GNSS-R soil humidity inversion method and device, storage medium and computer equipment

Abstract

The invention discloses a GNSS-R soil humidity inversion method and device, a storage medium and computer equipment. Comprising the following steps: inputting a delay-Doppler graph obtained by detecting a target area by a satellite-borne GNSS-R into a feature extraction model corresponding to the target area, extracting a global dependency relationship of a feature sequence of the delay-Doppler graph through the feature extraction model, and outputting global scattering feature data of the target area based on the global dependency relationship; inputting the ground environment data of the target area and the observation parameters of the satellite-borne GNSS-R into a geographically weighted regression model corresponding to the target area, performing fitting calculation on the spatial correlation of the ground environment data and the observation parameters through the geographically weighted regression model, and outputting prior inversion data of the soil humidity of the target area; and performing nonlinear transformation and feature fusion on the global scattering feature data, the prior inversion data and the observation parameters based on a soil humidity model to obtain a soil humidity inversion result. Therefore, the adaptability and inversion accuracy of the model in different geographical environments are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing monitoring, and in particular to a GNSS-R soil moisture inversion method and device, a storage medium and a computer equipment. BACKGROUND

[0002] Under the background of global climate change, extreme climate events such as drought and flood occur frequently. High-precision soil moisture monitoring can provide key scientific basis for agricultural production, water resource management and disaster warning, and has great significance for guaranteeing food security and ecological security.

[0003] At present, soil moisture monitoring methods mainly include two types: direct measurement based on ground observation and indirect inversion using remote sensing technology. Although the ground observation method has high accuracy, it has limitations such as limited spatial coverage and high cost. In contrast, remote sensing inversion technology can realize large-scale and periodic soil moisture monitoring. The L-band signal source carried by the Global Navigation Satellite System-Reflectometry (GNSS-R) technology has strong penetration ability for surface vegetation, soil and snow, and the spatial resolution can reach 3km, which has significant advantages compared with traditional microwave remote sensing (such as SMAP and SMOS satellites). However, the spatial heterogeneity of surface parameters and the nonlinear characteristics of observation data make the existing GNSS-R soil moisture inversion methods have two key limitations: one is that the detailed information of the original delay-Doppler map (DDM) is not fully utilized, and the other is that the spatial heterogeneity of surface parameters is not effectively modeled. Therefore, the traditional inversion method is difficult to fully exploit the deep information of the data, especially under the demand of high-precision inversion. SUMMARY

[0004] Therefore, the present application provides a GNSS-R soil moisture inversion method, device, storage medium and computer equipment, which can overcome the local limitation of traditional convolutional neural network in feature extraction, and finely depict the spatial distribution pattern of soil moisture, thereby improving the adaptability and inversion accuracy of the model in different geographical environments.

[0005] According to one aspect of the present application, a GNSS-R soil moisture inversion method is provided, comprising:

[0006] inputting a delay-Doppler map obtained by a spaceborne GNSS-R probe target region into a feature extraction model corresponding to the target region, extracting global dependency relationships of a feature sequence of the delay-Doppler map through the feature extraction model, and outputting global scattering feature data of the target region based on the global dependency relationships;

[0007] input ground environment data of the target area and observation parameters of the spaceborne GNSS-R into a geographical weighted regression model corresponding to the target area, respectively fit and calculate spatial correlations of the ground environment data and the observation parameters through the geographical weighted regression model, and output prior inversion data of soil moisture in the target area;

[0008] perform nonlinear transformation and feature fusion on the global scattering feature data, the prior inversion data, and the observation parameters based on a soil moisture model, and obtain the soil moisture inversion result.

[0009] Optionally, the feature extraction model comprises a segmentation unit and a multi-layer encoder, each layer of the encoder comprising a multi-head self-attention sublayer and a multi-layer perceptron sublayer, the global dependence relationship of the feature sequence of the delay-Doppler map is extracted through the feature extraction model, and global scattering feature data of the target area is output based on the global dependence relationship, comprising:

[0010] the delay-Doppler map is segmented by the segmentation unit to obtain a pixel block sequence, and position encoding is added to the pixel block sequence to form the feature sequence;

[0011] the multi-head self-attention sublayer is used to capture the correlation between different pixel blocks in the feature sequence;

[0012] the correlation output by different self-attention sublayers is subjected to residual connection and layer normalization processing to determine the global dependence relationship;

[0013] the multi-layer perceptron sublayer is used to perform nonlinear transformation on the global dependence relationship, and the context feature data of the enhanced feature sequence;

[0014] the context features output by different perceptron sublayers are subjected to residual connection and layer normalization processing to determine the global scattering feature data.

[0015] Optionally, the fitting calculation of the spatial correlations of the ground environment data and the observation parameters of the spaceborne GNSS-R through the geographical weighted regression model comprises:

[0016] the spatial coordinates of the sampling points corresponding to the ground environment data and the observation parameters are used to determine the spatial weights of the sampling points through a kernel function of the geographical weighted regression model;

[0017] the spatial weights are fitted and calculated based on a weighted least squares method to determine regression coefficients of the ground environment data and the observation parameters;

[0018] the prior inversion data is determined based on the regression coefficients.

[0019] Optionally, the geographically weighted regression model is represented as:

[0020]

[0021] where (u i ,v i ) is the spatial coordinate of the i-th sampling point, β j (u i ,v i ) is the j-th independent variable regression coefficient of the sampling point i, ε i is the error term of the sampling point i, and ε i ~N(0,σ 2 ), Cov(ε i ,ε j )=0(i≠j).

[0022] Optionally, the GNSS-R soil moisture inversion method further comprises:

[0023] initializing the kernel function of the geographically weighted regression model according to the density of the sample points in the target area;

[0024] inputting the ground environment sample data and the sample land surface reflectance into the geographically weighted regression model to determine the theoretical soil moisture;

[0025] adjusting the bandwidth and the initial parameters of the kernel function based on the residual error of the theoretical soil moisture and the actual soil moisture associated with the sample points.

[0026] Optionally, the observation parameters include the land surface reflectance, the leading edge slope, the trailing edge slope, and the incident angle.

[0027] The land surface reflectance is determined by using the following calculation method:

[0028]

[0029] where Γ is the land surface reflectance, P r is the uncorrected peak value of the delay-Doppler diagram, P t is the transmitted right-handed circular polarization power, G t is the gain of the transmitting antenna, R ts is the distance between the transmitter and the specular reflection point, R sr is the distance between the specular reflection point and the receiver, G r is the gain of the receiving antenna, and λ is the radar wavelength.

[0030] Optionally, the GNSS-R soil moisture inversion method further comprises:

[0031] obtaining the area of the region to be studied;

[0032] if the area of the region to be researched is greater than a preset area, the region to be researched is divided into a plurality of target regions based on the preset area;

[0033] if the area of the region to be researched is less than or equal to a preset area, the region to be researched is taken as the target region.

[0034] According to another aspect of the present application, a GNSS-R soil moisture inversion device is provided, comprising:

[0035] a feature extraction module configured to input a delay-Doppler map obtained by a spaceborne GNSS-R probe target region into a feature extraction model corresponding to the target region, extract global dependency of a feature sequence of the delay-Doppler map through the feature extraction model, and output global scattering feature data of the target region based on the global dependency;

[0036] a preprocessing module configured to input ground environment data of the target region and observation parameters of the spaceborne GNSS-R into a geographically weighted regression model corresponding to the target region, respectively fit and calculate spatial correlation of the ground environment data and the observation parameters through the geographically weighted regression model, and output prior inversion data of soil moisture of the target region;

[0037] an inversion module configured to perform nonlinear transformation and feature fusion on the global scattering feature data, the prior inversion data and the observation parameters based on a soil moisture model, and obtain the soil moisture inversion result.

[0038] Optionally, the feature extraction model comprises a segmentation unit and a multi-layer encoder, each layer of the encoder comprises a multi-head self-attention sublayer and a multi-layer perceptron sublayer, and the feature extraction module is specifically configured to perform segmentation processing on the delay-Doppler map through the segmentation unit to obtain a pixel block sequence, add position encoding to the pixel block sequence to form the feature sequence, capture the correlation between different pixel blocks in the feature sequence through the multi-head self-attention sublayer, determine the global dependency by performing residual connection and layer normalization processing on the correlation output by different self-attention sublayers, perform nonlinear transformation on the global dependency through the multi-layer perceptron sublayer to obtain enhanced context feature data of the feature sequence, and determine the global scattering feature data by performing residual connection and layer normalization processing on the context features output by different perceptron sublayers.

[0039] Optionally, the preprocessing module is specifically configured to determine, by a kernel function of the geographically weighted regression model, a spatial weight of a sampling point corresponding to the ground environment data and the observation parameter based on spatial coordinates of the sampling point; determine a regression coefficient of the ground environment data and the observation parameter by fitting calculation on the spatial weight based on a weighted least square method; and determine the prior inversion data based on the regression coefficient.

[0040] Optionally, the geographically weighted regression model is represented as:

[0041]

[0042] wherein, (u i ,v i ) is spatial coordinates of the i th sampling point, β j (u i ,v i ) is a j th independent variable regression coefficient of the sampling point i, ε i is an error term of the sampling point i, and ε i ~N(0,σ 2 ), Cov(ε i ,ε j )=0(i≠j).

[0043] Optionally, the GNSS-R soil moisture inversion device further comprises:

[0044] an initialization module configured to initialize a kernel function of the geographically weighted regression model according to a density of sample points in a target area;

[0045] an optimization module configured to input ground environment sample data and sample land surface reflectivity into the geographically weighted regression model to determine a theoretical soil moisture, and adjust a bandwidth and an initial parameter of the kernel function based on a residual error of the theoretical soil moisture and actual soil moisture associated with the sample points.

[0046] Optionally, the observation parameter comprises land surface reflectivity, leading edge slope, trailing edge slope and incident angle.

[0047] The land surface reflectivity is determined by using the following calculation method:

[0048]

[0049] wherein, Γ is the land surface reflectivity, P r is an uncorrected peak value of a delay-Doppler diagram, P t is transmitted right-hand circular polarization power, G t is a gain of a transmitting antenna, R ts is a distance between a transmitter and a specular reflection point, and R srG is the distance between the specular reflection point and the receiver, G r G is the gain of the receiving antenna, and λ is the radar wavelength.

[0050] Optionally, the GNSS-R soil moisture inversion device further comprises:

[0051] The partition module is configured to acquire an area of a region to be researched; if the area of the region to be researched is greater than a preset area, the region to be researched is divided into a plurality of target regions based on the preset area; and if the area of the region to be researched is less than or equal to the preset area, the region to be researched is taken as the target region.

[0052] According to another aspect of the present application, a readable storage medium is provided, which stores a program or instructions, and the program or instructions are executed by a processor to implement the steps of the GNSS-R soil moisture inversion method.

[0053] According to still another aspect of the present application, a computer device is provided, which comprises a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements the steps of the GNSS-R soil moisture inversion method when executing the program.

[0054] By means of the above technical solution, the dependence relationship of the global feature sequence in the delay-Doppler map (DDM) of the target region is deeply mined by the feature extraction model corresponding to the target region, the global scattering feature data in the DDM can be completely captured in fine granularity, and the DDM feature data can more comprehensively reflect the overall law of mutual influence between different positions of the ground scattering. The observation parameters of the GNSS-R are analyzed by means of the DDM, which are then input into the geographic weighted regression (GWR) model together with the ground environment data corresponding to the target region. The GWR model is used to adaptively fit the spatial heterogeneity of the data for different spatial positions, more truly reflects the spatial variation relationship between the ground environment and reflection characteristics and the soil moisture, and generates prior inversion data that can reflect the distribution difference of the soil moisture at different spatial positions, thereby making up for the lack of constraints on the spatial continuity of the DDM. The global scattering feature data, the prior inversion data, and the observation parameters of the GNSS-R are nonlinearly transformed and fused by means of the soil moisture model, multi-source information is cooperated, the complex correlation between the composite information is fully mined, the model can spontaneously learn how to offset or correct the influence of the vegetation and the roughness, and thus the model has stronger robustness under the complex underlying surface conditions and enhanced performance and robustness of the soil moisture inversion.

[0055] The above description is only a summary of the technical solutions of the present application. In order to more clearly understand the technical means of the present application, the specific embodiments of the present application can be implemented in accordance with the content of the description, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described. Attached Figure Description

[0056] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0057] Figure 1 A schematic flowchart of the GNSS-R soil moisture inversion method provided in this application embodiment is shown;

[0058] Figure 2 The delayed-Doppler graphs in the embodiments of this application are shown;

[0059] Figure 3 The unbiased root mean square error distribution of soil moisture inversion results of each model in the embodiments of this application is shown in the figure.

[0060] Figure 4 The spatial distribution map showing the correlation between the soil moisture inversion results of each model in the embodiments of this application and the SMAP soil moisture is shown.

[0061] Figure 5 This paper shows a scatter plot comparing the soil moisture inversion results of each model in the embodiments of this application with the SMAP soil moisture data.

[0062] Figure 6 This paper presents a schematic diagram comparing the sample size of different land cover types and the corresponding evaluation indicators of different CYGNSS soil moisture products in embodiments of this application.

[0063] Figure 7 The spatial distribution of correlation of soil moisture inversion results for different seasons in the embodiments of this application is shown;

[0064] Figure 8 This illustration shows the spatial distribution characteristics of the downscaled soil moisture inversion results in an embodiment of this application.

[0065] Figure 9 This paper presents a schematic diagram showing the comparative analysis of measured soil moisture and soil humidity inversion results at the ISMN ground observation station at a 3km scale in the embodiments of this application.

[0066] Figure 10 A structural block diagram of the GNSS-R soil moisture inversion device provided in an embodiment of this application is shown;

[0067] Figure 11 A schematic diagram of the electronic structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0068] Hereinafter, the application will be described in detail with reference to the drawings and examples. It should be noted that the examples in the application and the features in the examples can be combined with each other without conflict.

[0069] The embodiments of the application are described in detail below with reference to the accompanying drawings, in which the same or similar elements are denoted by the same or similar reference numerals throughout. The embodiments described below by reference to the accompanying drawings are exemplary and are only used to explain the application and cannot be interpreted as a limitation on the application.

[0070] Those skilled in the art can understand that, unless specifically stated, the singular form "a", "an", "said" and "the" used herein can also include the plural form. It should be further understood that the phrase "comprising" used in the specification of the application means that the features, integers, steps, operations, elements and / or components exist, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we say that an element is "connected" or "joined" to another element, it can be directly connected or joined to the other element, or there can be intermediate elements. In addition, "connected" or "joined" used herein can include wireless connection or wireless connection. The phrase "and / or" used herein includes all or any single unit and all combinations of the associated listed items.

[0071] Now, exemplary embodiments according to the present application will be described in more detail with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in various different forms, and should not be interpreted as being limited to only the embodiments set forth herein. It should be understood that these embodiments are provided so that the disclosure of the present application is thorough and complete, and the concepts of these exemplary embodiments are fully conveyed to those skilled in the art.

[0072] In the present embodiment, a GNSS-R soil moisture inversion method is provided, as shown in the figure, the method comprises: Figure 1

[0073] Step 101, input the delay-Doppler map obtained by the spaceborne GNSS-R detection target area into the feature extraction model corresponding to the target area, extract the global dependence relationship of the feature sequence of the delay-Doppler map through the feature extraction model, and output the global scattering feature data of the target area based on the global dependence relationship.

[0074] ​In this embodiment, Level 1 data collected by GNSS-R is used, which can be downloaded from the satellite official website. The L1 data stores time-delay-Doppler map (DDM), surface reflectivity, incident angle, front edge slope, tail edge slope, antenna gain and other related information in NetCDF format. The delay-Doppler map (DDM) contains bistatic radar cross section (BRCS), effective scattering area and simulated power information.

[0075] Specifically, the feature extraction model is established based on a Vision Transformer (ViT) network architecture. The feature extraction model is used to process an image into sequence data, and uses a multi-layer encoder (Encoder) of the Transformer to process the sequence, and finally completes feature extraction and other tasks.

[0076] In this embodiment, the spatio-temporal joint information contained in all pixel points of the entire delay-Doppler map (DDM) can effectively capture the mutual influence and spatial correlation between different positions, rather than only a few manually selected feature points. Compared with traditional GNSS-R soil moisture inversion methods relying on CNN models, while learning and identifying information in the DDM, the global correlation between pixels is strengthened, effectively reducing the inversion deviation caused by local noise and signal fluctuations, and greatly improving the expression ability and accuracy of global scattering features.

[0077] In an actual application scenario, the feature extraction model includes a segmentation unit and a multi-layer encoder, and each layer of the encoder includes a multi-head self-attention sublayer and a multi-layer perceptron sublayer. Then step 101 specifically includes the following steps:

[0078] Step 101-1, the delay-Doppler map is segmented by the segmentation unit to obtain a pixel block sequence, and position encoding is added to the pixel block sequence to form a feature sequence.

[0079] In this embodiment, the DDM is split into multiple smaller pixel blocks, which helps to capture the detailed features of the target or ground object in the local space, so as to accurately distinguish the scattering features of different physical phenomena. And add position encoding to each pixel block to distinguish the spatial position of each pixel block in the DDM, realize the modeling of spatial dependence relationship, and prevent the feature sequence from only relying on the content of the pixel block itself and ignoring its specific spatial distribution.

[0080] Step 101-2, the multi-head self-attention sublayer is used to capture the correlation between different pixel blocks in the feature sequence.

[0081] Step 101-3, residual connection and layer normalization processing are performed on the correlation output by the different self-attention sublayers to determine the global dependence relationship.

[0082] It can be understood that the global dependency is used to describe the correlation statistical information between the pixel blocks in the feature sequence.

[0083] In step 101-4, the global dependency is nonlinearly transformed by the multilayer perceptron sublayer, and the context feature data of the enhanced feature sequence.

[0084] In step 101-5, the context features output by different perceptron sublayers are connected in residual and normalized, and the global scattering feature data is determined.

[0085] In this embodiment, the multi-head attention sublayer can simultaneously calculate the attention weight (correlation) of any two pixel blocks from multiple different perspectives, and through residual connection and normalization, the features output by each sublayer are added to each other and then mean-variance standardized to form the global dependency. Thus, when the model is stacked with multiple layers of sublayers, it can still learn effective features stably, rather than causing feature degradation due to gradient decay, thereby improving the balance of feature expression. Further, the multilayer perceptron is used to convert the linear correlation information into more complex nonlinear feature expression, thereby characterizing the nonlinear relationship and obtaining the context feature data. Similarly, the context feature data output by each sublayer is integrated by using residual connection and normalization to obtain the global scattering feature containing rich correlation relationships, thereby providing high-quality feature input for subsequent core tasks such as soil moisture inversion.

[0086] For example, the bistatic radar cross section (BRCS), effective scattering area, and simulated power are taken as inputs of the ViT feature extraction model. The ViT feature extraction model performs pixel-by-pixel segmentation on the DDM two-dimensional image (17x11 pixels), rather than dividing it into non-overlapping small blocks to form a series of pixel blocks. After tiling, a classification token (ClassToken) is added as a learnable embedding vector to form a pixel sequence.

[0087]

[0088] In the formula, x p is the pixel sequence; N = H x W is the number of pixels; x c is the classification token, which is randomly initialized and connected to the front end of the pixel sequence, and can be fine-tuned during the training process and used as an auxiliary vector to aggregate global features from each DDM pixel.

[0089] The position information is added to the pixel sequence to form a feature sequence for the final input of the Transformer encoder layer, which can be represented as:

[0090] x o = x p + x pos ;

[0091] where x o is a feature sequence; x pos is a position encoding, which can be updated during network training.

[0092] The feature sequence is input into multiple Transformer encoder layers. Each Transformer encoder layer contains a multi-head self-attention (MSA) and a multi-layer perceptron (MLP) sublayer. Finally, the output representation of the encoder layer is normalized by a LN layer, and a flattened MLP head is used to generate features extracted from the DDM.

[0093] In an embodiment, before step 101, the GNSS-R soil moisture inversion method further comprises: obtaining the area of the region to be studied; if the area of the region to be studied is greater than a preset area, dividing the region to be studied into multiple target regions based on the preset area; if the area of the region to be studied is less than or equal to the preset area, taking the region to be studied as a target region.

[0094] For example, the preset area can be 36x36 kilometers, that is, the study area is divided into 36x36 kilometer basic modeling units. Each grid unit must contain at least 300 valid samples before model training, and areas with insufficient samples will be excluded. The data in each 36 kilometer grid unit is used to train an independent deep learning model, focusing on capturing the mapping relationship between the DDM features and soil moisture specific to the region. Within the 36 kilometer grid unit, all 9x9 kilometer sub-grid units use the model trained in the unit for prediction.

[0095] In this embodiment, a large area is divided into multiple smaller target regions, and feature extraction and geographical weighted regression are performed based on each target region. Not only can the mapping relationship between the DDM features and soil moisture specific to the region be captured, but the soil moisture prediction accuracy can also be improved. Moreover, multiple target regions can be independently or in parallel processed, supporting distributed computing to speed up the overall processing speed, balancing computing efficiency and inversion accuracy.

[0096] Step 102, input the ground environment data and the observation parameters of the spaceborne GNSS-R of the target region into the geographical weighted regression model corresponding to the target region, and fit and calculate the spatial correlation of the ground environment data and the observation parameters through the geographical weighted regression model, to output the prior inversion data of the soil moisture of the target region.

[0097] The ground environment data can be collected by the SMAP satellite. The ground environment data includes, but is not limited to, temperature, soil type, vegetation coverage, terrain characteristics and other environmental variables. The observation parameters include surface reflectivity (SR), leading edge slope (LES), trailing edge slope (TES) and incident angle obtained by the CYGNSS satellite task.

[0098] Specifically, the surface reflectivity is determined by using the following calculation method:

[0099]

[0100] wherein, Γ is the surface reflectivity, P r is the uncorrected peak value of the delay-Doppler map, P t is the transmitted right-handed circular polarization power, G t is the gain of the transmitting antenna, R ts is the distance between the transmitter and the specular reflection point, R sr is the distance between the specular reflection point and the receiver, G r is the gain of the receiving antenna, and λ is the radar wavelength.

[0101] In this embodiment, a local regression model is independently established for each pixel block by using a geographic weighted regression (GWR) model, and the ground environment measured data and the GNSS-R based observation parameters are used to complement and quantify the relationship between the soil moisture and the environmental factors and the observation parameters, and the change of the relationship with the geographical position. Thus, the spatial distribution pattern of the soil moisture can be finely described, and the output prior inversion data is more in line with the actual soil moisture initial field with rich spatial details, and the misjudgment caused by the environmental background differences such as terrain and land use is effectively alleviated.

[0102] In the actual application scenario, step 102 specifically includes the following steps:

[0103] Step 102-1, by using the kernel function of the geographic weighted regression model, the spatial weight of the sampling point is determined based on the spatial coordinates of the sampling point corresponding to the ground environment data and the observation parameters.

[0104] The sampling point can be the collection position of the ground environment data and the observation parameters with a clear spatial coordinate (such as latitude and longitude) in the target region, or can be a pixel block after the delay-Doppler map is split.

[0105] Specifically, the kernel function can be a Gaussian kernel function or an exponential kernel function, which can be selected according to the actual scene. For example, the Gaussian kernel is suitable for smooth transition of spatial distribution, and the exponential kernel is suitable for local influence of rapid decay.

[0106] It should be noted that the closer the sampling points, the higher the spatial weight, and the farther the weight is smaller. The higher the spatial weight, the more likely the ground environment data and the observation parameters are similar.

[0107] Step 102-2, based on the weighted least squares method, the spatial weight is calculated, and the regression coefficient of the ground environment data and the observation parameters is determined.

[0108] Step 102-3, based on the regression coefficient, the prior inversion data is determined.

[0109] In this embodiment, the spatial correlation of the adjacent area is accurately captured by the kernel function, and the local features are obtained. For different positions, the soil moisture is taken as the target variable, and the ground environment data and the observation parameters are taken as the covariates. The weighted least squares method with different weights is used for fitting, and an exclusive optimal regression coefficient suitable for the surrounding environment of each to-be-inverted point is generated, which effectively reduces the negative influence of outliers and measurement errors on local fitting parameters, and quantifies the difference in the influence strength of environmental factors and observation parameters in different sub-regions on soil moisture. The final output of the prior inversion data can finely reflect the spatial heterogeneity, and overcome the problem that the traditional global regression model cannot cope with the spatial non-stationarity of geographical phenomena.

[0110] Further, the geographical weighted regression model is represented as:

[0111]

[0112] In the formula, y i is the prior inversion data (soil volume water content, unit: cm 3 / cm 3 ) of the i-th sampling point, (u i , v i ) is the spatial coordinates of the i-th sampling point, β j (u i , v i ) is the j-th independent variable regression coefficient of the sampling point i, x j is the j-th independent variable (ground environment data and observation parameters), ε i is the error term of the sampling point i, and ε i ~N(0, σ 2 ), Cov(ε i , ε j ) = 0 (i ≠ j). β j (u i , v i ) is evaluated using local weighted least squares method.

[0113] In an embodiment, before step 102, the GNSS-R soil moisture inversion method further comprises: initializing a kernel function of the geographically weighted regression model according to the density of sample points in the target area; inputting ground environment sample data and sample surface reflectivity into the geographically weighted regression model to determine theoretical soil moisture; and adjusting the bandwidth of the kernel function based on the residual error of the theoretical soil moisture and the actual soil moisture associated with the sample points.

[0114] wherein the initial parameters include a distance decay function of the kernel.

[0115] In this embodiment, the spatial weight function is optimized by an adaptive bandwidth selection algorithm, the fitting deviation caused by uneven sample distribution is alleviated, the accuracy and stability of local regression are improved, and high-precision prior inversion data is generated. This process not only preserves the spatial non-stationary characteristics of environmental variables, but also constructs a feature vector with clear geographical significance by calculating local regression coefficients.

[0116] Step 103, performing nonlinear transformation and feature fusion on the global scattering feature data, the prior inversion data and the observation parameters based on the soil moisture model to obtain the soil moisture inversion result.

[0117] The GNSS-R soil moisture inversion method provided in the embodiments of the present application can deeply mine the dependency relationship of the global feature sequence in the delay-Doppler map (DDM) of the target area through the feature extraction model corresponding to the target area, can completely capture the fine-grained global scattering feature data in the DDM, and can make the DDM feature data more comprehensively reflect the overall law of mutual influence between different positions of surface scattering. Meanwhile, the observation parameters of GNSS-R are analyzed from the DDM, and then the observation parameters and the ground environment data corresponding to the target area are jointly input into a geographically weighted regression (GWR) model. The GWR model is used to adaptively fit the spatial heterogeneity of data for different spatial positions, to more truly reflect the spatial variation relationship between the surface environment and reflection characteristics and the soil moisture, to generate prior inversion data that can reflect the distribution difference of soil moisture at different spatial positions, and to make up for the lack of spatial continuity constraint of the DDM. Nonlinear transformation and feature fusion are performed on the global scattering feature data, the prior inversion data and the observation parameters of GNSS-R by using the soil moisture model, multi-source information is cooperated, the complex correlation between composite information is fully mined, the model can spontaneously learn how to offset or correct the influence of vegetation and roughness, and thus the model has stronger robustness under complex conditions, and the performance ability and stability of soil moisture inversion are enhanced.

[0118] The observation parameters of GNSS-R are analyzed from the DDM, and then the observation parameters and the ground environment data corresponding to the target area are jointly input into a geographically weighted regression (GWR) model. The GWR model is used to adaptively fit the spatial heterogeneity of data for different spatial positions, to more truly reflect the spatial variation relationship between the surface environment and reflection characteristics and the soil moisture, to generate prior inversion data that can reflect the distribution difference of soil moisture at different spatial positions, and to make up for the lack of spatial continuity constraint of the DDM. Nonlinear transformation and feature fusion are performed on the global scattering feature data, the prior inversion data and the observation parameters of GNSS-R by using the soil moisture model, multi-source information is cooperated, the complex correlation between composite information is fully mined, the model can spontaneously learn how to offset or correct the influence of vegetation and roughness, and thus the model has stronger robustness under complex conditions, and the performance ability and stability of soil moisture inversion are enhanced.

[0119] It can be understood that the measured soil moisture is taken as a supervision signal to define a loss function of the full connection neural network model. Historical global scattering feature data, historical prior inversion data and historical observation parameters are input into the full connection neural network model for training; the loss function is used to calculate the difference between the predicted value and the actual supervision signal; the gradient is calculated by using the loss value, and is propagated from the output layer to the input layer to obtain back propagation data; based on the back propagation data, the model network structure and the hyperparameters are continuously updated by using the gradient update rule until the iteration ends, and a soil moisture model with high precision mapping is obtained.

[0120] It can be understood that the activation function of the soil moisture inversion model is ReLU, LeakyReLU, Tanh and the like.

[0121] It should be noted that the size of the serial number of each step in the above embodiment does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0122] Specifically, GNSS-R data from 2018 to 2022 is taken as input to invert soil moisture data from 2023 to 2024. The experimental soil moisture data comes from CYGNSS data labeled SMAP soil moisture data, with a time resolution of 1 day and a spatial resolution of 9 km.

[0123] Step 1, obtain DelayDoppler Maps (DDMs) as shown in Figure 2 DDMs are two-dimensional scattering characteristic maps generated by measuring the arrival time (Delay) and Doppler shift (Doppler Shift) of GPS signals reflected on the earth's surface, including bistatic radar cross section (BRCS), effective scattering area and simulated power. Each pixel value reflects the scattering intensity of the GPS signal on the ground surface, and its scattering characteristics are closely related to the physical properties of the ground surface (such as soil moisture, vegetation coverage, etc.).

[0124] In order to make more full use of the delay Doppler map (DDM) information, CYGNSS other variables are integrated to build a complete feature system for the specular reflection point. According to the assumption that the land surface usually shows coherent scattering caused by specular reflection. Based on the bistatic radar equation, the effective surface reflectivity (Γ) is calculated using the peak value of the calibrated DDM. The calculation method is as follows:

[0125]

[0126] In the formula, P r is the uncorrected peak value of each DDM product, P t is the transmitted right-handed circularly polarized power, G tG is the gain of the transmitting antenna, R ts G is the gain of the transmitting antenna, R sr G is the gain of the transmitting antenna, R r G is the gain of the transmitting antenna, R is the GPS wavelength (0.19 meters).

[0127] In addition, the CYGNSS Tail Edge Slope (TES) is considered as the slope of the reflectivity delay waveform, which can be input into the SM retrieval and the reflectivity delay waveform value at m+3 as an indicator of coherence / incoherence. In addition to the DDMs, the surface reflectivity, TES, incidence angle are input to the subsequent deep learning.

[0128] Step 2, select data from August 10, 2018 to December 31, 2022 for training, 80% of the data is randomly selected to form a training set, and the remaining data is used for testing. Select January 1, 2023 to August 10, 2024 as the evaluation period. The core input is a 17x11x3 three-dimensional tensor composed of three normalized delay Doppler maps (DDMs).

[0129] Step 3, four commonly used machine methods are selected as baseline models respectively, including random forest regression model (RF), convolutional neural network model (CNN), Vision Transformer (ViT) and geographic weighted regression model (GWR), four baseline models and the GNSS-R soil moisture retrieval (ViT-GWR) model proposed in the application are trained on the processed training data set, and each model is updated and optimized through the training set to learn the features and patterns of the samples; then evaluate each trained model through the validation set, adjust the hyperparameters and network structure of the model according to the performance of the model on the validation set; finally, test the three baseline models and the soil moisture model proposed in the embodiment on the test set, and evaluate the prediction accuracy of all samples.

[0130] In the hourly precipitation prediction experiment of each model, bias (BIAS), root mean square error (RMSE), unbiased root mean square error (ubRMSE), and Pearson correlation coefficient (Correlation) are used to evaluate the hourly precipitation prediction accuracy of each model. The four indicators selected are commonly used to evaluate precipitation prediction performance. Ideally, according to, the developed model should perform better than the basic prediction method on these indicators. In order to more intuitively evaluate the precipitation prediction, the absolute precipitation error is calculated to evaluate the predicted precipitation. The calculation methods of BIAS, RMSE, ubRMSE, Correlation are as follows

[0131]

[0132] where y i and are the measured and model predicted precipitation values, respectively, and are the mean values of the measured and model predicted precipitation samples, respectively. The closer the Correlation value is to 1, the better the model’s prediction. The closer the BIAS, ubRMSE, and RMSE values are to 0, the better the model’s prediction.

[0133] The statistical evaluation of each model is shown in Table 1.

[0134] Table 1

[0135] BIAS RMSE ubRMSE R ViT-GWR 0.0050 0.0486 0.0483 0.87 ViT 0.0044 0.0509 0.0507 0.86 CNN 0.0039 0.0543 0.0541 0.84 RF 0.0014 0.0503 0.0503 0.8611 GWR 0.0060 0.0595 0.0592 0.81

[0136] The results in Table 1 show that the ViT-GWR model provided in the present application provides the lowest ubRMSD (0.0486 cm 3 / cm 3 ) and the highest correlation coefficient (0.87). In other words, the ViT in the ViT-GWR model can effectively utilize the information in the delay-Doppler map (DDM), and the prior features obtained by the GWR can also enable the DL model to better perceive the changes in the DDM, and the two complement each other, which significantly improves the estimation accuracy of soil moisture compared to the ViT or GWR alone.

[0137] Further, Figure 3 shows the spatial distribution of the unbiased root mean square error (ubRMSE, unit m 3 / m 3 ) of the inversion results of each model and the 9-kilometer grid of the SMAP soil moisture product in the CONUS region. The color from blue to red represents the error from small to large. The unbiased RMSE measures the size of the model prediction error, and the smaller the value, the better the performance. As Figure 3 shown, the ViT-GWR model provided in the embodiments of the present application corresponds to more blue areas, i.e., smaller ubRMSE, which indicates that the soil moisture inversion accuracy and spatial adaptability are improved by the combination of ViT and spatial weighted regression.

[0138] Figure 4 shows the spatial distribution of the correlation coefficient (R) of the inversion results of each model and the 9-kilometer grid of the SMAP soil moisture product in the CONUS region. The numerical range is 0-1, and the color from blue to red represents the correlation from low to high. The correlation coefficient reflects the linear correlation between the model prediction results and the actual observation data, and the closer the value is to 1, the better the model performs. As Figure 4 shown, the spatial correlation (R) of the ViT-GWR model provided in the embodiments of the present application is relatively balanced, which can better adapt to different characteristics between regions.

[0139] As Figure 5 shown, the overall results of soil moisture have good correlation with the results of soil moisture sampled by SMAP satellite, and the predicted values of each model are mostly symmetrically distributed along the 1:1 line. The ViT-GWR provided in the embodiments of the present application has some scattered data points, but most of the data points are distributed on the 1:1 line, showing smaller dispersion and less deviation compared with the inversion results of other models.

[0140] Figure 6 The number of samples of different land cover types (Forests: forest, Shrublands: shrub, Woody: forest area, Savannas: savanna, Grasslands: grassland, Croplands: farmland, Barren: barren land) and the corresponding evaluation indexes (correlation coefficient R, unbiased root mean square error ubRMSE, root mean square error RMSE, bias BIAS) of different CYGNSS soil moisture products are shown. As Figure 6 shown, the ViT-GWR (blue) provided in the embodiments of the present application shows the highest correlation in most land cover types and provides the smallest unbiased root mean square error (especially stable performance on forest, grassland and farmland types), indicating that the fitting effect is better than that of other models. In terms of root mean square error, the ViT-GWR provided in the embodiments of the present application and the ViT show relatively good performance, lower error and smaller change range.

[0141] Figure 7 The map of the spatial distribution of the correlation data obtained by the ViT-GWR method provided in the embodiments of the present application is shown. As Figure 7 shown, from the spatial pattern, although there are these spatial and seasonal differences, the model shows stable performance as a whole, and the soil moisture inversion error of most regions is controlled within a reasonable range of ±0.05cm 3 / cm 3 .

[0142] Table 2 shows the verification statistics of soil moisture inversion results of different models and soil moisture measurement values of the International Soil Observation Network (ISMN). All evaluation indexes are calculated separately at each independent site, and finally the average and median statistical results of all sites are reported. It can be seen that the ViT-GWR method provided in the embodiments of the present application has better prediction performance.

[0143] Table 2

[0144]

[0145] Figure 8The spatial distribution characteristics of the downscaling soil moisture product are shown, the daily average value of the ViT-GWR inversion result is synthesized, and the three kinds of soil moisture products are compared and analyzed: 9 km SMAP benchmark product (SMAP9km), 9 km ViT-GWR product (ViT-GWR9km) and 3 km downscaling ViT-GWR product (ViT-GWR3km). The results show that ViT-GWR9km and SMAP9km show a highly consistent spatial distribution pattern, verifying the reliability of the basic inversion algorithm; more importantly, the ViT-GWR3km product obtained by downscaling can maintain the spatial distribution characteristics of the original product while improving the spatial resolution from 9 km to 3 km, and does not introduce significant spatial distortion. The 3 km high-resolution product makes full use of the advantages of CYGNSS dense observation, and can more finely depict the spatial heterogeneity of soil moisture. While the 9 km medium-resolution product improves the spatial coverage of the product while maintaining data quality through an adaptive spatial aggregation algorithm. This multi-scale verification not only confirms the advantages of the ViT-GWR framework in maintaining spatial fidelity, but also provides a reliable technical approach for generating high-resolution soil moisture products.

[0146] Figure 9 The comparison between the ISMN ground observation site measured soil moisture and the ViT-GWR model inversion result at the scale of 3 km is shown, and it can be seen that the correlation (R=0.61) between ViT-GWR and ISMN ground observation data is still high.

[0147] In summary, the application innovatively integrates visual Transformer (ViT) and geographically weighted regression (GWR) based on the traditional CNN framework, and proposes a ViT-GWR hybrid model to improve the GNSS-R soil moisture inversion accuracy. To verify the model performance, the ViT-GWR is compared with the mainstream methods (CNN, Random Forest), achieving the highest accuracy. To analyze the source of the model advantage, the ablation experiment proves that ViT-GWR is significantly better than using ViT architecture alone and using GWR module alone (GWR-only), indicating that the synergistic effect of global feature modeling and spatial heterogeneity processing is the key to the accuracy breakthrough.

[0148] Further, as shown in Figure 10 As a specific implementation of the above GNSS-R soil moisture inversion method, the embodiments of the application provide a GNSS-R soil moisture inversion device 300, which comprises a feature extraction module 301, a preprocessing module 302 and an inversion module 303.

[0149] The feature extraction module 301 is configured to input the delay-Doppler map obtained by the spaceborne GNSS-R detection target region into a feature extraction model corresponding to the target region, extract the global dependence relationship of the feature sequence of the delay-Doppler map through the feature extraction model, and output the global scattering feature data of the target region based on the global dependence relationship.

[0150] The preprocessing module 302 is configured to input the ground environment data and the spaceborne GNSS-R observation parameter of the target region into a geographical weighted regression model corresponding to the target region, fit and calculate the spatial correlation of the ground environment data and the observation parameter through the geographical weighted regression model, and output the prior inversion data of the soil moisture of the target region.

[0151] The inversion module 303 is configured to perform nonlinear transformation and feature fusion on the global scattering feature data, the prior inversion data, and the observation parameter based on a soil moisture model, and obtain the soil moisture inversion result.

[0152] Further, the feature extraction model includes a segmentation unit and a multi-layer encoder, each layer of the encoder includes a multi-head self-attention sublayer and a multi-layer perceptron sublayer, and the feature extraction module 301 is specifically configured to perform segmentation processing on the delay-Doppler map through the segmentation unit to obtain a pixel block sequence, add position encoding to the pixel block sequence to form a feature sequence, capture the correlation between different pixel blocks in the feature sequence through the multi-head self-attention sublayer, determine the global dependence relationship through residual connection and layer normalization processing of the correlation output by different self-attention sublayers, perform nonlinear transformation on the global dependence relationship through the multi-layer perceptron sublayer, and obtain context feature data enhanced by the feature sequence, and determine the global scattering feature data through residual connection and layer normalization processing of the context features output by different perceptron sublayers.

[0153] Further, the preprocessing module 302 is specifically configured to determine the spatial weight of a sampling point based on the spatial coordinates of the sampling point corresponding to the ground environment data and the observation parameter through the kernel function of the geographical weighted regression model, determine the regression coefficient of the ground environment data and the observation parameter based on fitting calculation of the spatial weight through the weighted least squares method, and determine the prior inversion data based on the regression coefficient.

[0154] Further, the geographical weighted regression model is represented as:

[0155]

[0156] In the formula, (u i ,v i ) is the spatial coordinates of the i th sampling point, β j (u i ,v i ) is the j th independent variable regression coefficient of the sampling point i, and εi is the error term for sample point i, and ε i 2 ), Cov(ε i ,ε j ) = 0 (i≠j).

[0157] Further, the GNSS-R soil moisture inversion device 300 further comprises:

[0158] An initialization module (not shown in the figure) is configured to initialize a kernel function of the geographically weighted regression model according to the density of the sample points in the target region.

[0159] An optimization module (not shown in the figure) is configured to input the ground environment sample data and the sample ground surface reflectivity into the geographically weighted regression model to determine the theoretical soil moisture, and adjust the bandwidth and the initial parameters of the kernel function based on the residual error of the theoretical soil moisture and the actual soil moisture associated with the sample points.

[0160] Further, the observation parameters include the ground surface reflectivity, the leading edge slope, the trailing edge slope and the incident angle.

[0161] The ground surface reflectivity is determined by using the following calculation method:

[0162]

[0163] where Γ is the ground surface reflectivity, P r is the uncorrected peak value of the delay-Doppler diagram, P t is the transmitted right-handed circular polarization power, G t is the gain of the transmitting antenna, R ts is the distance between the transmitter and the specular reflection point, R sr is the distance between the specular reflection point and the receiver, G r is the gain of the receiving antenna, and λ is the radar wavelength.

[0164] Further, the GNSS-R soil moisture inversion device 300 further comprises:

[0165] A partition module (not shown in the figure) is configured to obtain the area of the region to be studied; if the area of the region to be studied is greater than a preset area, the region to be studied is divided into a plurality of target regions based on the preset area; if the area of the region to be studied is less than or equal to the preset area, the region to be studied is taken as a target region.

[0166] ​The specific definition of the GNSS-R soil moisture inversion device can refer to the definition of the GNSS-R soil moisture inversion method in the foregoing, and will not be repeated here. Each module in the GNSS-R soil moisture inversion device can be realized by software, hardware, and a combination thereof, in whole or in part. The above-mentioned modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to each of the above modules.

[0167] Based on the above method as shown in Figure 1 , accordingly, the embodiment of the present application also provides a readable storage medium, which stores a computer program, and the program is executed by a processor to realize the GNSS-R soil moisture inversion method as shown in Figure 1 .

[0168] Based on such understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a U disk, a mobile hard disk, etc.), and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the method described in each implementation scenario of the present application.

[0169] Based on the above method as shown in Figure 1 , and Figure 10 the virtual device embodiment, in order to achieve the above purpose, as shown in Figure 11 , the embodiment of the present application also provides a computer device, which includes a processor 401 and a memory 402, and the memory 402 stores a program or instruction that can be run on the processor 401, and the program or instruction is executed by the processor 401 to realize the GNSS-R soil moisture inversion method as shown in Figure 1 .

[0170] The memory 402 can be used to store software programs and various data. The memory 402 can mainly include a first storage area storing programs or instructions and a second storage area storing data, wherein the first storage area can store an operating system, application programs or instructions required by at least one function (such as a sound playing function, an image playing function, etc.), and the like. In addition, the memory 402 can include a volatile memory or a non-volatile memory, or the memory 402 can include both volatile and non-volatile memories. The non-volatile memory can be a Read-Only Memory (ROM), a Programmable ROM (PROM), an Erasable PROM (EPROM), an Electrically EPROM (EEPROM), or a flash memory. The volatile memory can be a Random Access Memory (RAM), a Static RAM (SRAM), a Dynamic RAM (DRAM), a Synchronous DRAM (SDRAM), a Double Data Rate SDRAM (DDR SDRAM), an Enhanced SDRAM (ESDRAM), a Synch link DRAM (SLDRAM), and a Direct Rambus RAM (DRRAM). The memory 402 in the embodiments of the present application includes but is not limited to these and any other suitable types of memories.

[0171] The processor 401 can include one or more processing units; optionally, the processor 401 integrates an application processor and a modem processor, wherein the application processor mainly processes operations related to an operating system, a user interface, and an application program, and the modem processor mainly processes wireless communication signals, such as a baseband processor. It can be understood that the above-mentioned modem processor can also not be integrated into the processor 401.

[0172] The computer device can specifically be a personal computer, a server, a network device, and the like.

[0173] Optionally, the computer device can further include a user interface, a network interface, a camera, a radio frequency (RF) circuit, a sensor, an audio circuit, a WI-FI module, etc. The user interface can include a display, an input unit such as a keyboard, etc. Optionally, the user interface can further include a USB interface, a card reader interface, etc. The network interface can optionally include a standard wired interface, a wireless interface (such as a Bluetooth interface, a WI-FI interface), etc.

[0174] Those skilled in the art can understand that the computer device structure provided by the embodiment does not constitute a limitation on the computer device, and can include more or fewer components, or combine certain components, or different component arrangements.

[0175] Those skilled in the art can understand that the drawings are only a schematic of a preferred implementation scenario, and the modules or flows in the drawings are not necessarily essential to implement the present application. Those skilled in the art can understand that the modules in the device in the implementation scenario can be distributed in the device in the implementation scenario according to the description of the implementation scenario, or can be changed and located in one or more devices different from the implementation scenario. The modules of the above implementation scenario can be combined as one module, or can be further split into multiple sub-modules.

[0176] The above application number is only for description, and does not represent the advantages and disadvantages of the implementation scenario. The above disclosure is only some specific implementation scenarios of the present application, but the present application is not limited thereto, and any changes that those skilled in the art can think of should fall within the protection scope of the present application.

Claims

1. A GNSS-R soil moisture inversion method, characterized in that, The method includes: The delayed-Doppler image obtained by the spaceborne GNSS-R probe of the target area is input into the feature extraction model corresponding to the target area. The global dependency of the feature sequence of the delayed-Doppler image is extracted by the feature extraction model, and the global scattering feature data of the target area is output based on the global dependency. The ground environment data of the target area and the observation parameters of the spaceborne GNSS-R are input into the geographic weighted regression model corresponding to the target area. The spatial correlation between the ground environment data and the observation parameters is fitted and calculated by the geographic weighted regression model, and the prior inversion data of soil moisture in the target area is output. The soil moisture inversion results are obtained by performing nonlinear transformation and feature fusion on the global scattering feature data, the prior inversion data, and the observation parameters based on the soil moisture model.

2. The GNSS-R soil moisture inversion method according to claim 1, characterized in that, The feature extraction model includes a segmentation unit and a multi-layer encoder. Each encoder layer includes a multi-head self-attention sublayer and a multi-layer perceptron sublayer. The feature extraction model extracts the global dependencies of the feature sequences of the delayed-Doppler image and outputs the global scattering feature data of the target region based on the global dependencies, including: The delayed-Doppler image is segmented by the segmentation unit to obtain a pixel block sequence, and position encoding is added to the pixel block sequence to form the feature sequence; The multi-head self-attention sublayer captures the correlation between different pixel blocks in the feature sequence; The correlation relationships output by different self-attention sub-layers are subjected to residual connection and layer normalization to determine global dependencies; The global dependency is nonlinearly transformed through the multilayer perceptron sublayer, resulting in enhanced contextual feature data of the feature sequence. The context features output from different perceptron sublayers are subjected to residual connection and layer normalization to determine global scattering feature data.

3. The GNSS-R soil moisture inversion method according to claim 1, characterized in that, The process of fitting and calculating the spatial correlation between the ground environment data and the observation parameters of the spaceborne GNSS-R using the geographic weighted regression model includes: The spatial weight of the sampling point is determined by the kernel function of the geographic weighted regression model, based on the spatial coordinates of the sampling points corresponding to the ground environment data and the observation parameters. The spatial weights are fitted and calculated using the weighted least squares method to determine the regression coefficients of the ground environment data and the observation parameters; The prior inversion data is determined based on the regression coefficients.

4. The GNSS-R soil moisture inversion method according to claim 3, characterized in that, The geographically weighted regression model is expressed as follows: In the formula, y i It is the prior inversion data of the i-th sampling point, (u i ,v i ) represents the spatial coordinates of the i-th sampling point, and β j (u i ,v i ) is the regression coefficient of the j-th independent variable at sampling point i, x j It is the j-th independent variable, ε i It is the error term for sampling point i, and ε i ~N(0,σ 2 ), Cov(ε i ,ε j )=0(i≠j).

5. The GNSS-R soil moisture inversion method according to claim 3, characterized in that, The method further includes: The kernel function of the geographic weighted regression model is initialized based on the density of sample points within the target area; Ground environmental sample data and sample surface reflectance are input into the geographic weighted regression model to determine theoretical soil moisture; Based on the residual between the theoretical soil moisture and the actual soil moisture associated with the sample points, the bandwidth and initial parameters of the kernel function are adjusted.

6. The GNSS-R soil moisture inversion method according to claim 1, characterized in that, The observation parameters include surface reflectance, leading edge slope, trailing edge slope, and incident angle; The surface reflectance is determined using the following calculation method: Where Γ is the surface reflectance, P r It is the uncorrected peak of the delay-Doppler plot, P t It is the right-hand circularly polarized power of the emitted signal, G. t It is the gain of the transmitting antenna, R ts R is the distance between the transmitter and the specular reflection point. sr G is the distance between the specular reflection point and the receiver. r λ is the gain of the receiving antenna, and λ is the radar wavelength.

7. The GNSS-R soil moisture inversion method according to any one of claims 1 to 5, characterized in that, The method further includes: Obtain the area of ​​the region to be studied; If the area of ​​the region to be studied is larger than the preset area, then the region to be studied is divided into multiple target regions based on the preset area; If the area of ​​the region to be studied is less than or equal to the preset area, then the region to be studied is taken as the target region.

8. A GNSS-R soil moisture inversion device, characterized in that, The device includes: The feature extraction module is used to input the delayed-Doppler image obtained by the spaceborne GNSS-R probe of the target area into the feature extraction model corresponding to the target area, extract the global dependency of the feature sequence of the delayed-Doppler image through the feature extraction model, and output the global scattering feature data of the target area based on the global dependency. The preprocessing module is used to input the ground environment data of the target area and the observation parameters of the spaceborne GNSS-R into the geographic weighted regression model corresponding to the target area, and to perform fitting calculations on the spatial correlation of the ground environment data and the observation parameters through the geographic weighted regression model, and output the prior inversion data of soil moisture in the target area. The inversion module is used to perform nonlinear transformation and feature fusion on the global scattering feature data, the prior inversion data and the observation parameters based on the soil moisture model to obtain the soil moisture inversion result.

9. A readable storage medium having a program or instructions stored thereon, characterized in that, When the program or instructions are executed by the processor, they implement the GNSS-R soil moisture inversion method as described in any one of claims 1 to 7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the GNSS-R soil moisture inversion method as described in any one of claims 1 to 7.

Images