorophyll content in g / m 2 .

[0137] Step S502: Generate multiple simulated spectra and empirical property simulation values ​​corresponding to each simulated spectrum using the PROSAIL model.

[0138] It should be noted that the measured values ​​of the above three types of empirical traits are all verifiable empirical trait parameters. By interactively using the following data, a hybrid inversion model integrating active learning can be established: simulated spectra generated by the PROSAIL model (1000, 1500, 2000, 2500 and 3000 samples), and the corresponding simulated values ​​of the empirical traits, and the measured values ​​of the empirical traits of each survey sample point.

[0139] Step S503: construct a sample data set from each simulated spectrum and the corresponding empirical property simulation value, and divide the sample data set into a plurality of subsets of different sizes.

[0140] In step S503, in order to reduce the influence of sample representativeness on the accuracy of the inversion model, 1000, 1500, 2000, 2500 and 3000 samples are randomly selected from the above 10,000 simulated spectra and their corresponding empirical trait simulation values ​​to construct 5 subsets of different sizes.

[0141] Similarly, it is necessary to use the principal component analysis method to generate five principal components to achieve dimensionality reduction in the spectral domain. Please refer to the above embodiment for details.

[0142] Step S504: for each subset, input the subset into the empirical trait parameter inversion model to be trained that integrates active learning, and obtain the empirical trait parameter inversion model to be selected corresponding to the subset.

[0143] In step S504, the heuristic iterative characteristics of active learning (AL) are used to selectively identify and annotate the samples with the most information to improve the model efficiency. The specific process is: 20 simulated spectra are randomly selected from subsets of different sizes (1000, 1500, 2000, 2500 and 3000 samples) as the initial pool, and 1 sample is added one by one according to the selected active learning standard. If the new sample improves the performance of the model (the empirical trait parameter inversion model that integrates active learning (AL)), it will be retained, otherwise it will be eliminated, and the above process will be iterated until all remaining simulated spectra are traversed.

[0144] Here, the squared Euclidean distance is used as the active learning criterion, and simulated spectra that are far away from existing samples in the training set are selected from the pool according to Formula 2.

[0145] (2)

[0146] Where X u is a sample in the candidate set; X l are samples in the training set. All distances between samples are calculated and the farthest distance is selected.

[0147] Step S505: For each candidate empirical trait parameter inversion model, the hyperspectral data corresponding to each survey sample point is input to obtain multiple empirical trait estimation values, and performance evaluation is performed based on each empirical trait estimation value and each empirical trait measured value to determine the empirical trait parameter inversion model with the best performance.

[0148] In step S505, the candidate empirical trait parameter inversion models constructed by the five lookup table sizes (LUTs), i.e., subsets, are tested successively. The coefficient R between the empirical trait measured value and the empirical trait estimated value output by the current test model is determined. 2 The root mean square error RMSE is shown in the following table:

[0149]

[0150] Specifically, the optimal inversion models corresponding to LCC, LAI, and CCC are not directly determined by the number of samples in the lookup table, but by their representativeness for the measured data. For example, when the test set size is 2000, 1500, and 2500, the inversion models corresponding to LCC, LAI, and CCC are optimal, respectively. 2 are 0.350, 0.577 and 0.596, and the RMSE are 3.943, 0.486 and 0.160.

[0151] It is understandable that the respective optimal empirical property parameter inversion models are selected to obtain LCC, LAI and CCC.

[0152] It should be noted that considering that the canopy chlorophyll content is more helpful in capturing the comprehensive impact of stripe rust at the canopy scale, LCC was abandoned when constructing the vegetation functional trait set.

[0153] Step S105: extracting the canopy temperature value corresponding to each survey sample point from the thermal image.

[0154] In step S105, the canopy temperature mean of all pixels within a range of 0.6m*0.6m around each survey sample point is extracted by Python programming as the canopy temperature value of the survey sample point.

[0155] Step S106: constructing a vegetation functional trait set based on each non-measured trait parameter value, each empirical trait parameter value and each canopy temperature value.

[0156] Step S107: combining the vegetation index set, the texture feature set and the vegetation functional trait set to obtain a combined feature set, and constructing a wheat stripe rust monitoring model based on the combined feature set and each disease index value.

[0157] In the specific implementation process of step S107, the combined feature set and each disease index value are input into the LASSO algorithm for training to obtain a wheat stripe rust monitoring model.

[0158] The LASSO algorithm uses L1 regularization in the regression process, and uses the penalty term to make the coefficients of local features tend to zero, which means that irrelevant or redundant features can be effectively removed, reducing overfitting and improving generalization ability.

[0159] It should be noted that the single features (vegetation index set VIs, texture feature set TFs or vegetation functional trait set PTs) and their combined features (VIs+TFs, VIs+PTs, TFs+PTs and VIs+TFs+PTs) were respectively input into the LASSO algorithm to construct 7 types of wheat stripe rust monitoring models. The test results showed that the optimal model was obtained when the three-type feature combination was used for LASSO regression.

[0160] The evaluation indexes of these seven wheat stripe rust monitoring models were calculated using R 2 , RMSE and mean absolute error (MAE), and a 6-fold cross-validation test was used for performance evaluation. The R 2 , RMSE, and MAE are 0.628, 8.03%, and 6.57% respectively. Compared with the data sources of PTs, TFs, VIs, PTs+TFs, PTs+VIs, and TFs+VIs, the model accuracy is improved by 5.90%, 18.71%, 16.08%, 2.11%, 4.49%, and 8.09% respectively. The details are shown in the following table:

[0161]

[0162] Based on the method for constructing a wheat stripe rust monitoring model with multi-source feature collaboration disclosed in the above-mentioned embodiment of the present invention, the functional traits of vegetation that closely reflect disease stress are comprehensively acquired based on hyperspectral images and thermal images, and combined feature modeling is performed in combination with the vegetation index set and the texture feature set to obtain a wheat stripe rust monitoring model. The combined feature modeling results show that multi-feature collaboration helps to improve the accuracy of the wheat stripe rust monitoring model. In addition, the LASSO regression algorithm removes irrelevant or redundant features through L1 regularization, effectively giving play to the advantages of multi-feature collaboration.

[0163] Corresponding to the method for constructing a wheat stripe rust monitoring model with multi-source feature coordination disclosed in the above embodiment of the present invention, as Figure 2 As shown, it is a structural diagram of a construction device of a multi-source feature coordinated wheat stripe rust monitoring model disclosed in an embodiment of the present invention, including: an acquisition unit 201, a first construction unit 202, a first extraction unit 203, an inversion unit 204, a second extraction unit 205, a second construction unit 206 and a model construction unit 207.

[0164] The acquisition unit 201 is used to acquire the hyperspectral image, thermal image and disease index value corresponding to each survey sample point of the target wheat field; the survey sample point is preset in the target wheat field; the disease index value is used to characterize the severity of the disease at the survey sample point;

[0165] The first construction unit 202 is used to construct a vegetation index set and a texture feature set based on the hyperspectral image.

[0166] In one embodiment, the first building unit 202 includes:

[0167] The first image generation subunit is used to calculate and obtain multiple vegetation index images based on the hyperspectral image and the calculation formulas corresponding to the multiple vegetation indices;

[0168] The first set construction subunit is used to extract the vegetation index value of each vegetation index at each survey sample point based on multiple vegetation index images, and screen each vegetation index, and construct a vegetation index set based on the vegetation index values ​​corresponding to the screened vegetation index;

[0169] The second image generation subunit is used to perform dimensionality reduction processing on the hyperspectral image to obtain multiple principal component maps, and respectively extract texture feature images corresponding to multiple texture features from each principal component map using a gray level co-occurrence matrix;

[0170] The second set construction subunit is used to extract the texture feature value of each texture feature at each survey sample point based on multiple texture feature images, screen each texture feature, and construct a texture feature set based on the texture feature values ​​corresponding to the screened texture features.

[0171] In one embodiment, the first set constructs a subunit, specifically for:

[0172] For each vegetation index, the corresponding vegetation index image and Python programming method are used to obtain the vegetation index value corresponding to each survey sample point;

[0173] For each vegetation index, the corresponding vegetation index values ​​and the disease index values ​​corresponding to each survey sample point are regressed to obtain the Pearson correlation coefficient, and the vegetation index with a Pearson correlation coefficient less than the correlation coefficient threshold is eliminated;

[0174] For each remaining vegetation index, variance inflation factor analysis is performed based on the corresponding vegetation index values ​​and the disease index values ​​corresponding to each survey sample point to obtain the VIF value;

[0175] If there is a VIF value greater than or equal to the VIF threshold, the vegetation index corresponding to the largest VIF value will be eliminated, and the step of performing variance inflation factor analysis on each remaining vegetation index based on the corresponding vegetation index values ​​and the disease index values ​​corresponding to each survey sample point to obtain the VIF value will be returned until each VIF value is less than the VIF threshold. Based on the vegetation index values ​​of the remaining vegetation indices at each survey sample point, a vegetation index set is constructed.

[0176] In one embodiment, the second set constructs a subunit, specifically for:

[0177] For each texture feature, the corresponding texture feature image and Python programming method are used to obtain the texture feature value corresponding to each survey sample point;

[0178] For each texture feature, the corresponding texture feature values ​​and the disease index values ​​corresponding to each survey sample point are regressed to obtain the Pearson correlation coefficient, and the texture features whose Pearson correlation coefficient is less than the correlation coefficient threshold are eliminated;

[0179] For each remaining texture feature, variance inflation factor analysis is performed based on the corresponding texture feature values ​​and the disease index values ​​corresponding to each survey sample point to obtain the VIF value;

[0180] If there is a VIF value greater than or equal to the VIF threshold, the texture feature corresponding to the largest VIF value will be eliminated, and the process will return to execute the variance inflation factor analysis for each remaining texture feature based on the corresponding texture feature values ​​and the disease index values ​​corresponding to each survey sample point to obtain the VIF value. This step will continue until each VIF value is less than the VIF threshold. Then, based on the texture feature values ​​of the remaining texture features at each survey sample point, a texture feature set will be constructed.

[0181] The first extraction unit 203 is used to extract the hyperspectral data corresponding to each survey sample point from the hyperspectral image.

[0182] The inversion unit 204 is used to input each hyperspectral data into a pre-built non-measured property parameter inversion model to invert a plurality of non-measured property parameter values, and input each hyperspectral data into a pre-built empirical property parameter inversion model to invert a plurality of empirical property parameter values.

[0183] In one embodiment, the device further comprises:

[0184] A first inversion model building unit is used to generate a plurality of simulated spectra and non-measured property simulation values ​​corresponding to each simulated spectrum using the PROSAIL model;

[0185] Each simulated spectrum and the corresponding simulated value of the non-measured trait are used as sample data to construct a sample data set, and the sample data set is divided into a training set and a test set;

[0186] The training set is input into the Gaussian process regression algorithm for training to obtain the inversion model of the non-measured trait parameters to be evaluated;

[0187] The test set is used to evaluate the inversion model of the non-measured trait parameters to be evaluated. If the evaluation passes, the trained inversion model of the non-measured trait parameters is obtained.

[0188] In one embodiment, the device further comprises:

[0189] The second inversion model building unit is used to detect and obtain the measured value of the empirical trait corresponding to each survey sample point;

[0190] The PROSAIL model is used to generate multiple simulated spectra and the empirical trait simulation values ​​corresponding to each simulated spectrum;

[0191] A sample data set is constructed by combining each simulated spectrum and the corresponding empirical trait simulation value, and the sample data set is divided into a plurality of subsets of different sizes;

[0192] For each subset, the subset is input into the empirical trait parameter inversion model to be trained that integrates active learning, so as to obtain the empirical trait parameter inversion model to be selected corresponding to the subset;

[0193] For each candidate empirical trait parameter inversion model, the hyperspectral data corresponding to each survey sample point is input to obtain multiple empirical trait estimation values. Performance evaluation is performed based on each empirical trait estimation value and each empirical trait measured value to determine the empirical trait parameter inversion model with the best performance.

[0194] The second extraction unit 205 is used to extract the canopy temperature value corresponding to each survey sample point from the thermal image.

[0195] The second constructing unit 206 is used to construct a vegetation functional trait set based on each non-measured trait parameter value, each empirical trait parameter value and each canopy temperature value.

[0196] The model building unit 207 is used to combine the vegetation index set, the texture feature set and the vegetation functional trait set to obtain a combined feature set, and to build a wheat stripe rust monitoring model based on the combined feature set and each disease index value.

[0197] In one embodiment, the model building unit 207 for building a wheat stripe rust monitoring model based on the combined feature set and each disease index value is specifically used to:

[0198] The combined feature set and each disease index value were input into the LASSO algorithm for training to obtain the wheat stripe rust monitoring model.

[0199] Based on the above-mentioned device for constructing a wheat stripe rust monitoring model with multi-source feature collaboration disclosed in the embodiment of the present invention, the functional traits of vegetation that closely reflect disease stress are comprehensively acquired based on hyperspectral images and thermal images, and combined feature modeling is performed in combination with the vegetation index set and the texture feature set to obtain a wheat stripe rust monitoring model. The combined feature modeling results show that multi-feature collaboration helps to improve the accuracy of the wheat stripe rust monitoring model. In addition, the LASSO regression algorithm removes irrelevant or redundant features through L1 regularization, effectively giving play to the advantages of multi-feature collaboration.

[0200] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can refer to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system or system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiment. The system and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without creative work.

[0201] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0202] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing a wheat stripe rust monitoring model with multi-source characteristics, characterized in that: The method comprises: Acquire a hyperspectral image, a thermal image and a disease index value corresponding to each survey sample point of a target wheat field; the survey sample point is preset in the target wheat field; the disease index value is used to characterize the severity of the disease at the survey sample point; Based on the hyperspectral image, construct a vegetation index set and a texture feature set; Extracting the hyperspectral data corresponding to each of the survey sample points from the hyperspectral image; Inputting each of the hyperspectral data into a pre-constructed non-measured trait parameter inversion model to obtain a plurality of non-measured trait parameter values ​​through inversion, and inputting each of the hyperspectral data into a pre-constructed empirical trait parameter inversion model to obtain a plurality of empirical trait parameter values ​​through inversion; Extracting the canopy temperature value corresponding to each of the survey sample points from the thermal image; Based on each of the non-measured trait parameter values, each of the empirical trait parameter values ​​and each of the canopy temperature values, a vegetation functional trait set is constructed; The vegetation index set, the texture feature set and the vegetation functional trait set are combined to obtain a combined feature set, and a wheat stripe rust monitoring model is constructed based on the combined feature set and each disease index value.

2. The method according to claim 1, characterized in that The step of constructing a vegetation index set and a texture feature set based on the hyperspectral image comprises: Based on the hyperspectral image and calculation formulas corresponding to the multiple vegetation indices, multiple vegetation index images are calculated; Based on the plurality of vegetation index images, the vegetation index value of each vegetation index at each survey sample point is extracted, and each vegetation index is screened, and a vegetation index set is constructed based on the vegetation index values ​​corresponding to the screened vegetation indexes; Performing dimensionality reduction processing on the hyperspectral image to obtain a plurality of principal component graphs, and respectively extracting texture feature images corresponding to a plurality of texture features from each of the principal component graphs using a gray level co-occurrence matrix; Based on the plurality of texture feature images, the texture feature value of each texture feature at each survey sample point is extracted, and each texture feature is screened, and a texture feature set is constructed based on the texture feature values ​​corresponding to the screened texture features.

3. The method according to claim 2, characterized in that The step of extracting the vegetation index value of each vegetation index at each survey sample point based on the plurality of vegetation index images, screening each vegetation index, and constructing a vegetation index set based on the vegetation index values ​​corresponding to the screened vegetation indexes includes: For each of the vegetation indices, using the corresponding vegetation index image and Python programming method, the vegetation index value corresponding to the vegetation index at each of the survey sample points is obtained; For each of the vegetation indices, the corresponding vegetation index values ​​and the disease index values ​​corresponding to the survey sample points are regressed to obtain a Pearson correlation coefficient, and the vegetation indices whose Pearson correlation coefficients are less than a correlation coefficient threshold are eliminated; For each of the remaining vegetation indices, a variance inflation factor analysis is performed based on the corresponding vegetation index values ​​and the disease index values ​​corresponding to the survey sample points to obtain a VIF value; If there is a VIF value greater than or equal to the VIF threshold, the vegetation index corresponding to the largest VIF value is eliminated, and the step of performing variance inflation factor analysis on each of the remaining vegetation indices based on the corresponding vegetation index values ​​and the disease index values ​​corresponding to each survey sample point to obtain the VIF value is returned to the step, until each VIF value is less than the VIF threshold, and then a vegetation index set is constructed based on the vegetation index values ​​of the remaining vegetation indices at each survey sample point.

4. The method according to claim 2, characterized in that: The method of extracting a texture feature value of each texture feature at each survey sample point based on the plurality of texture feature images, screening each texture feature, and constructing a texture feature set based on the texture feature values ​​corresponding to the screened texture features includes: For each of the texture features, using the corresponding texture feature image and Python programming method, the texture feature value corresponding to the texture feature at each survey sample point is obtained; For each of the texture features, the corresponding texture feature values ​​and the disease index values ​​corresponding to the survey sample points are regressed to obtain a Pearson correlation coefficient, and the texture features whose Pearson correlation coefficient is less than a correlation coefficient threshold are eliminated; For each of the remaining texture features, a variance inflation factor analysis is performed based on the corresponding texture feature values ​​and the disease index values ​​corresponding to the survey sample points to obtain a VIF value; If there is a VIF value greater than or equal to the VIF threshold, the texture feature corresponding to the largest VIF value is eliminated, and the process returns to the step of performing variance inflation factor analysis on each of the remaining texture features based on the corresponding texture feature values ​​and the disease index values ​​corresponding to each of the survey sample points to obtain the VIF value, until each of the VIF values ​​is less than the VIF threshold, and then a texture feature set is constructed based on the texture feature values ​​of the remaining texture features at each of the survey sample points.

5. The method according to claim 1, characterized in that The construction process of the non-measured property parameter inversion model includes: Generate a plurality of simulated spectra and non-measured property simulation values ​​corresponding to each of the simulated spectra using the PROSAIL model; Taking each of the simulated spectra and the corresponding simulated value of the non-measured property as sample data, constructing a sample data set, and dividing the sample data set into a training set and a test set; Inputting the training set into a Gaussian process regression algorithm for training to obtain an inversion model of non-measured trait parameters to be evaluated; The non-measured trait parameter inversion model to be evaluated is evaluated using the test set. If the evaluation passes, a trained non-measured trait parameter inversion model is obtained.

6. The method according to claim 1, characterized in that The construction process of the empirical property parameter inversion model includes: Detecting and obtaining the actual measured value of the empirical trait corresponding to each of the survey sample points; Generate a plurality of simulated spectra and empirical property simulation values ​​corresponding to each of the simulated spectra using the PROSAIL model; Constructing a sample data set from each of the simulated spectra and the corresponding simulated value of the empirical property, and dividing the sample data set into a plurality of subsets of different sizes; For each of the subsets, the subset is input into the empirical trait parameter inversion model to be trained that integrates active learning, so as to obtain the empirical trait parameter inversion model to be selected corresponding to the subset; For each of the candidate empirical trait parameter inversion models, the hyperspectral data corresponding to each of the survey sample points is input to obtain multiple empirical trait estimation values, and performance evaluation is performed based on each of the empirical trait estimation values ​​and each of the empirical trait measured values ​​to determine the empirical trait parameter inversion model with the best performance.

7. The method according to any one of claims 1 to 6, characterized in that: The wheat stripe rust monitoring model constructed based on the combined feature set and each of the disease index values ​​includes: The combined feature set and each of the disease index values ​​are input into the LASSO algorithm for training to obtain a wheat stripe rust monitoring model.

8. A device for constructing a wheat stripe rust monitoring model with multi-source feature coordination, characterized in that: The device comprises: An acquisition unit, used for acquiring a hyperspectral image, a thermal image and a disease index value corresponding to each survey sample point of a target wheat field; the survey sample point is preset in the target wheat field; the disease index value is used for characterizing the severity of the disease at the survey sample point; A first construction unit is used to construct a vegetation index set and a texture feature set based on the hyperspectral image; A first extraction unit, used for extracting the hyperspectral data corresponding to each of the survey sample points from the hyperspectral image; An inversion unit, used for inputting each of the hyperspectral data into a pre-constructed non-measured property parameter inversion model to obtain a plurality of non-measured property parameter values ​​through inversion, and inputting each of the hyperspectral data into a pre-constructed empirical property parameter inversion model to obtain a plurality of empirical property parameter values ​​through inversion; A second extraction unit is used to extract the canopy temperature value corresponding to each of the survey sample points from the thermal image; A second construction unit is used to construct a vegetation functional trait set based on each of the non-measured trait parameter values, each of the empirical trait parameter values ​​and each of the canopy temperature values; The model building unit is used to combine the vegetation index set, the texture feature set and the vegetation functional trait set to obtain a combined feature set, and to build a wheat stripe rust monitoring model based on the combined feature set and each disease index value.

9. The device according to claim 8, characterized in that The first building block comprises: A first image generation subunit, configured to calculate and obtain a plurality of vegetation index images based on the hyperspectral image and calculation formulas corresponding to a plurality of vegetation indices; The first set construction subunit is used to extract the vegetation index value of each vegetation index at each survey sample point based on the multiple vegetation index images, and screen each vegetation index, and construct a vegetation index set based on the vegetation index values ​​corresponding to the screened vegetation index; The second image generation subunit is used to perform dimensionality reduction processing on the hyperspectral image to obtain a plurality of principal component graphs, and respectively extract texture feature images corresponding to a plurality of texture features from each of the principal component graphs using a gray level co-occurrence matrix; The second set construction subunit is used to extract the texture feature value of each texture feature at each survey sample point based on the multiple texture feature images, screen each texture feature, and construct a texture feature set based on the texture feature values ​​corresponding to the screened texture features.

10. The device according to claim 9, characterized in that The first set construction subunit is specifically used for: For each of the vegetation indices, using the corresponding vegetation index image and Python programming method, the vegetation index value corresponding to the vegetation index at each of the survey sample points is obtained; For each of the vegetation indices, the corresponding vegetation index values ​​and the disease index values ​​corresponding to the survey sample points are regressed to obtain a Pearson correlation coefficient, and the vegetation indices whose Pearson correlation coefficients are less than a correlation coefficient threshold are eliminated; For each of the remaining vegetation indices, a variance inflation factor analysis is performed based on the corresponding vegetation index values ​​and the disease index values ​​corresponding to the survey sample points to obtain a VIF value; If there is a VIF value greater than or equal to the VIF threshold, the vegetation index corresponding to the largest VIF value is eliminated, and the step of performing variance inflation factor analysis on each of the remaining vegetation indices based on the corresponding vegetation index values ​​and the disease index values ​​corresponding to each survey sample point to obtain the VIF value is returned to the step, until each VIF value is less than the VIF threshold, and then a vegetation index set is constructed based on the vegetation index values ​​of the remaining vegetation indices at each survey sample point.

Images