Grassland vegetation coverage inversion method based on polarimetric SAR data

Abstract

The invention discloses a grassland vegetation coverage inversion method based on polarimetric SAR (Synthetic Aperture Radar) data in the technical field of remote sensing. The grassland vegetation coverage inversion method comprises the following steps of S1, acquiring the polarimetric SAR data and preprocessing the polarimetric SAR data; s2, obtaining a polarization coherence matrix through a similar pixel statistical method; s3, performing polarization decomposition on the polarization coherence matrix, and calculating RVI; s4, correcting the RVI by adopting an H / A parameter, and carrying out polarization weighting on the corrected RVI to obtain an ERVI; and S5, estimating the grassland vegetation coverage through ERVI inversion. The method is based on polarimetric SAR data, combines similar pixel statistics and polarimetric weighting methods, performs RVI correction and weighting to construct ERVI, realizes high-precision estimation of the grassland vegetation coverage, has the advantages of high precision, strong robustness, wide applicability and the like, is obviously improved in spatial precision and time precision, and is suitable for popularization and application. And particularly, the estimation effect in a low-coverage and sparse vegetation region is more accurate.

Description

Technical Field

[0001] This invention relates to the field of remote sensing technology, and in particular to a method for inverting grassland vegetation cover based on polarimetric SAR data. Background Technology

[0002] Grassland vegetation cover (FVC) is a key indicator for assessing grassland ecosystem health, land degradation, and biodiversity changes. FVC estimation typically relies on ground-based measurements and remote sensing techniques. However, ground-based measurements suffer from high costs, limited spatial coverage, and require large-scale manual sampling, making efficient and widespread monitoring difficult.

[0003] With the development of remote sensing technology, satellite remote sensing data has become the main means of large-scale vegetation monitoring, especially optical remote sensing and synthetic aperture radar (SAR) data, which are widely used for FVC estimation. Optical remote sensing technology can estimate FVC through indicators such as the Normalized Difference Vegetation Index (NDVI). However, optical remote sensing data is easily affected by factors such as cloud cover and climate change, resulting in intermittent data acquisition and limiting its application under complex weather and low-light conditions. Unlike optical remote sensing, SAR has all-weather, all-time imaging capabilities and can penetrate clouds and weather interference, giving it a significant advantage for vegetation monitoring in low-coverage areas such as grasslands. Radar Vegetation Index (RVI), as a vegetation cover estimation method based on SAR data, has been widely used in various vegetation monitoring applications. The traditional RVI method is based on the backscattering coefficient in SAR data and calculates vegetation characteristics between different pixels through simple arithmetic operations (such as addition, subtraction, multiplication, division, difference, and differentiation).

[0004] Grasslands, as an important ecosystem type, differ significantly from forests and farmland, primarily in their generally low vegetation cover, short vegetation height, and irregular spatial distribution. Due to these characteristics, the representation of grassland information in remote sensing monitoring exhibits significant uncertainty and spatial heterogeneity, posing a considerable challenge to the accurate inversion of vegetation parameters. In studies using Relative Visor Analyzer (RVI) for FVC estimation, traditional RVI is mainly constructed based on scattering characteristics within a spatial neighborhood, suitable for densely vegetated and uniformly covered forests and farmland. However, in sparsely vegetated and heterogeneous grassland environments, vegetated areas often contain a large amount of non-vegetated components, and vegetation cover varies significantly across different locations, easily leading to heterogeneous effects between pixels. This causes information mixing in the construction method based on neighborhood pixel statistics, affecting the RVI's responsiveness to vegetation cover. Furthermore, most existing methods for improving RVI focus on enhancing its applicability in high-density vegetation areas. However, for low-biomass areas such as grasslands, due to the low height and sparse cover of the vegetation canopy, the radar volume scattering component is weak, and RVI is not sensitive enough to subtle changes in vegetation, thus limiting its estimation accuracy and application effectiveness in grassland areas.

[0005] Therefore, with the increasing demand for grassland ecological monitoring, there is an urgent need to construct a vegetation cover estimation method that can effectively enhance the expression of scattering features and reduce inter-pixel interference, given the low coverage and high heterogeneity of grasslands, in order to improve the stability and accuracy of remote sensing inversion results. Summary of the Invention

[0006] This application provides a grassland vegetation cover inversion method based on a novel radar vegetation index (ERVI), which solves the problem of low accuracy of traditional RVI vegetation cover estimation methods in grassland areas and achieves more accurate estimation in areas with low coverage and sparse vegetation.

[0007] This application provides a method for inverting grassland vegetation cover based on polarimetric SAR data, including the following steps: S1: Acquire polarimetric SAR data and perform preprocessing; S2: Obtain the polarization coherence matrix through similar pixel statistical methods; S3: Perform polarization decomposition on the polarization coherence matrix and calculate RVI; S4: The RVI is corrected using the H / A parameter, and the corrected RVI is polarized-weighted to obtain the ERVI; S5: Estimate grassland vegetation cover using ERVI inversion.

[0008] The beneficial effects of the above embodiments are as follows: This method, based on polarimetric SAR data and combining similar pixel statistics and polarimetric weighting, corrects and weights RVI to construct ERVI, achieving high-precision estimation of grassland vegetation cover. It has advantages such as high accuracy, strong robustness, and wide applicability. This method achieves significant improvements in both spatial and temporal accuracy, especially in the estimation of low-coverage and sparse vegetation areas.

[0009] Based on the above embodiments, this application can be further improved as follows: In one embodiment of this application, in step S1, the polarimetric SAR image data is GF-3 fully polarimetric SAR data. Technical effect: Utilizing the high resolution and all-weather imaging capabilities of the GF-3 satellite, combined with a standardized data processing workflow, high-quality polarimetric scattering foundational data is provided for subsequent analysis.

[0010] In one embodiment of this application, step S1, the preprocessing includes radiometric calibration, filtering, terrain correction, and geocoding. The geocoding uses SRTM DEM data to register the image with geographic coordinates. Technical effect: High-precision terrain data correction ensures accurate matching between the polarimetric coherence matrix and the geospatial location, reducing the impact of terrain interference on subsequent index calculations.

[0011] In one embodiment of this application, step S2, the similarity pixel statistics method includes: S2.1: Set a fixed search window centered on each pixel; within the search window, filter out a set of pixels with similar characteristics to the central pixel by calculating the similarity of neighboring pixels and the central pixel in terms of polarization features; S2.2: Assign weights based on similarity and normalize them; S2.3: Perform a weighted average on the set of similar pixels to generate a weighted polarization coherence matrix. Technical effects: Suppresses random noise interference, preserves spatial structure features of ground features, and improves the stability and reliability of polarization information.

[0012] In one embodiment of this application, step S3 includes: S3.1: The polarization decomposition method is used to process the polarization coherence matrix to obtain the surface scattering power (PS), secondary scattering power (PD), and volume scattering power (PV) of each pixel. S3.2: Calculate RVI as follows: RVI = PV / (PS + PD + PV).

[0013] In one embodiment of this application, in step S3.1, the polarization decomposition method is HTCD (hybrid three-component decomposition) polarization decomposition. Technical effect: It obtains the surface scattering power (PS), secondary scattering power (PD), and volume scattering power (PV) for each pixel, providing a basis for subsequent exponential calculations.

[0014] In one embodiment of this application, step S4 includes: S4.1: Divide the study area into sub-regions and extract the H and A parameters of each sub-region using the H / A / alpha polarization decomposition method; S4.2: Using the NDVI-FVC calculated from the optical image as reference data, calculate the correlation coefficients of H and A parameters with FVC in each sub-region; S4.3: Select the parameter with the higher absolute value of the correlation coefficient from the H and A parameters as the correction parameter for each sub-region; S4.4: Introduce H / A decomposition parameters to correct RVI; S4.5: The modified RVI is weighted by volume scattering polarization to obtain the ERVI; specifically: ERVI=RVI*H*PV or ERVI=RVI*(1-A)*PV.

[0015] Technical effects: Strengthening the weight of volume scattering information, improving the sensitivity of the index to sparse vegetation through regional adaptive parameter correction, and expanding the dynamic range. Whether each sub-region adopts equation (3) or equation (4) is determined by the correction parameters of each sub-region in step S4.3.3; thereby realizing regional differentiated parameter optimization, ensuring high correlation between ERVI and vegetation cover, and improving the estimation adaptability.

[0016] In one embodiment of this application, step S4.5 involves introducing a power transformation to amplify ERVI in a non-linear manner, specifically as follows: ERVI=[RVI*(1-A)*PV] m Or ERVI = [RVI * (1 - A) * PV] m .

[0017] Technical effect: The original values ​​are usually distributed within a small range, resulting in a limited overall dynamic range. Introducing a power transform amplifies the originally weak but physically directional signal in a non-linear manner, enhancing its expressive power. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0019] Figure 1 This is a flowchart illustrating the steps of a grassland vegetation cover inversion method based on polarimetric SAR data in an embodiment of this application. Figure 2 This is an optical image of the study area in the embodiments of this application; Figure 3 This is a SAR image of the study area in the embodiments of this application; Figure 4 This is a schematic diagram of similar pixel feature extraction in an embodiment of this application; Figure 5 This is a schematic diagram of the optical reference FVC in the embodiments of this application. Detailed Implementation

[0020] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0021] Example: like Figure 1 As shown, a grassland vegetation cover inversion method based on polarimetric SAR data consists of the following steps: S1. Acquire polarimetric SAR data and perform preprocessing.

[0022] Preprocessing includes radiometric calibration, filtering, terrain correction, and geocoding.

[0023] The study area selected in this embodiment is located in the Xilingol League Grassland Ecological Zone of Inner Mongolia, and the terrain features in the study area are predominantly grassland. Grassland types include Stipa grandis and Leymus chinensis, etc. Figure 2 As shown. Acquire GF-3 fully polarimetric SAR image data covering the above area, as shown. Figure 3As shown, the satellite is equipped with a flexible imaging system, supporting multiple polarization modes (including HH, HV, VH, and VV) and various imaging modes, providing all-weather, all-day Earth observation capabilities. The GF-3 fully polarimetric SAR data used in this embodiment was acquired on June 10, 2023, and is C-band (5.4 GHz) fully polarimetric data. The spatial resolution of this data is approximately 2.25 m × 5 m, with a coverage area of ​​approximately 17 km × 38 km. To ensure consistency in imaging geometry and reduce variations in scattering characteristics caused by differences in observation angles, the incident angle of the selected data was controlled within the range of 35°–45°. The data acquisition date was chosen during the middle of the grassland growing season (June to August) to avoid bias caused by seasonal differences in vegetation. In radiometric calibration, the calibration parameters were extracted from the metadata file (.xml) accompanying the image.

[0024] To achieve accurate matching with geospatial locations, SAR data images were geocoded using SRTM (Shuttle Radar Topography Mission) digital elevation model (DEM) data released by NASA, with a resolution of 30 meters, to achieve accurate registration of images with geographic coordinates.

[0025] S2. Obtain the polarization coherence matrix through similar pixel statistics.

[0026] First, based on GF-3 fully polarimetric SAR data, an initial polarimetric coherence matrix is ​​constructed according to standard procedures to describe the scattering correlation of pixels in different polarimetric channels.

[0027] Secondly, to reduce the influence of heterogeneous pixels, a similar pixel statistical method is introduced. This method selects pixels with similar polarization characteristics to the central pixel within a local area for weighted statistical analysis, thereby obtaining more stable and reliable polarization information. The specific steps are as follows: A fixed search window (e.g., 5×5 or 7×7 pixels) is set around each pixel. Within this window, a set of pixels with similar characteristics to the central pixel is selected by calculating the similarity of neighboring pixels with the central pixel in terms of polarization characteristics (e.g., differences in scattering intensity, polarization ratio, etc.).

[0028] Each pixel is assigned a different weight based on its similarity to the central pixel; pixels with higher similarity have higher weights, while pixels with lower similarity have lower weights. All weights are normalized to ensure physical consistency in the overall statistics.

[0029] After determining the weight distribution, a weighted average operation is performed on the selected set of similar pixels to generate a weighted polarization coherence matrix. Compared with the traditional mean statistical method, this method can effectively reduce the interference of heterogeneous pixels, so that the polarization information of each pixel is mainly determined by pixels with similar scattering characteristics, thus more accurately reflecting the structural attributes and vegetation characteristics of the pixel itself.

[0030] Applying this method to all pixels of the entire image generates a polarization coherence matrix based on similar pixel statistics, providing a reliable data foundation for subsequent polarization decomposition and radar vegetation index (RVI) calculation.

[0031] S3. Perform polarization decomposition on the polarization coherence matrix and calculate the RVI (radar vegetation index).

[0032] S3.1. In order to fully characterize the electromagnetic scattering characteristics of grassland vegetation, this embodiment uses the polarization decomposition method to process the polarization coherence matrix to obtain the surface scattering power (PS), secondary scattering power (PD) and volume scattering power (PV) of each pixel, providing a basis for subsequent index calculation.

[0033] This study employs the Hybrid Three-Component Decomposition (HTCD) algorithm as the polarization decomposition method. Compared to traditional polarization decomposition algorithms, it combines matrix rotation theory with a hybrid scattering model, enabling a more rational allocation of scattering power and thus achieving appropriate adaptation to different types of vegetation structures. Furthermore, when different volume scattering models are selected for decomposition, negative power can be eliminated to varying degrees, providing a strong guarantee for the accuracy of RVI in FVC estimation in grassland areas.

[0034] S3.2. Calculate RVI as follows: (1); RVI primarily reflects the contribution of volume scattering to total scattering. PV was chosen as the molecule because in grassland ecosystems, vegetation volume scattering is particularly sensitive to C-band SAR signals, while surface scattering and secondary scattering are significantly affected by soil, topography, or background clutter. Therefore, RVI can more directly reflect the degree of grassland vegetation cover.

[0035] S4. The RVI is corrected using the H / A parameter, and the corrected RVI is polarized weighted to obtain the ERVI.

[0036] In the H / A / α (polarization entropy / mean scattering angle / anisotropy) decomposition method, the three parameters H (polarization entropy), A (anisotropy), and α (mean scattering angle) describe the complexity of ground cover scattering, the heterogeneity of scattering mechanism distribution, and the directionality of the main scattering mechanism, respectively. Since there are significant differences in ground cover types and vegetation cover in different regions, the sensitivity of these parameters to grassland vegetation cover (FVC) also varies. Therefore, it is necessary to select parameters based on regional characteristics to ensure that the selected parameters have a high correlation with vegetation (Pearson correlation analysis is used). This step specifically involves: S4.1: Preliminary analysis of regional characteristics. The study area is divided into sub-regions, and the vegetation type, coverage, and spatial heterogeneity of each sub-region are analyzed.

[0037] Extract the H / A / α parameters. Apply the H / A / α decomposition method to each pixel to obtain the spatial distribution map of the three parameters: polarization entropy H, anisotropy A, and average scattering angle α.

[0038] S4.2: Correlation Analysis. Using NDVI-FVC calculated from optical images as reference data, statistical analysis was performed on the correlation between each parameter and FVC in each sub-region. The explanatory power of H, A, and α for FVC was evaluated by calculating correlation coefficients (such as Pearson correlation coefficient).

[0039] S4.3: Parameter selection. Based on the statistical results, the parameter with the highest correlation to FVC is selected as the representative parameter for this region (in practice, α has a poor correlation, so only one is selected from H and A).

[0040] For the entire study area, polarization feature matrices can be constructed using H / A parameter combinations selected from different sub-regions to ensure that the parameters used in each region have maximum sensitivity to FVC.

[0041] S4.4: Considering the significant differences in land cover features and vegetation structure in different regions, this embodiment introduces H / A decomposition parameters in RVI for correction.

[0042] Polarization entropy H describes the complexity and randomness of the scattering mechanism. It is highly sensitive to volume scattering in sparsely vegetated areas and can reflect subtle changes in vegetation structure.

[0043] Anisotropy A describes the non-uniformity of power distribution for different scattering mechanisms, and helps to distinguish the contributions of surface scattering, secondary scattering, and volume scattering in complex or high-coverage regions.

[0044] After selecting appropriate parameters based on regional characteristics, H or A is added to the ERVI calculation to achieve adaptive enhancement of the region. The specific form is as follows: RVI * H or RVI * (1-A); In practical assessments, the indicator most significantly correlated with FVC was selected. In this study area, (1-A) was chosen to correct RVI. Introducing (1-A) can suppress RVI values ​​in non-vegetation scattering-dominated areas, thereby highlighting the signal contrast in vegetated areas.

[0045] S4.5: The modified RVI is weighted by volume scattering polarization to obtain the ERVI.

[0046] The corresponding formula is: ERVI=RVI*H*PV or ERVI=RVI*(1-A)*PV(2); RVI is defined as the ratio of volume scattering power to total power, and its value ranges from 0 to 1. Volume scattering power typically also falls within this range. Therefore, RVI*PV reduces the overall value of the original RVI, but this reduction is relative and exhibits a selective modulation effect. Specifically, in non-vegetated areas, the proportion of volume scattering is extremely low, and the product of RVI and PV significantly suppresses the original RVI; while in vegetated areas, the high proportion of volume scattering enhances the original RVI. Therefore, this method possesses the characteristic of "low-value suppression and high-value enhancement," which helps distinguish between grassland vegetated areas and non-vegetated areas. It is worth noting that the value of RVI*PV is usually distributed within a relatively small range, resulting in a limited overall dynamic range.

[0047] To further enhance its expressive power, this embodiment introduces a power transform to amplify the originally weak but physically directional signal in a non-linear manner. The corresponding formula is: ERVI = [RVI * H * PV] m Or ERVI = [RVI * (1 - A) * PV] m (3); To reasonably determine the power exponent *m* introduced in ERVI, this study uses the mean and variance of the features after exponential transformation as the main criteria. Specifically, within the range of *m* = 0.05–0.50, values ​​are sequentially taken in steps of 0.05 to perform exponential transformations on volume scattering-related features, and the mean and variance of the transformed features are calculated. The mean is used to determine whether the exponential transformation leads to excessive amplification or compression of the overall feature values; the variance is used to measure whether the differences between pixels after the exponential transformation still have sufficient discriminative power. An ideal power exponent should maintain both a moderate mean level and a reasonable variance to avoid overly concentrated or divergent feature distributions.

[0048] Systematic testing of all candidate m values ​​reveals that the mean and variance distributions of the features are most ideal around m = 0.15: the mean remains stable, avoiding an overall overestimation due to a small exponent, and also avoiding compression to an excessively low range due to a large exponent; simultaneously, the variance remains moderate, effectively reflecting the differences between pixels. Therefore, m = 0.15 achieves the best balance between mean stability and variance preservation under the data conditions of this study, and was ultimately selected as the power exponent parameter in the ERVI construction.

[0049] It should be noted that the optimal value of the exponent m is somewhat data-dependent. The statistical characteristics after exponential transformation may change under different regions, different sensors, or different vegetation types. Therefore, the choice of m can be appropriately adjusted based on the mean-variance trend of the actual data. Thus, the final formula is: ERVI = [RVI * H * PV] 0.15 Or ERVI = [RVI * (1 - A) * PV] 0.15 (4).

[0050] Refer to the calculation of FVC: For the study area, Sentinel-2 Level-2A data from June 13, 2023, were selected. All selected data underwent preprocessing steps including radiometric calibration, geometric correction, atmospheric correction, and data resampling, ensuring high accuracy and facilitating subsequent NDVI extraction and analysis. After calculating NDVI, vegetation cover (FVC) was further retrieved based on the pixel-dichotomy model (PDM), such as... Figure 5 As shown. The formulas for calculating NDVI and FVC are as follows: (5); In the formula: NIR and RED are the reflectance values ​​of the near-infrared and infrared bands of the Landsat 8 image, respectively; (6); In the formula: NDVI veg It is the NDVI value of pure vegetation pixels; NDVI soil This refers to the NDVI value of a pure, unfiltered pixel. In this study, 5% and 95% confidence intervals were selected to determine the NDVI. veg and NDVI soil .

[0051] S5. Estimate grassland vegetation cover (FVC) using ERVI inversion.

[0052] In the final stage, the obtained ERVI is used as input to a linear regression model to invert FVC. Specifically, the obtained sample data is first divided into a training set and a test set in a 70%:30% ratio. The training set is used for fitting and calibrating the model parameters, while the test set is used to independently validate the model performance. The training set contains samples covering different growth states, vegetation heights, and coverage levels to ensure that the model can fully learn the statistical relationship between grassland vegetation scattering characteristics and FVC during training. The test set covers samples from the same region that were not used for training and is used to evaluate the model's generalization ability and prediction accuracy for unknown samples.

[0053] In the regression modeling process, the ERVI of the training set samples is linearly fitted with the corresponding FVC to obtain the regression coefficients and intercepts. Subsequently, the trained model is applied to the test set to predict its FVC value, and by comparing it with the reference FVC, the coefficient of determination (R²), root mean square error (RMSE), and other indicators are calculated to comprehensively evaluate the model's estimation accuracy, stability, and applicability.

[0054] It needs to be further explained that: 1) Polarimetric Coherence Matrix: The polarimetric coherence matrix is ​​a matrix composed of the complex scattering coefficients of each polarimetric channel of SAR data. It reflects the statistical scattering characteristics of the target and provides a foundation for subsequent polarimetric decomposition and grassland vegetation feature extraction. Its specific expression is as follows: (7); In the formula, the superscript * indicates complex conjugate; Represents the average of sets in time and space; The backscattering coefficients represent different polarization modes.

[0055] 2) Similarity Pixel Statistical Method: Traditional RVI construction process usually assumes that local areas can be regarded as statistically homogeneous units, but this assumption is often difficult to hold in grassland areas. For example, the intermingling of sparse grasslands and dense grasses, as well as the frequent transitions between different vegetation types and bare land, make grassland areas exhibit more significant heterogeneity.

[0056] Therefore, this paper introduces a statistical method based on similar pixels. In this method, pixels are no longer integrated solely based on fixed information from their geometrically neighboring regions. Instead, they are actively sought within a certain range to identify pixel groups exhibiting high similarity in scattering structure, serving as reference units for feature extraction. This method borrows the statistical strategy of NLM (Nonlocal Means). The mathematical expression is as follows: (8); In the formula: v(j) is a pixel in the image patch that is similar to pixel i; the weight w(i,j) is determined based on the similarity between pixel i and pixel j.

[0057] The similarity between pixel i and pixel j is calculated by the Gaussian weighted Euclidean distance d(i,j), which is expressed as: (9); In the formula, N i Let N represent a square neighborhood of fixed size centered at pixel i, and N represent the square neighborhood of fixed size centered at pixel i. j Let represent the similar neighborhood centered at pixel j; 'a' is the standard deviation of the Gaussian kernel, and 'a > 0'. The smaller the Gaussian weighted Euclidean distance, the greater the similarity in the image, and the greater the weight of the corresponding pixel. The weight is defined as: (10); In the formula, Z(i) is a normalization constant; the parameter h controls the decay rate of the function.

[0058] Subsequent methods have further evolved into matching based on the polarization coherence matrices of similar scattering characteristics within local regions, using adaptive weighted fusion to enhance expressive power. The specific formula is as follows: (11); In the formula, x0 is the current pixel position; T(x i ) is the polarization coherence matrix; ω(x0, x i ) is the weight.

[0059] (12); In the formula, f s (x0, x i ) represents spatial weights; f r (x0, x i ) is the radiation weight.

[0060] (13); In the formula, ||x0-x i ‖ is the distance between the center pixel and the search pixel; r s Parameters that control spatial decay.

[0061] (14); In the formula, Used to measure the distance between two matrices; r is the covariance matrix corresponding to the pixel position; r It is a parameter that controls radiation similarity.

[0062] For the HPD (Hermitian Positive Definite Matrix) of each pixel in polarimetric SAR data, the LE distance (Log-Euclidean Distance) can be used to measure matrix similarity: (15); by Figure 4 For example, taking pixel x0 as the center, several pixels with similar scattering characteristics, such as x1, x2, and x3, are searched within its neighborhood window. In this framework, the similarity between each pair of pixels is determined by the corresponding weight ω(x0, x...). i As represented by the formula, the weight depends on the similarity of the weight to the center pixel in terms of scattering characteristics. The selection of similar pixels is not limited to geometric proximity, but is based on the consistency of statistical properties (such as covariance matrix or polarization eigenvector), thereby achieving dual nonlocal information fusion in spatial and feature dimensions.

[0063] 3) The formula for the HTCD algorithm is as follows: (16); here This represents the result after the polarization coherence matrix has been rotated. This represents the total matrix of surface scattering and double-bounce scattering. The volume scattering matrix is ​​represented by the volume scattering model proposed by An et al. , , These represent the weighting coefficients for surface scattering, double-bounce scattering, and volume scattering, respectively. Depending on the dielectric constant of the surface and the angle of incidence, This represents the scattering phase of surface scattering.

[0064] when When the scattering power is at that time, it can be expressed as: , and (17); when When the scattering power is at that time, it can be expressed as: , and (18); Where Pv, Pd and Ps represent the volume scattering power, surface scattering power and secondary scattering power, respectively; φ and θ are different matrix rotation angles; λ1, λ2 and λ3 are the eigenvalues ​​of the matrix.

[0065] 4) The mean reflects the degree of compression of the overall amplitude. Variance measures the dispersion of the data. The specific formula is as follows: (19); (20); in, It is the sample variance. It is the sample mean; It is the i-th sample point; n is the total number of samples.

[0066] 5) Pearson Correlation Coefficient: The Pearson correlation coefficient reflects the degree of linear correlation between two random variables. This method assumes that the relationship between the variables follows a normal distribution, and under the premise that the variables satisfy linear characteristics, it can accurately characterize their changing trends. Let the Pearson correlation coefficient be r, and its calculation formula is as follows: (twenty one); in, and Let x and y represent the i-th sample values, respectively. and These are the means of the corresponding variables.

[0067] 5) Univariate Linear Regression Model: Based on the correlation analysis, a univariate linear regression model is further constructed to evaluate the predictive ability of ERVI for FVC. Its fitted linear regression equation is: (16); Where y represents the predicted FVC value, and x represents the input vegetation index. and These are the intercept and slope terms of the regression model, respectively. This is the error term.

[0068] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: 1. Minimize the impact of heterogeneous pixels. This invention uses a nonlocal mean algorithm to calculate the polarization coherence matrix and constructs ERVI in regions with similar scattering mechanisms, thereby reducing the confusion between pixels with different vegetation cover.

[0069] 2. Enhance sensitivity in low biomass areas. ERVI enhances the vegetation scattering response of RVI by selecting features extracted by polarization decomposition and performing product and power transformation.

[0070] 3. Multi-parameter enhanced formula ERVI; This invention proposes a more comprehensive enhanced ERVI formula. The combination of multiple parameters makes ERVI more applicable and accurate under different vegetation types, geographical regions, and biomass conditions.

[0071] 4. Experimental verification and optimization selection; This invention evaluates the correlation between different enhancement parameters (such as H, A, and volumetric scattering power) and grassland FVC by comparing with an optical reference FVC, and optimizes the selection of key parameters. In practical applications, by selecting parameters highly correlated with FVC for optimization, the high-precision estimation capability of ERVI is ensured.

[0072] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for retrieving grassland vegetation coverage based on polarimetric SAR data, characterized in that, The method comprises the following steps: S1: obtaining and preprocessing polarimetric SAR data; S2: obtaining a polarimetric coherence matrix by a similar pixel statistical method; S3: performing polarimetric decomposition on the polarimetric coherence matrix and calculating RVI; S4: modifying the RVI by using H / A parameters and performing polarimetric weighting on the modified RVI to obtain ERVI; S5: estimating grassland vegetation coverage FVC by ERVI inversion.

2. The grassland vegetation coverage inversion method according to claim 1, characterized in that: In the step S1, the polarimetric SAR image data is GF-3 full-polarimetric SAR data.

3. The grassland vegetation coverage inversion method according to claim 1, characterized in that: In the step S1, the preprocessing comprises radiation calibration, filtering, terrain correction and geocoding.

4. The grassland vegetation coverage inversion method according to claim 3, characterized in that: In the step S1, the geocoding is implemented by using SRTM DEM data to register the image and geographic coordinates.

5. The grassland vegetation coverage inversion method according to claim 1, characterized in that: In the step S2, the similar pixel statistical method comprises: S2.1: setting a fixed search window centered on each pixel; in the search window, a set of pixels similar to the central pixel in polarization characteristics is selected by calculating the similarity of the neighborhood pixels and the central pixel in polarization characteristics; S2.2: assigning a weight value according to the similarity and normalizing; S2.3: performing weighted average on the similar pixel set to generate a weighted polarimetric coherence matrix.

6. The grassland vegetation coverage inversion method according to claim 1, characterized in that: In the step S3, specifically: S3.1: using a polarimetric decomposition method to process the polarimetric coherence matrix to obtain surface scattering power PS, secondary scattering power PD and volume scattering power PV of each pixel; S3.2: calculating RVI in the following manner: RVI = PV / (PS + PD + PV).

7. The grassland vegetation coverage inversion method according to claim 6, characterized in that: In the step S3.1, the polarimetric decomposition method is HTCD polarimetric decomposition.

8. The grassland vegetation coverage inversion method according to claim 6, characterized in that: The step S4 comprises: S4.1: dividing the region into sub-regions and extracting H and A parameters of each sub-region by using the H / A / alpha polarimetric decomposition method; S4.2: calculating the correlation coefficient of H and A parameters in each sub-region with FVC by using NDVI-FVC calculated from the optical image as reference data; S4.3: selecting the parameter with higher absolute value of the correlation coefficient as the correction parameter of each sub-region in the H or A parameter; S4.4: introducing the H or A parameter to modify the RVI; S4.5: performing volume scattering polarimetric weighting on the modified RVI to obtain ERVI; specifically: ERVI = RVI*H*PV or ERVI = RVI*(1-A)*PV.

9. The grassland vegetation coverage inversion method according to claim 8, characterized in that: In the step S4.5, a power transformation is introduced to amplify ERVI in a nonlinear manner, specifically: ERVI = [RVI * H * PV] m or ERVI = [RVI * (1 - A) * PV] m .

Images