Satellite-borne GNSS-R vegetation water content inversion method of BiGRU-CNN-Net deep learning model based on attention mechanism

Abstract

The invention discloses a satellite-borne GNSS-R vegetation water content inversion method of a BiGRU-CNN-Net deep learning model based on an attention mechanism, the model combines the advantages and characteristics of a bidirectional gating recursive unit neural network (Bi-GRU), the attention mechanism and a convolutional neural network (CNN), the CNN can effectively extract the spatial features of DDM, the Bi-GRU is good at capturing context information, the attention mechanism is good at capturing context information, and the depth of the BiGRU-CNN-Net deep learning model is good at capturing context information. Therefore, the input sequence can be better understood to solve the problem that the CNN model may ignore sequence information or time sequence dependency, and meanwhile, the attention mechanism can further help the model to highlight important positions or features in the sequence. According to the method, a mixed deep learning model is combined with satellite-borne GNSS-R data and multi-source satellite remote sensing data to cooperatively invert the local high-temporal-spatial-resolution vegetation water content for the first time, a new vegetation water content acquisition means is provided, and a satellite-borne GNSS-R vegetation water content inversion method considering multiple influence factors is realized. And the accuracy of satellite-borne GNSS-R inversion is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of vegetation moisture content inversion of GNSS reflection signals, and specifically relates to a satellite-borne GNSS-R vegetation moisture content inversion method based on a BiGRU-CNN-Net deep learning model with an attention mechanism. Background Art

[0002] Vegetation is an important component of terrestrial ecosystems and plays an important role in carbon fixation and oxygen release. Vegetation Water Content (VWC), as a key variable of vegetation phenology, can be used to guide agricultural irrigation, predict yields, drought assessment, and predict forest fires and other natural disasters. Therefore, high-precision, long-time series VWC estimation plays an important role in analyzing environmental changes, studying abnormal climate phenomena, and monitoring vegetation growth status. The traditional VWC monitoring method is to measure on-site, which can only obtain data from a small range and cannot obtain historical data. It is destructive, time-consuming, and labor-intensive. Global Navigation Satellite System-Reflectometry (GNSS-R) technology provides us with a technology that can replace traditional monitoring methods to obtain in-situ long-time series VWC information, because the amplitude of direct and reflected GNSS interferometric signals is related to the change of VWC. With the development of GNSS-R technology, GNSS-R technology has been proven to have the ability to monitor geophysical parameters such as soil moisture, wind speed, swell, sea ice, ocean tides, snow depth, and vegetation.

[0003] At present, the research on vegetation parameter inversion using satellite GNSS-R technology mainly focuses on the inversion of AGB, while there are few quantitative studies on VWC, and most of them use satellite GNSS-R observation data to establish sensitivity analysis with VWC, which belongs to the early stage of qualitative analysis. Therefore, exploring a better VWC quantitative inversion method is an urgent problem to be solved. Although there have been recent cases of using machine learning models to invert vegetation water content, they all use a single model, and no hybrid deep learning model has been discussed to invert VWC.

[0004] In view of this, the present invention is proposed. Summary of the invention

[0005] In order to solve the shortcomings of the existing CYGNSS satellite onboard GNSS-R vegetation moisture inversion method, the present invention provides a onboard GNSS-R vegetation moisture inversion method based on a BiGRU-CNN-Net deep learning model with an attention mechanism. The model combines the advantages and characteristics of a bidirectional gated recurrent unit neural network (Bi-GRU), an attention mechanism and a convolutional neural network (CNN). CNN can effectively extract the spatial features of DDM, and Bi-GRU is good at capturing contextual information, thereby better understanding the input sequence to solve the problem that the CNN model may ignore sequence information or temporal dependency. At the same time, the attention mechanism can further help the model highlight important positions or features in the sequence. On this basis, the DDM images (BRCS, effective scattering area) CYGNSS variable parameters and surface auxiliary parameters (such as SMAP surface roughness, GPM rainfall, ECMWF data surface temperature and AMSRU data soil moisture) were integrated to build and train a BiGRU-CNN-Net deep learning model based on the attention mechanism. The results were compared with the VWC inverted by the LightGBM and XGBoost models, which effectively improved the inversion accuracy of the local vegetation moisture content of the spaceborne GNSS-R.

[0006] To achieve the above object, the basic concept of the technical solution adopted by the present invention is:

[0007] A method for inverting vegetation moisture content using a spaceborne GNSS-R vegetation moisture inversion method based on a BiGRU-CNN-Net deep learning model with an attention mechanism includes the following steps:

[0008] Step S1, obtaining CYGNSS satellite-borne GNSS-R, SMAP, GPM-IMERG rainfall, ECMWF data, AMSRU data, and global surface water data;

[0009] Step S2, extracting observation variables from the onboard GNSS-R L1 level data, calculating characteristic observation values, extracting parameters in auxiliary products, and performing quality control on all acquired data sets;

[0010] Step S3, performing spatiotemporal matching and data division on all data sets;

[0011] Step S4, construction and training of the BiGRU-CNN-Net deep learning model based on the attention mechanism for vegetation water content inversion;

[0012] In step S5, the test data set is input into the trained BiGRU-CNN-Net deep learning model based on the attention mechanism to obtain the inverted vegetation moisture content value, and the result is compared with the vegetation moisture content inverted by the previous model (LightGBM and XGBoost model) for performance evaluation.

[0013] Furthermore, the GNSS-R data onboard the CYGNSS satellites used include: normalized bistatic radar cross section (NBRCS in LI-level metadata); effective scattering area (eff_scatter in LI-level metadata); signal-to-noise ratio (ddm_snr in LI-level metadata); receiver antenna gain signal-to-noise ratio (sp_rx_gain in LI-level metadata); equivalent isotropic radiated power (gps_eirp in LI-level metadata); distance between the receiver and the mirror reflection point (rx_to_sp_range in LI-level metadata); distance between the mirror reflection point and the GNSS transmitter (tx_to_sp_range in LI-level metadata); latitude of the mirror reflection point (sp_lat in LI-level metadata); longitude of the mirror reflection point (sp_lon in LI-level metadata); incident angle of the mirror reflection point (sp_inc_angle in LI-level metadata), etc.; the vegetation moisture content verification data is SMAP from MODIS data. VWC data; surface auxiliary data include: GPM-IMERG rainfall, ECMWF data, AMSRU data, global surface water data

[0014] Furthermore, the specific steps of extracting observation variables from the satellite-borne GNSS-R L1 data, calculating characteristic observation values, extracting parameters in auxiliary products, and performing quality control on all acquired data sets as described in step S2 are as follows:

[0015] Step 2.1, extract observation variables from the spaceborne GNSS-R L1 data, such as: bistatic radar cross section NBRCS, distance between the specular reflection point and the GNSS transmitter and GNSS-R receiver, latitude of the specular reflection point, longitude of the specular reflection point, incident angle of the specular reflection point, receiver antenna gain, equivalent isotropic radiated power and signal-to-noise ratio.

[0016] Step 2.2, using the extracted GNSS-R observation variables to calculate characteristic observation values, including: surface reflectivity, range correction gain (RCG), leading edge slope (LES) and trailing edge slope (TES), where the surface reflectivity is calculated as shown in formula (4), and the range correction gain is calculated as shown in formula (5):

[0017]

[0018] In the formula, R t is the distance between the mirror reflection point and the GNSS transmitter (tx_to_sp_range in L1), R r is the distance between the receiver and the specular reflection point (rx_to_sp_range in L1) and Pbrcs is BRCS (the variable name in L1 is brcs). In addition, the trailing edge slope (LES) is calculated based on formula (4) according to the reflectivity delay waveform value at delay bin m (peak delay bin) and m+3.

[0019] In addition, the distance correction gain RCG is calculated as follows:

[0020]

[0021] In the formula, G r is the receiver antenna gain pattern, the unit of RCG is 10 27 ×dBi×m -4 , in fact, the meaning of the parameters is shown in the above formula.

[0022] Step 2.3, extract variable parameters from auxiliary products. The variable indicators required by the present invention are extracted from different data products: "vegetation moisture content" and "surface roughness" are extracted from SMAP products; daily rainfall data are extracted from GPM rainfall; "surface temperature" indicators are extracted from ECMWF data; soil moisture products are extracted from AMSRU data products; and seasonal inland water indicators are extracted from global surface water data.

[0023] Step 2.4, perform data screening and quality control on all acquired data. In order to ensure the accuracy of the inverted VWC, we need to screen the data with poor quality. The specific screening items are as follows:

[0024] 1) The GYGNSS satellite-borne CYGNSS-R data removed the data of "S-band transmitter power on", "large spacecraft attitude error", "blackbody DDM", "DDM test mode" and "low confidence in GPS equivalent isotropic radiated power (EIRP) estimate" according to the quality flag in the CYGNSS L1 metadata ("quality-flags" in L1). In addition, CYGNSS observations with antenna gain <0, BRCS uncertainty greater than 1, range correction gain (RCG) less than 0, and observations with an angle of incidence greater than 65° were filtered to avoid including noisy DDM. 2) For the quality control of auxiliary data products, first, for SMAP products, we use the "retrieval_qual_flag" variable in the L1 data to retain high-quality inversion values ​​with "retrieval_qual_flag" equal to 0 (binary "0000") or equal to 8 (binary "1000"), and remove values ​​less than 0.02 and -9999 in the L1 metadata. In addition, all data products exclude data containing NAN values. Second, for global surface water data, we use a 2% threshold to remove 3km grid pixels of inland water bodies, and about 11.80% of the data samples are excluded.

[0025] Furthermore, step S3 performs spatiotemporal matching and data division on all data sets. The specific steps are as follows:

[0026] Step 3.1: All data products are unified into the EASE-Grid 2.09km×9km grid. First, the surface auxiliary parameter data (GPM rainfall data, surface temperature data in ECMWF data, soil moisture data in AMSRU products, and inland water data in global surface water data) are unified with the SMAP 9km×9km grid resolution in time and space dimensions. Then, the CYGNSS data are reprocessed and projected into the EASE-Grid 2.09km×9km grid used by the SMAP data to generate a data set with the same number of grid points as the target area. All data products are resampled using linear interpolation methods in time scale to keep them highly consistent, and bilinear interpolation methods are used in space.

[0027] Step 3.2: The dataset that has been spatiotemporally matched and screened is divided into a training set, a validation set, and a test set according to the order of observation time. 60% of the data is used for training, 25% for testing, and 15% for validation.

[0028] Furthermore, the BiGRU-CNN-Net deep learning model based on the attention mechanism for vegetation water content inversion described in step S4 is constructed and trained. The BiGRU-CNN-Net deep learning model based on the attention mechanism comprises three layers (Bi-GRU layer, attention mechanism layer and CNN layer). The Bi-GRU layer can extract context features from the forward and reverse directions to maximize the extraction of feature variable information; the attention mechanism can enhance the role of input parameters in the VWC inversion process, and the CNN layer is based on a convolutional neural network, which can quickly and accurately extract key spatial features in DDM.

[0029] Furthermore, step S5 inputs the test data set into the trained BiGRU-CNN-Net deep learning model based on the attention mechanism to obtain the inverted vegetation moisture content value, and compares the result with the vegetation moisture content inverted by the previous model (LightGBM and XGBoost model) for performance evaluation. The training process of the BiGRU-CNN-Net deep learning model based on the attention mechanism is as follows:

[0030] Step S5.1, divide the quality-controlled data, and input the divided training set into the BiGRU-CNN-Net deep learning model based on the attention mechanism;

[0031] Step S5.2, input all filtered data including two DDM images (NBRCS, effective scattering area) of 17×11 two-dimensional matrix, GNSS-R observation variables and surface auxiliary parameters into different models, such as: NBRCS, effective scattering area, surface reflectivity, SNR, LES, receiver antenna gain, ERIP, distance between receiver and specular reflection point, distance between specular reflection point and GNSS transmitter, location information of specular reflection point (longitude and latitude), incident angle of specular reflection point, range correction gain (RCG), soil moisture (SM), surface temperature (SST), surface roughness (RC), rainfall (GPM);

[0032] In step S5.3, we conducted a large number of experiments on the preprocessed data set, and used trial and error and grid search methods to optimize the hyperparameters of the BiGRU-CNN-Net model based on the attention mechanism, and finally determined the optimal hyperparameter settings of the model, namely, CNN convolution filters (CNN Filters) are 16, BiGRU units (BiGRU Units) are 64, convolution kernels (kernelsize) are 3, and the ratio of discarded neurons (dropout ratio) is 0.3; the number of iterations (epochs) is 150, and the optimization model parameter algorithm (optimizer) selects Adam; the activation function (Activation function) selects ReLU and Sigmoid function. The patent uses MSE as the loss function, which is specifically defined as follows:

[0033]

[0034] Where N represents the number of samples, represents the predicted VWC value, and u i Indicates SMAP VWC.

[0035] In step S5.4, the SMAP VWC value is input as the target value into the BiGRU-CNN-Net deep learning model based on the attention mechanism for training; the BiGRU-CNN-Net deep learning model based on the attention mechanism is a hybrid model that fully utilizes the advantages of Bi-GRU, attention mechanism and CNN, and combines these three models into a new VWC estimation model to improve the VWC inversion performance. We fully utilize the convolutional neural network to quickly and accurately extract key spatial features in DDM, but may not be able to capture long-range dependencies and contextual information as well as the inherent temporal relationship in the time series data in VWC, which may cause the model to predict incorrectly when facing complex problems. The Bi-GRU model can extract contextual features from the forward and reverse directions to maximize the extraction of feature variable information. At the same time, the attention mechanism can strengthen the role of input parameters in the VWC inversion process.

[0036] Step S5.5, after the training is completed, the test data set is input into the trained BiGRU-CNN-Net deep learning model based on the attention mechanism to obtain the inverted vegetation water content value, and the result is verified with the SMPA VWC data from the MODIS product. Finally, the VWC inverted by the LightGBM and XGBoost models are compared.

[0037] After adopting the above technical scheme, the present invention has the following beneficial effects compared with the prior art.

[0038] The novel BiGRU-CNN-Net deep learning neural network model based on the attention mechanism constructed in the present invention combines the advantages and characteristics of the bidirectional gated recurrent unit neural network (Bi-GRU), the attention mechanism and the convolutional neural network (CNN). CNN can effectively extract the spatial features of DDM, and Bi-GRU is good at capturing contextual information to better understand the input sequence to solve the problem that the CNN model may ignore sequence information or temporal dependencies. At the same time, the attention mechanism can further help the model highlight important positions or features in the sequence. At the same time, we integrate the proposed model with DDM images (bistatic radar cross section (BRCS), effective scattering area (Effective Scattering Area), CYGNSS variable parameters and surface auxiliary parameters to construct a deep learning network model for local high spatiotemporal resolution vegetation moisture inversion. This invention realizes for the first time the use of a hybrid deep learning model combined with spaceborne GNSS-R data and multi-source satellite remote sensing data to collaboratively invert local high spatiotemporal resolution vegetation moisture content, provides a new means of obtaining vegetation moisture content, and realizes a method for spaceborne GNSS-R inversion of vegetation moisture content taking into account multiple influencing factors, thereby improving the accuracy of spaceborne GNSS-R inversion.

[0039] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The accompanying drawings are part of this application and are used to provide a further understanding of the present invention. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but do not constitute an improper limitation of the present invention. Obviously, the drawings described below are only some embodiments. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0041] Figure 1 It is a schematic diagram of the inversion method flow in the implementation case of the present invention;

[0042] Figure 2 The flowchart of the BiGRU-CNN-Net deep learning model algorithm based on the attention mechanism provided in the implementation case of the present invention;

[0043] Figure 3 This is a test effect diagram of the BiGRU-CNN-Net model based on the attention mechanism provided in the implementation case of the present invention on the training set, test set and validation set.

[0044] Figure 4 It is a scatter density diagram of vegetation water content inversion using three models provided in the implementation case of the present invention;

[0045] Figure 5 This is a local deviation distribution map of vegetation moisture content inverted by the three models provided in the implementation case of the present invention.

[0046] Figure 6 It is a histogram of the deviation distribution between the three model inversion VWC and SMAP VWC provided in the implementation case of the present invention.

[0047] It should be noted that these drawings and textual descriptions are not intended to limit the conceptual scope of the present invention in any way, but are intended to illustrate the concept of the present invention for those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The following embodiments are used to illustrate the present invention but are not used to limit the scope of the present invention.

[0049] Example 1

[0050] The present invention provides a method for inverting vegetation water content of a satellite GNSS-R based on a BiGRU-CNN-Net deep learning model with an attention mechanism. In order to verify the advantages of the method proposed in the present invention, the VWC inverted by the BiGRU-CNN-Net deep learning model based on the attention mechanism is compared with the vegetation water content results inverted by previous models (extreme gradient boosting (XGBoost) algorithm and lightweight gradient boosting machine (Light Gradient Boosting Machine, LightGBM) model).

[0051] The experiments of the implementation cases of the present invention were conducted on a computer equipped with an Intel(R) Core(TM) i5-12400F processor, 32.0GB RAM and NVIDIA GeForce RTX 3060 GPU, using Python 3.8. The basic software packages include numpy 1.24.4, scikit-learn 0.22.1, scipy 1.10.1, tdqm 4.66.4 and Pillow 9.0.0.

[0052] A method for inverting vegetation moisture content in space-borne GNSS-R based on the BiGRU-CNN-Net deep learning model with attention mechanism. The technical solution implementation process is as follows: Figure 1 As shown, the following steps are included:

[0053] Step S1, obtaining CYGNSS satellite-borne GNSS-R, SMAP, GPM-IMERG rainfall, ECMWF data, AMSRU data, and global surface water data;

[0054] Step S2, extracting observation variables from the onboard GNSS-R L1 level data, calculating characteristic observation values, extracting parameters in auxiliary products, and performing quality control on all acquired data sets;

[0055] Step S3, performing spatiotemporal matching and data division on all data sets;

[0056] Step S4, construction and training of the BiGRU-CNN-Net deep learning model based on the attention mechanism for vegetation water content inversion;

[0057] In step S5, the test data set is input into the trained BiGRU-CNN-Net deep learning model based on the attention mechanism to obtain the inverted vegetation moisture content value, and the result is compared with the vegetation moisture content inverted by the previous model (LightGBM and XGBoost model) for performance evaluation.

[0058] As an implementation of this embodiment, in step S1, the CYGNSS satellite-borne GNSS-R includes the following observation variable parameters:

[0059] Table 1 CYGNSS observation variables used in the present invention

[0060]

[0061] The surface auxiliary parameters used in the present invention include the following data products:

[0062] Table 2 Surface auxiliary parameter data products used in the present invention

[0063]

[0064] As an implementation manner of this embodiment, preferably, the extraction of observation variables from the satellite-borne GNSS-R L1 level data, calculation of characteristic observation values, extraction of parameters in auxiliary products, and quality control of all acquired data sets described in step S2 are specifically performed as follows:

[0065] Step 2.1, extract observation variables from the spaceborne GNSS-R L1 data, such as: bistatic radar cross section NBRCS, distance between the specular reflection point and the GNSS transmitter and GNSS-R receiver, latitude of the specular reflection point, longitude of the specular reflection point, incident angle of the specular reflection point, receiver antenna gain, equivalent isotropic radiated power and signal-to-noise ratio.

[0066] Step 2.2, use the extracted GNSS-R observation variables to calculate the characteristic observation values, including: surface reflectivity, range correction gain (RCG), leading edge slope (LES) and trailing edge slope (TES), etc. The calculation of surface reflectivity is shown in formula (7), and the calculation of range correction gain is shown in formula (8):

[0067]

[0068] Where R t is the distance between the mirror reflection point and the GNSS transmitter (tx_to_sp_range in L1), R r is the distance between the receiver and the specular reflection point (rx_to_sp_range in L1) and P brcs is BRCS (the variable name in L1 is brcs). In addition, the trailing edge slope (LES) is calculated based on formula (7) according to the reflectivity delay waveform value at delay bin m (peak delay bin) and m+3.

[0069] In addition, the calculation formula of the distance correction gain RCG is as follows:

[0070]

[0071] In the formula, G r is the receiver antenna gain pattern, the unit of RCG is 10 27 ×dBi×m -4 , in fact, the meaning of the parameters is shown in the above formula.

[0072] Step 2.3, extract variable parameters from auxiliary products. The variable indicators required by the present invention are extracted from different data products: "vegetation moisture content" and "surface roughness" are extracted from SMAP products; daily rainfall data are extracted from GPM rainfall; "surface temperature" indicators are extracted from ECMWF data; soil moisture products are extracted from AMSRU data products; and seasonal inland water indicators are extracted from global surface water data.

[0073] Step 2.4, perform data screening and quality control on all acquired data. In order to ensure the accuracy of the inverted VWC, we need to screen the data with poor quality. The specific screening items are as follows:

[0074] 1) The GYGNSS satellite-borne CYGNSS-R data removed the data of "S-band transmitter power on", "large spacecraft attitude error", "blackbody DDM", "DDM test mode" and "low confidence in GPS equivalent isotropic radiated power (EIRP) estimate" according to the quality flags in the CYGNSS L1 metadata ("quality-flags" in L1). In addition, we also filtered out CYGNSS observations with antenna gain < 0, BRCS uncertainty greater than 1, range correction gain (RCG) less than 0, and removed observations with an incidence angle greater than 65° to avoid including noisy DDM. 2) For the quality control of auxiliary data products, first, for SMAP products, we use the "retrieval_qual_flag" variable in the L1 data to retain high-quality inversion values ​​with "retrieval_qual_flag" equal to 0 (binary "0000") or equal to 8 (binary "1000"), and remove values ​​less than 0.02 and -9999 in the L1 metadata. In addition, all data products exclude data containing NAN values. Second, for global surface water data, we use a 2% threshold to remove 3km grid pixels of inland water bodies, and about 11.80% of the data samples are excluded.

[0075] As an implementation of this embodiment, preferably, step S3 performs spatiotemporal matching and data division on all data sets, and the specific steps are as follows:

[0076] Step 3.1: All data products are unified into the EASE-Grid 2.09km×9km grid. First, the surface auxiliary parameter data (GPM rainfall data, surface temperature data in ECMWF data, soil moisture data in AMSRU products, and inland water data in global surface water data) are unified with the SMAP 9km×9km grid resolution in time and space dimensions. Then, the CYGNSS data are reprocessed and projected into the EASE-Grid 2.09km×9km grid used by the SMAP data to generate a data set with the same number of grid points as the target area. All data products are resampled using linear interpolation methods in time scale to keep them highly consistent, and bilinear interpolation methods are used in space.

[0077] Step 3.2: The dataset that has been spatiotemporally matched and screened is divided into a training set, a validation set, and a test set according to the order of observation time. 60% of the data is used for training, 25% for testing, and 15% for validation.

[0078] As an implementation mode of this embodiment, preferably, the construction and training of the BiGRU-CNN-Net deep learning model based on the attention mechanism for vegetation water content inversion described in step S4, the BiGRU-CNN-Net deep learning model based on the attention mechanism includes three layers (Bi-GRU layer, attention mechanism layer and CNN layer), the Bi-GRU layer can extract context features from the forward and reverse directions to maximize the extraction of feature variable information; the attention mechanism can strengthen the role of input parameters in the VWC inversion process, and the CNN layer is based on a convolutional neural network, which can quickly and accurately extract key spatial features in DDM, such as Figure 2 shown.

[0079] As an implementation of this embodiment, preferably, step S5 inputs the test data set into the trained BiGRU-CNN-Net deep learning model based on the attention mechanism to obtain the inverted vegetation moisture content value, and compares the result with the vegetation moisture content inverted by the previous model (LightGBM and XGBoost model) for performance evaluation. The training process of the BiGRU-CNN-Net deep learning model based on the attention mechanism is as follows:

[0080] Step S5.1, divide the quality-controlled data, and input the divided training set into the BiGRU-CNN-Net deep learning model based on the attention mechanism. The test effect of the BiGRU-CNN-Net model based on the attention mechanism on the training set, test set and validation set is shown in the figure below: Figure 3 As shown;

[0081] Step S5.2, input all filtered data including two DDM images (NBRCS, effective scattering area) of 17×11 two-dimensional matrix, GNSS-R observation variables and surface auxiliary parameters into different models, such as: NBRCS, effective scattering area, surface reflectivity, SNR, LES, receiver antenna gain, ERIP, distance between receiver and specular reflection point, distance between specular reflection point and GNSS transmitter, location information of specular reflection point (longitude and latitude), incident angle of specular reflection point, range correction gain (RCG), soil moisture (SM), surface temperature (SST), surface roughness (RC), rainfall (GPM);

[0082] In step S5.3, we conducted a large number of experiments on the preprocessed data set, and used trial and error and grid search methods to optimize the hyperparameters of the BiGRU-CNN-Net model based on the attention mechanism, and finally determined the optimal hyperparameter settings of the model, namely, CNN convolution filters (CNN Filters) are 16, BiGRU units (BiGRU Units) are 64, convolution kernels (kernelsize) are 3, and the ratio of discarded neurons (dropout ratio) is 0.3; the number of iterations (epochs) is 150, and the optimization model parameter algorithm (optimizer) selects Adam; the activation function (Activation function) selects ReLU and Sigmoid function. The patent uses MSE as the loss function, which is specifically defined as follows:

[0083]

[0084] Where N represents the number of samples, represents the predicted VWC value, and u i Indicates SMAP VWC.

[0085] In step S5.4, the SMAP VWC value is input as the target value into the BiGRU-CNN-Net deep learning model based on the attention mechanism for training; the BiGRU-CNN-Net deep learning model based on the attention mechanism is a hybrid model that fully exploits the advantages of Bi-GRU, attention mechanism and CNN, and combines these three models into a new VWC estimation model to improve the VWC inversion performance. We make full use of the convolutional neural network to quickly and accurately extract key spatial features in DDM, but may not be able to capture long-range dependencies and contextual information as well as the inherent temporal relationship in the time series data in VWC, which may cause the model to predict incorrectly when facing complex problems. The Bi-GRU model can extract contextual features from both forward and reverse directions to maximize the extraction of feature variable information. At the same time, the attention mechanism can strengthen the role of input parameters in the VWC inversion process.

[0086] Step S5.5, when the training is completed, the test data set is input into the trained BiGRU-CNN-Net deep learning model based on the attention mechanism to obtain the inverted vegetation water content value, and the result is verified with the SMPA VWC data from the MODIS product. Finally, the VWC inverted by the LightGBM and XGBoost models are compared.

[0087] The root mean square error (RMSE), Pearson correlation coefficient (R), mean absolute error (MAE) and mean absolute percentage error (MAPE) are used as indicators to evaluate the inversion performance of the model. First, using SMAP VWC as reference data, the BiGRU-CNN-Net deep learning network model based on the attention mechanism is compared and evaluated with the existing LightGBM and XGBoost models. The final results are shown in Figure 4 As shown in Table 3, the scatter density map and inversion accuracy statistics of VWC inverted by the BiGRU-CNN-Net deep learning network model based on the attention mechanism, LightGBM and XGBoost model are given respectively. The following conclusions can be drawn from the table:

[0088] from Figure 4 It can be seen that the VWC and SMAP inverted by the proposed model (BiGRU-CNN-Net) method Compared with LightGBM and XGBoost models, VWC has the best correlation. The BiGRU-CNN-Net model method has the most points distributed around the 1:1 reference line y=x, while there are relatively few scattered points, indicating that the BiGRU-CNN-Net model has the best performance in inverting local VWC, followed by LightGBM, and the worst is the XGBoost model. From Table 3, it can be concluded that: 1) In terms of RMSE, the BiGRU-CNN-Net model based on the attention mechanism has improved the accuracy by 14.86% and 25.00% respectively compared with the other two models (LightGBM and XGBoost models); in terms of MAE, the accuracy has improved by 30.65% and 38.57% respectively; in terms of MAPE, the accuracy has improved by 54.94% and 58.50% respectively; in terms of R, the R value of the BiGRU-CNN-Net deep learning network model is as high as 0.85, which is 7.59% higher than the XGBoost model.

[0089] Table 3 Accuracy of VWC inversion by different models on the test dataset

[0090]

[0091] To further evaluate the performance of the BiGRU-CNN-Net model in inverting VWC, Figure 5The deviation distribution diagram of the proposed model (BiGRU-CNN-Net model) and the LightGBM and XGBoost models is shown. From the figure, it can be seen that the local deviation of the proposed model is small, while the local deviation of the LightGBM and XGBoost models is large, that is to say, the VWC inverted by the proposed model is closer to the reference value than the VWC inverted by LightGBM and XGBoost. At the same time, the areas with smaller errors of the proposed model are in the central and western parts of Australia, and the larger errors are in the eastern part. In order to further verify this conclusion, Figure 6 The deviation distribution histogram of the three model inverted VWC and SMAP VWC is given (the figure shows the mean deviation (μ), standard deviation (σ), mean absolute error (MAE) and 80% quantile of the deviation (Qua), which can be seen from the figure that this is consistent with the above conclusions.

[0092] The above is only a preferred embodiment of the present invention, and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment, it is not used to limit the present invention. Any technician familiar with this patent can make some changes or modify the technical contents suggested above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the scope of the solution of the present invention.

Claims

1. A method for inverting vegetation moisture content using a spaceborne GNSS-R deep learning model based on an attention mechanism BiGRU-CNN-Net, characterized in that: The following steps are involved: Step S1, obtaining CYGNSS satellite-borne GNSS-R, SMAP vegetation moisture content verification data and surface auxiliary data; Step S2, extracting observation variables from the onboard GNSS-R L1 level data, calculating characteristic observation values, extracting parameters in auxiliary products, and performing quality control on all acquired data sets; Step S3, performing spatiotemporal matching and data division on all data sets; Step S4, construction and training of the BiGRU-CNN-Net deep learning model based on the attention mechanism for vegetation water content inversion; In step S5, the test data set is input into the trained BiGRU-CNN-Net deep learning model based on the attention mechanism to obtain the inverted vegetation moisture content value, and the result is compared with the vegetation moisture content inverted by the previous model (LightGBM and XGBoost model) for performance evaluation.

2. According to claim 1, a method for inverting vegetation moisture content using a BiGRU-CNN-Net deep learning model based on an attention mechanism, characterized in that: The CYGNSS satellite-borne GNSS-R data used in step S1 include: normalized bistatic radar cross section (NBRCS in LI-level metadata), effective scattering area (eff_scatter in LI-level metadata), signal-to-noise ratio (ddm_snr in LI-level metadata), receiver antenna gain signal-to-noise ratio (sp_rx_gain in LI-level metadata), equivalent isotropic radiated power (gps_eirp in LI-level metadata), distance between receiver and specular reflection point (rx_to_sp_range in LI-level metadata), distance between specular reflection point and GNSS transmitter (tx_to_sp_range in LI-level metadata), latitude of specular reflection point (sp_lat in LI-level metadata), longitude of specular reflection point (sp_lon in LI-level metadata), and incident angle of specular reflection point (sp_inc_angle in LI-level metadata); The SMAP vegetation water content verification data is the SMAP VWC data from MODIS data; Surface auxiliary data include: GPM-IMERG rainfall, ECMWF data, AMSRU data, and global surface water data.

3. The method for inverting vegetation moisture content using a space-borne GNSS-R deep learning model based on an attention mechanism according to claim 1, characterized in that: Step S2 specifically includes: Step 2.1, extract observation variables from the onboard GNSS-R L1 data, such as bistatic radar cross section NBRCS, distance between the specular reflection point and the GNSS transmitter and GNSS-R receiver, latitude of the specular reflection point, longitude of the specular reflection point, incident angle of the specular reflection point, receiver antenna gain, equivalent isotropic radiated power, and signal-to-noise ratio; Step 2.2, using the extracted GNSS-R observation variables to calculate characteristic observation values, including: surface reflectivity, range correction gain (RCG), leading edge slope (LES) and trailing edge slope (TES), where the surface reflectivity is calculated as shown in formula (1), and the range correction gain (RCG) is calculated as shown in formula (2): In the formula, R t is the distance between the mirror reflection point and the GNSS transmitter (tx_to_sp_range in L1), R r is the distance between the receiver and the specular reflection point (rx_to_sp_range in L1) and P brcs is BRCS (the variable name in L1 is brcs); In addition, the trailing edge slope (LES) is calculated based on formula (1) according to the reflectivity delay waveform value at delay bin m (peak delay bin) and m+3; In addition, the calculation formula of distance correction gain (RCG) is: In the formula, G r is the receiver antenna gain pattern, the unit of RCG is 10 27 ×dBi×m -4 ; Step 2.3, extract variable parameters from auxiliary products; extract variable indicators required by the present invention from different data products respectively: extract "vegetation moisture content" and "surface roughness" from SMAP products; extract daily rainfall data from GPM rainfall; extract "surface temperature" indicators from ECMWF data; extract soil moisture products from AMSRU data products; extract seasonal inland water indicators from global surface water data; Step 2.4, screen all acquired data and perform quality control. In order to ensure the accuracy of the inverted VWC, it is necessary to screen data with poor quality. The specific screening items are as follows: 2.4.1) The GYGNSS satellite-borne CYGNSS-R data has removed the data of "S-band transmitter power on", "large spacecraft attitude error", "blackbody DDM", "DDM test mode" and "low confidence in GPS equivalent isotropic radiated power (EIRP) estimate" according to the quality flag in the CYGNSS L1 metadata ("quality-flags" in L1); in addition, CYGNSS observations with antenna gain < 0, BRCS uncertainty greater than 1, range correction gain (RCG) less than 0, and observations with an incidence angle greater than 65° are filtered to avoid containing noisy DDM; 2.4.2) Quality control of auxiliary data products; First, for SMAP products, the "retrieval_qual_flag" variable in the L1 data is used to retain high-quality inversion values ​​with "retrieval_qual_flag" equal to 0 (binary "0000") or equal to 8 (binary "1000"), and values ​​less than 0.02 and -9999 in the L1 metadata are removed; in addition, data containing NAN values ​​are removed from all data products; Secondly, for the global surface water data, a threshold of 2% was used to remove the 3km grid pixels of inland water bodies, and about 11.80% of the data samples were excluded.

4. The method for inverting vegetation moisture content using a space-borne GNSS-R deep learning model based on an attention mechanism according to claim 1, characterized in that: Step S3 specifically includes the following steps: Step 3.1, unify all data products into the EASE-Grid 2.09km×9km grid; First, the surface auxiliary parameter data are unified with the SMAP 9km×9km grid resolution in time and space dimensions. The surface auxiliary parameter data here include GPM rainfall data, surface temperature data in ECMWF data, soil moisture data in AMSRU products, and inland water data in global surface water data; then the CYGNSS data are reprocessed and projected into the EASE-Grid 2.09km×9km grid used by SMAP data to generate a data set with the same number of grid points as the target area; all data products use linear interpolation methods in time scale to keep them highly consistent, and use bilinear interpolation methods for resampling in space; Step 3.2: The dataset that has been spatiotemporally matched and screened is divided into a training set, a validation set, and a test set according to the order of observation time. 60% of the data is used for training, 25% for testing, and 15% for validation.

5. The method for inverting vegetation moisture content using a space-borne GNSS-R deep learning model based on an attention mechanism according to claim 1, characterized in that: In step S4, The BiGRU-CNN-Net deep learning model based on the attention mechanism consists of three layers: the Bi-GRU layer, the attention mechanism layer, and the CNN layer. The Bi-GRU layer can extract context features from the forward and reverse directions to maximize the extraction of feature variable information. The attention mechanism layer can strengthen the role of input parameters in the VWC inversion process, and the CNN layer is based on convolutional neural network, which can quickly and accurately extract key spatial features in DDM.

6. The method for inverting vegetation moisture content using a space-borne GNSS-R deep learning model based on an attention mechanism according to claim 1, characterized in that: In step S5, The training process of the BiGRU-CNN-Net deep learning model based on the attention mechanism is as follows: Step S5.1, divide the quality-controlled data, and input the divided training set into the BiGRU-CNN-Net deep learning model based on the attention mechanism; Step S5.2, input all filtered data including two DDM images (NBRCS, effective scattering area) of 17×11 two-dimensional matrix, GNSS-R observation variables and surface auxiliary parameters into different models, such as: NBRCS, effective scattering area, surface reflectivity, SNR, LES, receiver antenna gain, ERIP, distance between receiver and specular reflection point, distance between specular reflection point and GNSS transmitter, location information of specular reflection point (longitude and latitude), incident angle of specular reflection point, range correction gain (RCG), soil moisture (SM), surface temperature (SST), surface roughness (RC), rainfall (GPM); Step S5.3, the trial and error method and grid search method are used to optimize the hyperparameters of the BiGRU-CNN-Net model based on the attention mechanism, and the optimal hyperparameter settings of the model are finally determined, that is, CNN convolution filters are 16, BiGRU units are 64, convolution kernels are 3, and the ratio of discarded neurons is 0.3; the number of iterations (epochs) is 150, and the optimizer algorithm (optimizer) selects Adam; the activation function selects ReLU and Sigmoid function; MSE is used as the loss function, which is defined as follows: Where N represents the number of samples, represents the predicted VWC value, and u i represents SMAP VWC; Step S5.4, inputting the SMAP VWC value as the target value into the BiGRU-CNN-Net deep learning model based on the attention mechanism for training; Step S5.5, after the training is completed, the test data set is input into the trained BiGRU-CNN-Net deep learning model based on the attention mechanism to obtain the inverted vegetation water content value, and the result is verified with the SMAPVWC data from the MODIS product. Finally, the VWC inverted by the LightGBM and XGBoost models are compared.

Images