Regional soil moisture inversion method and system for inverting canopy parameters

Abstract

The invention belongs to the technical field of agricultural remote sensing, and provides a regional soil moisture inversion method for inverting canopy parameters, which comprises the following steps: S1, collecting actual measurement data, and dividing the actual measurement data into a training set and a test set; s2, carrying out polarization decomposition on the SAR remote sensing image to obtain polarization decomposition characteristics with mechanical property; s3, screening a polarization decomposition characteristic combination with strong correlation with the leaf area index LAI and the plant water content VWC; s4, training a canopy parameter inversion model based on the screened features, and performing inversion to obtain a regional leaf area index LAI and a regional plant water content VWC; s5, water cloud model parameter calibration; s6, using the calibrated water cloud model, coupling the regional leaf area index LAI and the regional plant water content VWC to invert the soil moisture. According to the method, the limitation of a single-point measurement and optical inversion method on the acquisition of regional crop canopy parameters is overcome, and the water cloud model input parameters are ensured to have relatively high space consistency and accuracy.

Description

Technical Field

[0001] The invention belongs to the technical field of agricultural remote sensing, and in particular relates to a method and system for inverting regional soil moisture by inverting canopy parameters. Background Art

[0002] As an important component of the global ecosystem, soil moisture is a bridge connecting the land surface moisture and the atmospheric space and groundwater storage, and has a profound impact on the stability, functionality and diversity of the ecosystem. In the agricultural field, soil moisture in crop-covered areas is an important hydrological state variable that directly affects the growth and development of crops and the formation of yields. Therefore, accurate monitoring of soil moisture in crop-covered areas is of great significance for agricultural monitoring, resource utilization, disaster prevention and control, and crop yield improvement.

[0003] Traditional soil moisture monitoring mainly relies on instruments or manual fixed-point monitoring. The monitoring data is highly accurate, but the acquisition cost is also high, which cannot meet the needs of long-term and large-area soil moisture monitoring. Although the temporal and spatial distribution information of soil moisture can be obtained through model simulation, this method has strict requirements on the model input parameters and is not very operable. With the development of earth observation technology, remote sensing technology has gradually been applied to soil moisture monitoring. Optical remote sensing has the advantages of convenience and efficiency, high spatial resolution, and rich band information. It can effectively reflect the growth status of surface crops, but it is easily affected by weather and cannot achieve high-quality monitoring in rainy weather. Synthetic Aperture Radar (SAR) belongs to active microwave remote sensing, which has different imaging principles and penetration capabilities from optics, and can make up for the shortcomings of optical remote sensing observations. When the soil moisture content changes, the dielectric properties of the soil will change accordingly, which in turn affects the SAR backscattering coefficient. Therefore, SAR remote sensing has significant advantages in soil moisture inversion.

[0004] However, for areas covered by crops, the SAR backscatter coefficient is not only related to its own polarization mode, incident angle and soil moisture, but also affected by factors such as crop moisture content. At this time, the SAR backscatter includes the direct scattering contribution of the crop canopy, the ground scattering contribution that is attenuated twice by the crop canopy, and the scattering contribution of the complex interaction between crops and the ground surface. How to reduce the impact of crops is the key and difficulty of soil moisture inversion in crop-covered areas. In addition, soil moisture changes rapidly over time, and it is difficult to directly build a relationship with SAR remote sensing features for inversion. In comparison, crop canopy parameters change slowly over time, and indirect inversion of soil moisture by coupling regional crop canopy parameters helps to improve the robustness and robustness of the inversion model.

[0005] The water cloud model is one of the main models for inverting soil moisture based on crop canopy parameters. The water cloud model assumes the crop canopy as "water-cloud" and only focuses on the direct scattering of the crop layer and the soil layer, without considering the multiple scattering between the crop and the soil. Crop canopy parameters such as Leaf Area Index (LAI) and Vegetation Water Content (VWC) are important input parameters. At present, the acquisition of canopy parameters such as Leaf Area Index (LAI) and Vegetation Water Content (VWC) is mainly based on field single-point measurement and spatial interpolation or optical image inversion. Affected by crop varieties, soil properties and field management measures, crop canopy parameters have significant spatial variability, and the spatial interpolation method cannot accurately obtain the true value of crop canopy parameters for each pixel. In areas covered by clouds and rain, the quality of optical images is poor, and the regional crop canopy parameters inverted based on optical data have large errors.

[0006] In addition, the water cloud model's description of crop structure and stratification is too simple and its mechanism is weak. It is urgent to introduce new SAR remote sensing mechanism characteristics to improve the accuracy of soil moisture inversion. Summary of the invention

[0007] In view of the problems in the background technology, the present invention proposes a regional soil moisture inversion method for inverting canopy parameters.

[0008] Radar polarization characteristic values ​​can provide information about the scattering characteristics of ground objects, which are related to soil moisture. The present invention first performs polarization decomposition of SAR remote sensing data based on polarization decomposition technology and selects polarization characteristics that have mechanistic and physical significance and can better reflect the plant characteristics of crops. Then, the above-mentioned preferred polarization characteristics are input into the random forest algorithm to invert regional canopy parameters, and the crop canopy parameters obtained by inversion are substituted into the water cloud model to indirectly invert regional soil moisture.

[0009] The invention discloses a regional soil moisture inversion method for inverting canopy parameters, comprising: S1, collecting measured data, which is divided into a training set and a test set, wherein the measured data comprises: SAR backscattering coefficient, leaf area index LAI and plant water content VWC of measured sample points; S2, analyzing the microwave scattering characteristics of crops in the study area, and performing polarization decomposition on SAR remote sensing images to obtain polarization decomposition characteristics with mechanism: SAR backscattering coefficient components and radar polarization characteristic components; S3, random forest canopy parameter inversion: based on the random forest algorithm, combining ground sample point data, performing feature importance evaluation on the SAR backscattering coefficient components and the radar polarization characteristic components, and screening the components with leaf area index LAI, plant water content VWC, and leaf area index LAI. Polarization decomposition feature combination with strong correlation with water content VWC; S4, based on the screened polarization decomposition feature combination, based on the machine learning algorithm, the canopy parameter inversion model is trained to obtain the regional leaf area index LAI and the regional plant water content VWC; S5, water cloud model parameter calibration: the SAR backscattering coefficient, leaf area index LAI and plant water content VWC of the measured sample points in S1 are substituted into the water cloud model, and the parameters of the water cloud model are calibrated in combination with the particle swarm optimization algorithm and the trust region reflection algorithm; S6, using the calibrated water cloud model, coupling the regional leaf area index LAI obtained in S4 to invert soil moisture, and coupling the regional plant water content VWC to invert soil moisture.

[0010] The present invention also correspondingly proposes a regional soil moisture inversion system for inverting canopy parameters, comprising: a computer executable program, which completes the steps of the method described above when executed.

[0011] The beneficial effects of the present invention are as follows. On the one hand, affected by the rapid temporal change of soil moisture, the robustness of the soil moisture inversion model established by directly constructing the relationship between soil moisture and SAR scattering components using a machine learning model is poor. In order to solve this problem, the present invention selects relatively stable crop canopy parameters as input parameters of the semi-empirical inversion model water cloud model. The method of the present invention can not only improve the robustness of the model, but also enhance the adaptability of the model under different soil conditions. In addition, from a mechanistic point of view, the relationship between crop canopy parameters and radar polarization characteristics is relatively clear, which can improve the mechanistic rationality of the semi-empirical model for soil moisture inversion. On the other hand, it is very difficult to obtain the input parameters of the water cloud model within a regional range. At present, the acquisition of canopy parameters such as leaf area index (LAI) and plant water content (VWC) is mainly based on field single-point measurement and then obtained by spatial interpolation or based on optical image inversion. Affected by crop varieties, soil properties and field management measures, crop canopy parameters have significant spatial variability, and the spatial interpolation method cannot accurately obtain the true value of the crop canopy parameters of each pixel. In areas covered by clouds and rain, the optical image quality is poor, and the regional crop canopy parameter inversion based on optical data has large errors. In order to solve this problem, the present invention adopts a random forest model combined with radar polarization decomposition technology to effectively obtain crop canopy parameters such as regional leaf area index (LAI) and vegetation water content (VWC). The method of the present invention can overcome the limitations of single-point measurement and optical inversion methods in obtaining regional crop canopy parameters, and ensure that the water cloud model input parameters have high spatial consistency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to make the present invention more easily understood, the present invention will be described in more detail by referring to the specific embodiments shown in the accompanying drawings. These drawings only depict typical embodiments of the present invention and should not be considered as limiting the scope of protection of the present invention.

[0013] Figure 1 is a flow chart of the method of the present invention.

[0014] Figure 2 A technical roadmap for an embodiment of the method of the present invention.

[0015] Figure 3 This is a schematic diagram of the random forest model.

[0016] Figure 4 An overview map of a study area to which the method of the present invention is applied.

[0017] Figure 5 Schematic diagram of the alpha characteristic component obtained after polarization decomposition of SAR remote sensing images.

[0018] Figure 6Schematic diagram of the Freeman_Vol characteristic components obtained after polarization decomposition of SAR remote sensing images.

[0019] Figure 7 This is the distribution map of ground sampling points.

[0020] Figure 8 The results of random forest feature importance evaluation for LAI are shown.

[0021] Fig. 9 The results of random forest feature importance evaluation for VWC are shown.

[0022] Fig.10 The inversion accuracy of the training set for LAI inversion is shown.

[0023] Fig.11 The inversion accuracy of the test set for LAI inversion is shown.

[0024] Fig.12 The inversion accuracy of the training set for VWC inversion is shown.

[0025] Fig.13 The inversion accuracy of the test set for VWC inversion is shown.

[0026] Fig.14 The inversion accuracy of the soil moisture training set for LAI inversion using coupled radar polarimetric signatures is shown.

[0027] Fig.15 The inversion accuracy of the soil moisture training set of VWC coupled with radar polarimetric signature inversion is shown.

[0028] Fig.16 The regional soil moisture inversion results of the LAI coupled with the radar polarimetric signature are shown.

[0029] Fig.17 The regional soil moisture inversion results of VWC coupled with radar polarimetric signature inversion are shown.

[0030] Fig.18 The inversion accuracy of the soil moisture test set for LAI inverted by coupled radar polarimetric signature is shown.

[0031] Fig.19 The inversion accuracy of the soil moisture test set of VWC coupled radar polarimetric signature inversion is shown. DETAILED DESCRIPTION

[0032] The following describes the implementation modes of the present invention with reference to the accompanying drawings so that those skilled in the art can better understand the present invention and implement it. However, the enumerated embodiments are not intended to limit the present invention. In the absence of conflict, the following embodiments and the technical features in the embodiments may be combined with each other, wherein the same components are represented by the same figure marks.

[0033] Refer to the following Figure 1 One embodiment of the method of the present invention is described below.

[0034] S1, collect measured data, which are divided into training set and test set. The measured data include: SAR backscatter coefficient, leaf area index LAI and plant water content VWC of the measured sample points.

[0035] S2, analyze the microwave scattering characteristics of crops in the study area, and perform polarization decomposition of SAR remote sensing images to obtain mechanistic polarization decomposition characteristics: SAR backscattering coefficient component and radar polarization characteristic component.

[0036] S3, random forest canopy parameter inversion: Based on the random forest algorithm and combined with ground sample data, the feature importance of the SAR backscatter coefficient component and the radar polarization feature component are evaluated to screen the polarization decomposition feature combination with the strongest correlation with leaf area index LAI and plant water content VWC.

[0037] S4, based on the selected polarization decomposition feature combination and the machine learning algorithm, the measured sample data set of each period is divided into a 3:5 ratio as the training set and verification set for canopy parameter inversion, the canopy parameter inversion model is trained with the training set, and the verification set is used for verification, and the crop canopy parameter inversion results of the verification set are substituted into the water cloud model in the next step, and then divided into the training set and test set of the soil moisture inversion model again. Finally, the regional leaf area index LAI and the regional plant water content VWC are obtained.

[0038] S5, water cloud model parameter calibration: Substitute the SAR backscatter coefficient, leaf area index LAI and plant water content VWC of the measured sample points in S1 into the water cloud model, and calibrate the parameters of the water cloud model by combining the particle swarm optimization algorithm and the trust region reflectance algorithm.

[0039] S6, using the calibrated water cloud model, coupling the regional leaf area index LAI obtained in S4 to invert soil moisture, and coupling the regional plant water content VWC to invert soil moisture.

[0040] Preferably, in step S2, the full-polarization SAR remote sensing image is also pre-processed: radiation calibration, multi-view processing, coherent speckle filtering, geocoding and decibel conversion.

[0041] The polarization decomposition in step S2 in a preferred embodiment is described below.

[0042] The essence of polarization decomposition is to decompose complex polarization scattering signals into basic scattering mechanisms. Polarization parameters extracted by different polarization decomposition methods can characterize the scattering characteristics of ground objects and mine hidden ground object information in SAR data. The basic scattering mechanisms mainly include odd scattering (surface scattering), even scattering (dihedral scattering), volume scattering, etc. Polarization decomposition originates from the polarization target decomposition theorem. By decomposing the target scattering matrix, the complex polarization radar echo signal is simplified into intuitive and easy-to-understand parameters, thereby extracting the characteristic information of the target ground object.

[0043] Polarization decomposition methods are mainly divided into two categories, one is coherent target decomposition based on scattering matrix, such as Krogager decomposition, Pauli decomposition, Cameron decomposition, etc.; the other is incoherent target decomposition based on coherence matrix or covariance matrix, such as H / A / α decomposition, Freeman-Durden three-component scattering decomposition, Yamaguchi four-component scattering decomposition, etc. Considering that the crop-covered area in the study area is a distributed radar target, preferably, the present invention adopts H / A / α decomposition and Freeman-Durden three-component scattering decomposition method in the incoherent target decomposition method to perform polarization decomposition of SAR remote sensing images.

[0044] In the first embodiment, the polarization decomposition adopts H / A / α decomposition.

[0045] H / A / α decomposition is a commonly used polarization decomposition method. For pre-processed SAR remote sensing images, select H / A / Alpha Decomposition under Process in the toolbar of the PolSARpro software, select DecompositionParameters, and set the input and output paths in the dialog box. Check the polarization decomposition characteristic parameters to be generated as needed. For Radarsat-2 fully polarized data, Alpha, Entropy (H), and Anisotropy (A) can be selected as the required polarization decomposition characteristic parameters. Click Run to run the H / A / α polarization decomposition program, and obtain the polarization decomposition results stored in the form of an HDR file in the set output path.

[0046] Open the hdr file corresponding to the polarization decomposition characteristic parameter in ENVI software and convert it into TIFF file in turn. Import the TIFF file into ArcMap software, use the multi-value extraction to point tool in the extraction tool of the spatial analysis toolbox, match the polarization decomposition characteristic parameter with the measured sample point through the longitude and latitude coordinates, and obtain the value of each polarization decomposition characteristic parameter corresponding to each measured sample point, which is convenient for subsequent model training and parameter inversion.

[0047] The principle of H / A / α decomposition is shown in formula (1). Cloude and Pottier analyzed the eigenvalues ​​and eigenvectors of the SAR coherence matrix T3 to obtain parameters describing different scattering processes and their respective scattering proportions, mainly including polarization information entropy (H), calculated by formula (2), anisotropy (A), calculated by formula (3) and polarization scattering angle (α), thereby understanding the scattering mechanism and physical characteristics of the target object.

[0048] (1), (2), (3), In the formula, is the coherence matrix, is the characteristic matrix of rank 1 obtained by decomposing the coherence matrix, is the feature matrix The corresponding eigenvalue represents the scattering intensity corresponding to the corresponding eigenvector. is the feature matrix The corresponding eigenvector represents the polarization scattering characteristics corresponding to the corresponding eigenvalue. is the eigenvalue The pseudo probability obtained is , N is the base of the logarithm, N is 3 for a single-station radar, and N is 4 for a dual-station radar. and The eigenvalues ​​are obtained by polarization decomposition, and the eigenvalues ​​are sorted as follows: .

[0049] Through H / A / α decomposition, the polarization information entropy (H), anisotropy (A) and polarization scattering angle (α) are obtained.

[0050] Polarization information entropy (H) reflects the degree of randomness of various scattering mechanisms in a statistical sense, and its value ranges from 0 to 1. A lower polarization entropy H (H<0.3) indicates that the scattering mechanism of the target is certain, and the dominant scattering mechanism can be used as the scattering mechanism of the target; when the value of polarization entropy H is close to 1, it means that the scattering mechanism of the target is highly random, which usually corresponds to a complex scattering model.

[0051] Anisotropy (A) describes the relationship between two relatively weak scattering components obtained by polarization feature decomposition, and its value range is between 0 and 1. Anisotropy A is a supplement to polarization information entropy H, and is usually used to identify scattering mechanisms only when polarization entropy H is greater than 0.7. A value close to 0 indicates that the scattering intensities of the second and third scattering mechanisms of the target are close; A value close to 1 indicates that the second scattering mechanism of the target is absolutely dominant over the third scattering mechanism.

[0052] The polarization scattering angle (α) is a key parameter for identifying the main scattering mechanism, and its value range is between 0° and 90°. The scattering angle α value corresponds to the change of the scattering mechanism. When the α value is 0°, it indicates surface scattering, when the α value is 45°, it indicates volume scattering, and when the α value is 90°, it indicates even scattering.

[0053] Second embodiment - Freeman-Durden three-component decomposition.

[0054] Freeman-Durden three-component decomposition is a polarization decomposition method for full polarization data. For pre-processed SAR remote sensing images, click Process in the toolbar of PolSARpro software, select the Polarimetric Decompositions tool, run the FRE3: Freeman 3 Component Decomposition program, and obtain the polarization decomposition results stored in the form of hdr files in the set output path. Similar to H / A / α decomposition, the polarization decomposition feature parameters are converted into TIFF files in ENVI software, and the feature parameters are matched with the measured sample points in ArcMap software. Freeman and Durden proposed the Freeman-Durden three-component decomposition method, which decomposes the polarization covariance matrix into three basic scattering mechanisms, namely the first-order Bragg surface scattering model describing rough surface scattering, the even scattering model simulating the dihedral reflector, and the volume scattering model. This model is based on the scattering symmetry assumption and requires that there is no correlation between the co-polarization and cross-polarization of the target object, that is, it satisfies equation (4). The calculation method of Freeman-Durden decomposition is shown in equations (5) to (8).

[0055] (4), (5), (6), (7), (8), In the formula, and is the SAR backscatter coefficient of the same polarization channel, and is the SAR backscatter coefficient of the cross-polarization channel, is the surface scattering weighting coefficient, is the surface scattering covariance matrix, is the even-order scattering weighting coefficient, is the even-order scattering covariance matrix, is the volume scattering weighting coefficient, is the volume scattering covariance matrix, is the even-order scattering parameter, is the surface scattering parameter.

[0056] The random forest algorithm of step S3 is described below.

[0057] In S3, when screening the polarization decomposition feature combination with the strongest correlation with leaf area index LAI and plant water content VWC, the feature importance evaluation method in the random forest model, that is, the reduction of Gini impurity or the reduction of mean square error, is used to calculate the input feature importance score. According to the empirical division method, the threshold of the feature importance score is set to 0.01, and the feature values ​​with feature importance scores of 0.01 and above in the random forest feature importance evaluation are selected as the feature values ​​participating in the subsequent canopy parameter inversion.

[0058] The random forest model is an integrated learning method based on the combination of multiple decision trees. It improves the robustness and accuracy of the random forest model regression by integrating the prediction results of multiple decision trees. It also has great advantages in processing high-dimensional data, evaluating feature importance, and reducing overfitting.

[0059] The random forest model structure is as follows Figure 3 As shown. Random forest divides the measured sample data set into training set and test set for model training and evaluation. The random forest model uses the bagging bootstrap aggregation method to reduce the generalization error and improve the regression accuracy of the model. The random forest first uses the bootstrap sampling method to generate a decision tree, that is, a number of samples are uniformly and with replacement extracted from the original data set to form multiple training sets, and a separate decision tree is constructed for each training set. The nodes of the decision tree represent subsets of the data set. In the process of building a decision tree, when each node splits, the model will randomly select some features instead of all features to find the best split point to improve the diversity of the decision tree, avoid the model's dependence on a single feature and the risk of overfitting, reduce the model's sensitivity to noise and outliers, and thus improve the model's generalization ability and robustness. The target parameter inversion belongs to a regression problem. The random forest model trains the randomly selected feature subsets and data samples separately, performs regression analysis, takes the average of the prediction results of all decision trees through the combiner, and outputs the final prediction value of the target parameter. In addition, the model is verified and optimized by combining cross-validation, accuracy assessment, model optimization and other steps to achieve accurate and efficient inversion of soil moisture.

[0060] The process of inverting the target parameters using the random forest model is to use the random forest to establish the mapping relationship between the target parameters and the input covariates, which can be expressed as equations (9) to (12).

[0061] (9), (10), (11), (12), In the formula, is the target inversion parameter, is a function that uses the random forest model to invert canopy parameters, is the input feature dataset of the random forest model (polarization decomposition features obtained in step S2), is the inversion error term, which represents the random error or noise of the model, To include The feature vector of features, for dimensional real vector space, It is a feature At the new data point The final weight on It is Features in a tree For data points The weight of It is the first The randomized feature selection used by the trees, is the number of trees.

[0062] In one embodiment, step S3 includes: S31, dividing each period of measured sample points into a training sample set and a verification sample set for target parameter inversion. S32, for the training sample set data, inputting the measured data and the corresponding eigenvalues ​​into the random forest model for feature importance evaluation. S33, based on the evaluation results, eigenvalues ​​with feature importance scores greater than or equal to 0.01 are regarded as important eigenvalues ​​and retained. S34, training and optimizing the random forest regression model through grid search and 10-fold cross validation, and applying the optimal model to the verification set to perform regional inversion and accuracy verification of the target parameters.

[0063] In step S4, the random forest model is selected to invert the crop canopy parameters LAI and VWC. According to the result of the feature optimization in step S3, the measured sample data of each period is divided into a training data set and a validation data set for canopy parameter inversion in a ratio of 3:5. For the training set, the method of ten-fold cross validation combined with grid search is used to determine the number of decision trees, the maximum depth of the tree, the minimum number of samples required for internal node splitting, and the minimum number of samples for leaf nodes. The random forest inversion model is trained and optimized, and then applied to the validation set for the inversion of LAI and VWC to obtain the single-point inversion results of the validation set of LAI and VWC. The random forest inversion model constructed by the training set is applied to the regional LAI and VWC inversion to obtain the inversion results of the regional LAI and VWC.

[0064] The present invention adopts a water cloud model. In step S5, the parameters of the water cloud model are calibrated, and the calibrated water cloud model is used in S6. The water cloud model assumes that the crop is a "cloud" composed of a large number of uniformly distributed identical water particles. It only considers the volume scattering from the crop reflection and the backscattering of the ground after the double-layer attenuation of the crop, and ignores other forms of scattering in the crop and the soil. The variables in the model only include the radar incident angle, crop water content and soil humidity. The parameters of the water cloud model are easy to obtain and the form is simple. It can effectively separate the backscattering coefficients of the crop layer and the soil layer. The specific form is shown in Equation (13)-Equation (15). The backscattering coefficient of the soil layer Soil moisture content There is a simple linear relationship between them, see formula (16), therefore, the soil moisture inversion in the crop-covered area based on the water cloud model can be expressed by formula (17).

[0065] (13), (14), (15), (16), (17), In the formula, is the total backscatter coefficient of the vegetation canopy at any polarization, subscript , are the transmitting and receiving polarization modes, is the backscattering coefficient of the crop layer, is the backscattering coefficient of the soil layer, is the crop canopy parameter. The plant water content VWC and leaf area index LAI were taken respectively. is the double-layer attenuation factor of microwave penetration into the crop layer, is the radar incident angle, is the soil moisture content in the crop-covered area, and A, B, C, and D are the water cloud model parameters that need to be calibrated.

[0066] In the process of calibrating the water cloud model parameters, the function of the water cloud model is defined using formula (17). First, the canopy parameter validation set obtained by random forest inversion is divided into a training set and a test set, and the training set is divided into 10 folds using KFold. For each fold of the training set data, the particle swarm optimization algorithm (PSO) is first used to find the initial parameter combination of the water cloud model parameters, and then the trust region reflective algorithm (TRF) is used to finely solve the water cloud model in the local area based on the initial parameter combination of the water cloud model to determine the precise parameter combination of the water cloud model. Furthermore, the water cloud model is calibrated 10 times using the training set data, and the validation set error in the cross-validation is compared. The water cloud model parameter combination with the smallest error is taken as the optimal parameter combination, and the regional inversion of the water cloud model and the accuracy verification of the test set are performed.

[0067] The following describes an application example of the present invention. The method of the present invention has been verified. The research area selects City A, a major rapeseed producing area in China, as the research area to verify the feasibility and applicability of the regional soil moisture inversion technology for canopy parameters inverted by coupling radar polarization characteristics. City A has a subtropical monsoon climate with four distinct seasons and abundant rainfall. The main crop grown is rapeseed. An overview of the research area is as follows: Figure 4 shown.

[0068] The research data uses RADARSAT-2 fully polarized SAR remote sensing images of the study area on November 30, 2023, January 17, 2024, March 5, 2024, and March 29, 2024. RADARSAT-2 is a Canadian high-resolution SAR satellite equipped with a C-band sensor. It was launched in 2007, with a revisit period of 24 days and a maximum resolution of 1m. It includes four polarization modes: HH, HV, VH, and VV. It has the characteristics of short revisit period, high spatial resolution, strong data storage capacity, and accurate satellite positioning. It is currently widely used in agriculture, surveying and mapping, environmental resources and other fields.

[0069] In this embodiment, SNAP 9.0 software is used to preprocess the RADARSAT-2 fully polarized SAR remote sensing image. The main operation steps are as follows. (1) Radiometric calibration. The Calibrate tool in the SNAP software is used to perform radiometric calibration of the RADARSAT-2 SAR remote sensing image, and the DN value of the original image is converted into a backscattering coefficient with a clear physical meaning. (2) Multi-looking processing. The Multilooking tool in the SNAP software is used to perform multi-looking processing of the SAR remote sensing image. The number of rows and columns of the multi-looking processing is set to 2×3, that is, 2 in the range direction and 3 in the azimuth direction. Multi-looking processing helps to reduce the coherent speckle noise of the SAR remote sensing image, improve the signal-to-noise ratio, and enhance the data quality. (3) Coherent speckle filtering. The Single Product SpeckleFilter tool in the SNAP software is used to perform coherent speckle filtering, and the spatial filtering of the Refined Lee adaptive window is selected to smooth the coherent speckle noise in the image. (4) Geocoding. The Range Doppler Terrain Correction tool in the SNAP software is used for geocoding, and the DEM elevation data is SRTM 1 Sec HGT. Terrain correction converts SAR remote sensing images to a geographic coordinate system so that they can be aligned and analyzed with other geographic data. (5) Decibelization. Use the Converts bands to / from dB tool in the SNAP software to decibelize SAR remote sensing images. Decibelization is the process of converting the backscatter coefficient from a linear scale to a logarithmic scale (decibels) to facilitate the subsequent application of SAR backscatter coefficient parameters.

[0070] After obtaining the backscatter coefficients of each polarization channel of the SAR remote sensing image, the PolSARpro 6.0 software is used to obtain the polarization characteristic components of the SAR remote sensing image through the HA-α decomposition and Freeman-Durden polarization decomposition methods. The schematic diagram is shown in the figure. Figure 5 and Figure 6 As shown, Figure 5 Represents a schematic diagram of the alpha feature component, Figure 6 Schematic diagram showing the Freeman_Vol characteristic components.

[0071] Obtaining ground observation data: According to the distribution of rapeseed in the study area, 32 ground measurement sampling points were set up in the study area, and field observation experiments were carried out from December 1 to 3, 2023, January 13 to 15, 2024, March 2 to 4, 2024, and March 29 to 31, 2024, to obtain 128 measured sample data in four stages: seedling stage, bud stage, flowering stage, and silique maturity stage of rapeseed. The leaf area index LAI was measured using the LAI-2200 C measuring instrument. The LAI-2200 C uses five "fisheye" optical sensors with different inclination angles to measure the transmitted light at five angles above and below the crop canopy. The radiation transmission model is used to calculate the leaf area index, average inclination angle, and canopy porosity of the crop canopy. Repeat the measurement multiple times to obtain the average LAI value of the sampling point. The soil moisture content was measured using the TDR 350 meter. The TDR 350 is a portable rapid three-parameter soil moisture, temperature and conductivity meter. It measures the volumetric moisture content of the surface soil based on time domain reflectometry. Each sampling point was measured three times, and the average value was taken as the soil moisture measurement value. The fresh weight and dry weight of the crop are used to calculate the volumetric moisture content of the plant. The difference between the fresh weight and dry weight of the crop is first calculated, and then divided by the fresh weight of the crop to obtain the volumetric moisture content of the crop.

[0072] Determine the training set: set 10 sample points numbered as multiples of 3 in each period of measured sample points as soil moisture inversion test set sample points, and the remaining 22 sample points as training set sample points. In the training set, the ground sampling points are distributed as follows: Figure 7 shown.

[0073] Random forest feature importance assessment: The backscatter coefficient obtained after SAR remote sensing image preprocessing and the polarization feature component obtained after polarization decomposition are input into the random forest model as feature vectors, and feature importance assessment is performed for the inversion of LAI and VWC respectively. The feature importance score is not less than 0.01. The feature value is regarded as the feature value with high feature importance, and the subsequent random forest target parameter inversion is performed. Taking the first period of data on November 30, 2023 as an example, the feature importance assessment results are as follows: Figure 8 and Fig. 9 As shown. Figure 8 The importance evaluation results of the random forest features of LAI are shown. Fig. 9 The importance evaluation results of random forest features of VWC are shown. From the feature importance evaluation results, it can be concluded that the feature importance scores of all feature values ​​of LAI and VWC in each period are above 0.01. After the feature importance test, all feature values ​​are retained for the subsequent crop canopy parameter inversion.

[0074] Crop canopy parameter inversion modeling and accuracy verification based on random forest: (1) LAI inversion: Based on the feature importance evaluation results of LAI inversion, the feature values ​​with feature importance scores not less than 0.01 are input into the random forest model. The 12 sample points numbered 1 and 3n-1 in the training set are used as the training set for LAI inversion, and the remaining 10 sample points and the 10 test sets of soil moisture inversion are used as the validation set for LAI inversion (a total of 20 sample points). Combined with 10-fold cross-validation and grid search, the inversion modeling and accuracy verification of the LAI random forest regression model are carried out. Fig.10 and Fig.11 As shown in Figure 2, the LAI inversion accuracy based on random forest is higher. Fig.10 The inversion accuracy of the training set for LAI inversion is shown. Fig.11 The inversion accuracy of the validation set for LAI inversion is shown.

[0075] (2) VWC inversion: Based on the feature importance evaluation results of VWC inversion, the feature values ​​with feature importance scores not less than 0.01 were input into the random forest model. The training set and validation set of VWC inversion were divided in the same way as LAI inversion. The inversion modeling and accuracy verification of the VWC random forest regression model were performed by combining 10-fold cross validation and grid search. Fig.12 and Fig.13 As shown in Figure 2, the VWC inversion based on random forest has a higher accuracy. Fig.12 The random forest single-point inversion soil moisture results of the training set of VWC inversion are shown. Fig.13 The random forest single-point inversion soil moisture results for the test set of VWC inversion are shown.

[0076] Water cloud model parameter calibration: 10 sample points with serial numbers multiple of 3 in the canopy parameter validation set of each period of random forest inversion are used as the test set, and the remaining 10 sample points are used as the training set. The radar characteristic values ​​of the training set, the canopy parameter inversion results and the measured values ​​of soil moisture content are input into the water cloud model, and the parameters of the water cloud model are calibrated by combining 10-fold cross validation, particle swarm optimization algorithm and trust region reflective algorithm to select the best parameter combination of the water cloud model. The parameter calibration results of the water cloud model are shown in Table 1.

[0077] Table 1 Parameter calibration results of water cloud model

[0078] Regional soil moisture inversion of canopy parameters by coupling radar polarimetric characteristics: Substitute the LAI and VWC regional results of random forest inversion into the water cloud model after parameter calibration to obtain the regional results of soil moisture inversion of the water cloud model. In the test set, R² and nRMSE are used as accuracy evaluation indicators of soil moisture inversion by the water cloud model to verify the inversion accuracy. Figure 14-Figure 19Taking the VH polarization method as an example, the regional soil moisture inversion results of canopy parameters inverted by coupling radar polarization characteristics are shown. Figure 14-Figure 19 It can be seen that the regional soil moisture inversion accuracy of VWC coupled with radar polarization characteristics is relatively high, R² can reach 0.62, and the regional soil moisture inversion accuracy of LAI coupled with radar polarization characteristics can reach 0.46, which improves the robustness of the inversion model compared with the random forest direct inversion of soil moisture. Figure 14-15 They represent the soil moisture inversion training set accuracy of LAI and VWC inverted by coupled radar polarization characteristics, Figure 16-17 They respectively represent the regional inversion effect of soil moisture under VH polarization when LAI and VWC are used as canopy parameters for inversion. Figure 18-Figure 19 The soil moisture inversion test set accuracy of LAI and VWC inverted by coupled radar polarization characteristics are shown respectively.

[0079] The embodiments described above are only preferred specific implementations of the present invention. This specification uses the phrases "in one embodiment", "in another embodiment", "in yet another embodiment" or "in other embodiments" to refer to one or more of the same or different embodiments according to the present disclosure. Common changes and substitutions made by those skilled in the art within the scope of the technical solution of the present invention should be included in the protection scope of the present invention.

Claims

1. A regional soil moisture inversion method for inverting canopy parameters, characterized in that: include: S1, collect measured data, which are divided into training set and test set. The measured data include: SAR backscatter coefficient, leaf area index LAI and plant water content VWC of the measured sample points; S2, analyze the microwave scattering characteristics of crops in the study area, and perform polarization decomposition on SAR remote sensing images to obtain mechanistic polarization decomposition characteristics: SAR backscattering coefficient component and radar polarization characteristic component; S3, random forest canopy parameter inversion: Based on the random forest algorithm, combined with ground sample data, the feature importance of the SAR backscatter coefficient component and the radar polarization feature component is evaluated, and the polarization decomposition feature combination with strong correlation with leaf area index LAI and plant water content VWC is screened; S4, based on the selected polarization decomposition feature combination and based on a machine learning algorithm, a canopy parameter inversion model is trained to obtain a regional leaf area index LAI and a regional plant water content VWC; S5, water cloud model parameter calibration: Substitute the SAR backscatter coefficient, leaf area index LAI and plant water content VWC of the measured sample points in S1 into the water cloud model, and calibrate the parameters of the water cloud model by combining the particle swarm optimization algorithm and the trust region reflectance algorithm; S6 uses the calibrated water cloud model to couple the regional leaf area index LAI obtained in S4 to invert soil moisture, and couples the regional plant water content VWC to invert soil moisture.

2. The method according to claim 1, characterized in that In step S2, the mechanistic polarization decomposition feature is obtained by H / A / α decomposition or by three-star scattering component decomposition.

3. The method according to claim 1, characterized in that: In step S3, When screening the polarization decomposition feature combination with the strongest correlation with leaf area index LAI and plant water content VWC, the feature importance evaluation method in the random forest model was used to calculate the input feature importance score.

4. The method according to claim 3, characterized in that The eigenvalues ​​with feature importance scores of 0.01 and above in the random forest feature importance assessment were used as the eigenvalues ​​involved in the subsequent canopy parameter inversion.

5. The method according to claim 4, characterized in that Step S3 includes: S31, dividing each period of measured sample points into a training sample set and a verification sample set for target parameter inversion; S32, for the training sample set data, inputting the measured data and the corresponding feature values ​​into the random forest model to evaluate the feature importance; S33, according to the evaluation results, the feature values ​​with feature importance scores greater than or equal to 0.01 are regarded as important feature values ​​and retained; S34, a random forest regression model was trained and optimized through grid search and 10-fold cross validation, and the optimal model was applied to the validation set for regional inversion and accuracy verification of the target parameters.

6. The method according to claim 1, characterized in that Step S5 includes: S51, dividing the canopy parameter validation set obtained by random forest inversion into a training set and a test set, and performing 10-fold division on the training set; S52, for each fold of training set data, use the particle swarm optimization algorithm to find the initial parameter combination of the water cloud model parameters, and then use the trust region reflective algorithm to finely solve the initial parameter combination of the water cloud model in the local area to determine the precise parameter combination of the water cloud model; S53, using the training set data to calibrate the water cloud model 10 times, comparing the validation set error in cross-validation, and taking the water cloud model parameter combination with the smallest error as the optimal parameter combination.

7. The method according to claim 1, characterized in that The water cloud model is: , In the formula, is the total backscatter coefficient of the ground canopy covered by vegetation at any polarization, is plant water content VWC and leaf area index LAI, is the radar incident angle, is the soil moisture content in the crop-covered area, and A, B, C, and D are the water cloud model parameters that need to be calibrated.

8. The method according to claim 7, characterized in that Step S2 also includes: preprocessing the fully polarized SAR remote sensing image, including: radiation calibration, multi-view processing, coherent speckle filtering, geocoding and decibel conversion.

9. The method according to claim 5, characterized in that In step S4, According to the result of feature optimization in step S3, the measured sample data of each period is divided into a training data set and a verification data set for canopy parameter inversion in a ratio of 3:5; For the training set, a ten-fold cross validation combined with a grid search method was used to determine the number of decision trees, the maximum depth of the tree, the minimum number of samples required for internal node splitting, and the minimum number of samples for leaf nodes, and to train and optimize the random forest inversion model. It is then applied to the validation set to invert LAI and VWC, and the single-point inversion results of the validation set of LAI and VWC are obtained.

10. A regional soil moisture inversion system for inverting canopy parameters, characterized in that: The method comprises a computer executable program, and when the program is executed, the method according to any one of claims 1 to 9 is implemented.

Images