Sea surface wind speed inversion method and system based on GNSS-R forward model

Abstract

The invention relates to a sea surface wind speed inversion method and system based on a GNSS-R forward model, and belongs to the technical field of ocean remote sensing and environment monitoring. The method comprises the following steps: firstly, inputting four types of CYGNSS data to obtain a modeling DDM and a Jacobian matrix, secondly, carrying out feature extraction and data quality screening to obtain a plurality of features related to sea surface wind speed, then carrying out space-time matching on the features and the wind speed, inputting the features into an inversion model, and finally obtaining that the distribution of the DDM generated by a DDM forward model and the distribution of the observation DDM are relatively concentrated, and the overall simulation effect is good. According to the method, the problems that the neural network lacks physical interpretability and the inversion process is complex are solved, and a new thought is provided for later inversion of soil humidity and other indexes.

Description

Technical Field

[0001] The invention relates to a sea surface wind speed inversion method and system based on a GNSS-R forward model, and belongs to the technical field of ocean remote sensing and environmental monitoring. Background Art

[0002] Sea surface wind speed is an important parameter in marine environmental monitoring, and is of great significance for climate change research, marine meteorological forecasting and marine resource development. Traditional sea surface wind speed measurement usually relies on equipment such as buoys or radars, but these methods have problems such as high cost, limited coverage and complex deployment. GNSS-R is a passive remote sensing technology that inverts sea surface characteristics by receiving reflected waves of global navigation satellite signals. It has the advantages of low cost, wide coverage and all-weather operation. In recent years, GNSS-R technology has made significant progress in inverting sea surface wind speed, especially after combining physical models and data-driven algorithms, the inversion accuracy has been significantly improved. To this end, the present invention is proposed to apply GNSS-R technology to sea surface wind speed inversion. Summary of the invention

[0003] In view of the shortcomings of the prior art, the present invention provides a sea surface wind speed inversion method and system based on the GNSS-R forward model, aiming to overcome the problems of high cost and insufficient precision in the prior art. Compared with the prior art, the present invention has the advantages of low cost, wide coverage, and strong anti-interference ability. The present invention introduces the Jacobian matrix for sensitivity analysis and combines it with the attention mechanism in deep learning to further improve the accuracy and reliability of wind speed inversion.

[0004] Terminology explanation: CYGNSS: (Cyclone Global Navigation Satellite System) Cyclone Global Navigation Satellite System; CYGNSS L1 GNSS-R data: Cyclone Global Navigation Satellite System Level 1 GNSS-R data product; DDMI: (Delay Doppler Mapping Instrument) delayed Doppler imager; E2ES: (end-to-end simulator) end-to-end simulator; ERA5: (European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis) European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis data.

[0005] The technical solution of the present invention is as follows: The sea surface wind speed inversion method based on the GNSS-R forward model has the following steps: (1) In the GNSS-R receiver, the reflected signal is first cross-correlated with a local replica of the transmitted signal within the delay and Doppler frequency range to determine the model input, construct a DDM forward model based on the bistatic scattering model, and calculate the Jacobian matrix; (2) Prepare inversion model data, including CYGNSS L1 GNSS-R data, ERA5 wind speed and direction, extract observation variable parameters from GNSS-R data, perform data quality control, data filtering and model data set construction, and complete spatiotemporal matching; (3) Combined with the Jacobian matrix output by the DDM forward model, an inversion model is constructed, and the inversion results of the model with the Jacobian matrix introduced are evaluated and analyzed.

[0006] According to the preferred embodiment of the present invention, in step (1), the model input includes four types, namely: geometric shape, including the position and velocity of the transmitter, receiver and mirror reflection point; Metadata, including sampling time, GPS pseudo-random noise (PRN) code, and specular reflection point index; Power, including the effective isotropic radiated power (EIRP) of the GPS transmitter and the receiver’s antenna parameters; Wind field, a gridded wind field expressed in latitude and longitude coordinates.

[0007] Among them, geometry, metadata and power are all from CYGNSS level 1 data, and the gridded wind field is constructed by the cross-calibrated multi-platform ocean surface wind field product CCMP dataset.

[0008] Preferably, according to the present invention, in step (1), the DDM forward model building process is: The bistatic scattering model formula is: (1) Among them, the integral variable is the two-dimensional position vector defining the point on the sea surface, represents the coherent integration time of the receiver, is the wavelength of the GPS carrier, is The wind speed at is the GPS EIRP (effective isotropic radiated power), which is assumed to remain constant throughout the scattering region. denote the gains of the receiver antenna and transmitter antenna on the surface, respectively, Indicates the transmit power, is the additional rainfall attenuation, Respectively indicate points The path distance from the mean sea surface at to the transmitter and receiver, Surface weighted area for a given delay and Doppler; During the simulation, the semi-empirical Katzberg model is used to convert the wind speed into the mean square slope of the sea surface. In E2ES in the DDMI of CYGNSS, equation (1) is rewritten as a two-dimensional convolution with a delta function to calculate the expected value of the DDM. (2) For fast computation, the Fast Fourier Transform (FFT) method is applied to the two-dimensional convolution: (3) In the fuzziness function and

[0009] Between is the Dirac delta function, In the above formula A brief description of the representation, ** represents the convolution operation, and ,in, is the expected value of the power contribution at each location on the surface, is the mapping of this power from surface coordinates to delay-Doppler coordinates.

[0010] To calculate the above equation (3), the delay Doppler space needs to be discretized. The subsequent Jacobian matrix needs to be inverted together with the observed value. Therefore, the same delay Doppler coordinates as the observed DDM need to be used to accurately reflect the authenticity of the data. The mirror point position estimate generated by DDMI on CYGNSS is very rough and fixed as a discrete value. Therefore, the real mirror reflection point is usually not located at the center of a single DDM bin. The fractionally accurate mirror reflection index generated by using and , the delay-Doppler coordinates of the mobile modeled DDM to align with the observed DDM and other observed values, are discretized, and calculated using a 2D fast Fourier transform; (4) in, , , is the index in discrete delay-Doppler space, and are the delay and Doppler increment, and for CYGNSS, , , , is the second cycle of the GPSC / A code chip, Hz, is the discretized form of the surface integral.

[0011] According to the preferred embodiment of the present invention, in step (1), the Jacobian matrix H is the forward operator with respect to the input vector The partial derivative matrix of each wind speed in is expressed as follows: (5) The Jacobian matrix reflects the sensitivity of each sample of the DDM to the wind speed at each surface grid point.

[0012] Preferably, according to the present invention, in step (1), in order to improve the physical interpretability of the model and overcome the problem of limited accuracy improvement of deep learning methods, by modeling the GNSS-R basic model, a matrix that can represent the sensitivity of each sample of DDM to the wind speed of each surface grid point is output, and some positions in the matrix are zero because these positions correspond to the parts outside the flash area centered on the mirror reflection point. The dispersion power of some functions is very small and is gradually submerged by noise. The actual values ​​of all Jacobian matrices are zero, and the high absolute values ​​in the matrix represent the sensitivity of DDM to the wind speed at a specific position within the relative delay and Doppler range. The two dimensions of the matrix are DDM values ​​and surface wind speed, where DDM is expanded into a vector using delay and Doppler, and surface wind speed is expanded into a vector using latitude and longitude. In order to verify the availability of the model, the data of the model's input variables are obtained from the CYGNSS L1a product at a rate of 1 Hz. The receiver antenna pattern estimated by pre-launch measurements and on-orbit corrections is provided by the CYGNSS SOC. At each observation moment, the simulated DDM generated by the forward operator is compared with the DDM obtained from the L1a data. Two metrics are used to quantify the consistency between the modeled DDM (the DDM generated by the DDM forward model calculation) and the observed DDM (the observed DDM is the DDM obtained using the delayed Doppler imager processing in the CYGNSS L1 data), namely, the relative difference of the effective bins and the correlation coefficient between the two DDM effective bins, as follows:

[0013] (6) in, is the index of DDM, and is the observed DDM and the modeled DDM, and Represent the indexes of delay and Doppler shift respectively, N is the number of valid bins, It is to observe DDM in the delay and Doppler shift The power value at is to model DDM in the delay and Doppler shift The power value at It is Delayed sampling points, It is Doppler shift sampling points; (7) in, For the The correlation coefficient of the DDM, and is the average of all valid bins of the observed DDM and the modeled DDM, and Represent the indexes of delay and Doppler shift respectively, It is to observe DDM in the delay and Doppler shift The power value at is to model DDM in the delay and Doppler shift The power value at It is Delayed sampling points, It is The more similar the shapes of the observed DDM and the modeled DDM are, the closer the correlation coefficient is to 1.

[0014] Preferably, in step (3) of the present invention, the inversion model architecture is: a CNN module, an auxiliary feature processing module, a feature fusion layer and an output layer, the CNN module is used to extract the features of a single delay-Doppler correlation power map, and the CNN module is composed of 2 convolutional layers, 2 pooling layers and 1 flattening layer: the 2 convolutional layers contain 32 3×3 filters and 64 3×3 filters respectively, which are used to perform a convolution operation on the single delay-Doppler correlation power map to obtain a first feature map; the 2 pooling layers both use a maximum pooling operation to perform a maximum pooling operation on the first feature map to obtain a pooled feature map, the pooling window size is 2×2, and the step size is 2; the flattening layer is used to flatten the pooled feature map into a first feature sequence with a length of 128; The auxiliary feature processing module includes three fully connected layers, each of which is composed of 64 neurons, 128 neurons, and 256 neurons, respectively. The auxiliary feature processing module is used to extract features of auxiliary parameters such as incident angle, normalized bistatic radar cross section, and DDM front slope to obtain a second feature sequence. The feature fusion layer is composed of two fully connected layers, each of which contains 512 and 256 neurons, respectively, and is used to concatenate and fuse the first feature sequence and the second feature sequence to obtain a fused feature; The output layer is a fully connected layer, which is used to map the fused features into wind speed prediction values; The inversion model after the introduction of the Jacobian matrix adds: sensitivity weighted convolution layer, custom attention mechanism and dual attention fusion module; Finally, define the loss function.

[0015] According to the present invention, the sensitivity weighted convolution layer includes a sensitivity mask generation network and a dynamic feature modulation mechanism: the sensitivity mask generation network is composed of two layers of convolution networks, the first layer is a 1×1 convolution layer, the number of input channels is the number of Jacobian matrix channels, the number of output channels is 32, and it is used for channel mapping; the second layer is a 3×3 convolution layer, the number of input channels is 32, and the number of output channels is the same as the number of original feature channels, which is used to capture spatial correlation; the dynamic feature modulation mechanism includes a learnable scaling factor, which is used to dynamically adjust the generated sensitivity mask, thereby weighting the input features; The sensitivity mask is generated from the Jacobian matrix through a two-layer convolutional network: (8) Where J is the input Jacobian matrix, C1 is the number of channels of the first layer of convolution, which is set to 32, C2 is the number of channels of the second layer of convolution, which is set to the number of input channels, k is the size of the second layer of convolution kernel, which is set to 3. ReLU is the activation function; Dynamic feature modulation mechanism: (9) Where X is the input feature, α is a learnable scaling factor initialized to 1.0, and * represents term-by-term multiplication.

[0016] Preferably, according to the present invention, the custom attention mechanism consists of a multi-head attention structure and Jacobi-guided attention weights: the multi-head attention structure divides the input features into 4 attention heads, and each attention head independently calculates the attention weight to capture the feature correlation of different subspaces; the Jacobi-guided attention weight converts the Jacobi matrix information into attention modulation coefficients through a two-layer perceptron, the first layer output dimension is 64, and the second layer output dimension is the number of attention heads, which is used to dynamically adjust the attention distribution; Jacobian-guided attention weights: Att_weight = σ(MLP(J))(10) Att_out=Attention(Q,K,V) * (1+β* Att_weight) (11) in, is the sigmoid activation function, are scientific parameters. Q (Query), K (Key), and V (Value) are the three key components in the attention mechanism, namely query matrix, key matrix, and value matrix. MLP is a multi-layer perceptron: the first layer parameter amount = input_dim × 64 + 64 (bias term), the second layer parameter amount = 64×num_heads+num_heads (bias term), input_dim is the feature dimension of the Jacobian matrix, and 64 neurons are selected in the middle for feature transformation and dimensionality upgrade. The output is the attention weight of num_heads dimension. The main function of MLP is to convert the information of the Jacobian matrix into the weight coefficient of each head in the attention mechanism, thereby realizing sensitivity-based attention modulation.

[0017] Preferably, according to the present invention, the dual attention fusion module includes channel attention and spatial attention: channel attention calculates the importance weight of each channel through global average pooling and two-layer perceptron, the first layer of perceptron reduces the number of channels by 16 times, and the second layer of perceptron restores the original number of channels; Spatial attention uses a 7×7 convolutional layer to generate an attention map of the spatial dimension, with an output channel of 1. The outputs of the two attentions are applied to the original features through element-by-element multiplication to achieve feature enhancement.

[0018] Preferably, according to the present invention, the loss function is composed of a basic loss term, a sensitivity weighted term and a regularization term: the basic loss term uses the mean square error; the sensitivity weighted term calculates the sample weight based on the absolute value of the Jacobian matrix; the regularization term uses L2 regularization to prevent overfitting; The loss function is: (12) in, is the sensitivity weighted loss, is the absolute value of the Jacobian matrix, is the L2 regularization term, is the L2 norm of the weight, that is, the square root of the sum of squares of ownership, W is the weight parameter of the model, is the weight coefficient of sensitivity loss, set to 0.1, is the weight coefficient of L2 regularization, set to 0.001, and MSE is the mean square error: (13) in, is the wind speed value predicted by the model, is the actual wind speed value, is the sample size.

[0019] Preferably, according to the present invention, in step (3), the model inversion result is quantified by the following formula: (14) (15) in, is the predicted wind speed, is the reference wind speed from ERA5, and the overall RMSE and bias are calculated based on the total The root mean square error is determined by samples. It is applicable to all inversion architectures and MVE traditional models. RMSE (root mean square error) reflects the average deviation between the model prediction value and the true value. The smaller the value, the higher the model prediction accuracy. Bias (bias) reflects the systematic deviation of the predicted value relative to the true value, which is used to determine whether the model has a tendency to overestimate or underestimate. RMSE reflects the overall accuracy of the prediction, and Bias reflects the systematic error of the prediction.

[0020] The sea surface wind speed inversion system based on the GNSS-R forward model includes: Modeling module: used in the GNSS-R receiver to cross-correlate the reflected signal with the local copy of the transmitted signal within the delay and Doppler frequency range, determine the model input, construct the DDM forward model based on the bistatic scattering model, and calculate the Jacobian matrix; Data preparation module, used to prepare inversion model data, including CYGNSS L1 GNSS-R data, ERA5 wind speed and direction, extract observation variable parameters from GNSS-R data, perform data quality control, data filtering and model data set construction, and complete time-space matching; The inversion construction module is used to combine the Jacobian matrix output by the DDM forward model, construct the inversion model, and evaluate and analyze the inversion results of the model with the Jacobian matrix introduced.

[0021] The beneficial effects of the present invention are: The present invention first obtains the modeling DDM (Delay-Doppler Map, reflection delay Doppler map) and Jacobian matrix by inputting four types of CYGNSS data, and then performs feature extraction. After data quality screening, multiple features related to sea surface wind speed are obtained, and then the wind speed is matched in time and space and input into the inversion model. Finally, it is concluded that the distribution of DDM generated by the DDM forward model and the observed DDM is relatively concentrated, and the overall simulation effect is good. The relative error value of the sample is basically maintained below 20%, and the distribution of the correlation coefficient is mainly concentrated between 0.94 and 0.98, most of the values ​​are close to 1, and only a small number of samples have values ​​lower than 0.92. The correlation coefficient will decrease under high wind speed conditions. In summary, it can be concluded that the inversion model performs well, the performance of the model on most samples is very close to the observed data, and the modeled DDM is basically consistent with the DDM measured by CYGNSS. The inversion model also improves the RMSE of the model that does not introduce the Jacobian matrix by 16%, from 1.875m / s to 1.572m / s. The present invention solves the problem that the neural network lacks physical interpretability and the inversion process is complicated, and provides a new idea for the subsequent inversion of soil moisture and other indicators. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 This is a comparison diagram of the modeled DDM and observed DDM of the present invention, where: Figure 2 (a) is the observed DDM diagram, Figure 2 (b) is the modeling DDM diagram. DETAILED DESCRIPTION

[0023] The present invention will be further described below by way of embodiments in conjunction with the accompanying drawings, but is not limited thereto.

[0024] Embodiment 1: like Figure 1 As shown, this embodiment provides a sea surface wind speed inversion method based on the GNSS-R forward model, and the steps are as follows: (1) In the GNSS-R receiver, the reflected signal is first cross-correlated with a local replica of the transmitted signal within the delay and Doppler frequency range to determine the model input, construct a DDM forward model based on the bistatic scattering model, and calculate the Jacobian matrix; There are four types of model inputs, namely: geometry, including the positions and velocities of the transmitter, receiver, and specular reflection points; Metadata, including sampling time, GPS pseudo-random noise (PRN) code, and specular reflection point index; Power, including the effective isotropic radiated power (EIRP) of the GPS transmitter and the receiver’s antenna parameters; Wind field, a gridded wind field expressed in latitude and longitude coordinates.

[0025] Among them, geometry, metadata and power are all from CYGNSS level 1 data, and the gridded wind field is constructed by the cross-calibrated multi-platform ocean surface wind field product CCMP dataset.

[0026] The DDM forward model construction process is: The bistatic scattering model formula is: (1) Among them, the integral variable is the two-dimensional position vector defining the point on the sea surface, represents the coherent integration time of the receiver, is the wavelength of the GPS carrier, is The wind speed at is the GPS EIRP (effective isotropic radiated power), which is assumed to remain constant throughout the scattering region. denote the gains of the receiver antenna and transmitter antenna on the surface, respectively, Indicates the transmit power, is the additional rainfall attenuation, Respectively indicate points The path distance from the mean sea surface at to the transmitter and receiver, Surface weighted area for a given delay and Doppler; During the simulation, the semi-empirical Katzberg model is used to convert the wind speed into the mean square slope of the sea surface. In E2ES in the DDMI of CYGNSS, equation (1) is rewritten as a two-dimensional convolution with a delta function to calculate the expected value of the DDM. (2) For fast computation, the Fast Fourier Transform (FFT) method is applied to the two-dimensional convolution: (3) In the fuzziness function and

[0027] Between is the Dirac delta function, In the above formula A brief description of the representation, ** represents the convolution operation, and ,in, is the expected value of the power contribution at each location on the surface, is the mapping of this power from surface coordinates to delay-Doppler coordinates.

[0028] To calculate the above equation (3), the delay Doppler space needs to be discretized. The subsequent Jacobian matrix needs to be inverted together with the observed value. Therefore, the same delay Doppler coordinates as the observed DDM need to be used to accurately reflect the authenticity of the data. The mirror point position estimate generated by DDMI on CYGNSS is very rough and fixed as a discrete value. Therefore, the real mirror reflection point is usually not located at the center of a single DDM bin. The fractionally accurate mirror reflection index generated by using and , the delay-Doppler coordinates of the mobile modeled DDM to align with the observed DDM and other observed values, are discretized, and calculated using a 2D fast Fourier transform; (4) in, , , is the index in discrete delay-Doppler space, and are the delay and Doppler increment, and for CYGNSS, , , , is the second cycle of the GPSC / A code chip, Hz, is the discretized form of the surface integral.

[0029] The Jacobian matrix H is the forward operator with respect to the input vector The partial derivative matrix of each wind speed in is expressed as follows: (5) The Jacobian matrix reflects the sensitivity of each sample of the DDM to the wind speed at each surface grid point.

[0030] In order to improve the physical interpretability of the model and overcome the limited accuracy improvement of deep learning methods, the GNSS-R basic model is modeled and the output is a matrix that can represent the sensitivity of each sample of DDM to the wind speed of each surface grid point. Some positions in the matrix are zero because these positions correspond to the parts outside the flash area centered on the mirror reflection point. The dispersion power of some functions is very small and is gradually submerged by noise. The actual value of all Jacobian matrices is zero. The high absolute value in the matrix represents the sensitivity of DDM to the wind speed at a specific location within the relative delay and Doppler range. The two dimensions of the matrix are DDM values ​​and surface wind speed, where DDM is expanded into a vector using delay and Doppler, and surface wind speed is expanded into a vector using latitude and longitude. In order to verify the availability of the model, the data of the model's input variables are obtained from the CYGNSS L1a product at a rate of 1 Hz. The receiver antenna pattern estimated by pre-launch measurements and on-orbit corrections is provided by the CYGNSS SOC. At each observation moment, the simulated DDM generated by the forward operator is compared with the DDM obtained from the L1a data. Two metrics are used to quantify the consistency between the modeled DDM (the DDM generated by the DDM forward model calculation) and the observed DDM (the observed DDM is the DDM obtained using the delayed Doppler imager processing in the CYGNSS L1 data), namely, the relative difference of the effective bins and the correlation coefficient between the two DDM effective bins, as follows:

[0031] (6) in, is the index of DDM, and is the observed DDM and the modeled DDM, and Represent the indexes of delay and Doppler shift respectively, N is the number of valid bins, It is to observe DDM in the delay and Doppler shift The power value at is to model DDM in the delay and Doppler shift The power value at It is Delayed sampling points, It is Doppler shift sampling points; (7) in, For the The correlation coefficient of the DDM, and is the average of all valid bins of the observed DDM and the modeled DDM, and Represent the indexes of delay and Doppler shift respectively, It is to observe DDM in the delay and Doppler shift The power value at is to model DDM in the delay and Doppler shift The power value at It is Delayed sampling points, It is The more similar the shapes of the observed DDM and the modeled DDM are, the closer the correlation coefficient is to 1.

[0032] The comparison results are as follows Figure 2 As shown in the figure, the simulated DDM and Jacobian matrix are obtained through the DDM forward model. The model is evaluated and found that the relative difference of DDM is less than 30% and the correlation coefficient is greater than 90%, which ensures the consistency of the model. It also shows the corresponding relationship between the Jacobian matrix and DDM under different wind speeds. Ideally, the main reason for the difference in results in the DDM forward model is the difference between the model wind field and the real wind. In addition, there are some other uncertain factors that affect the performance of the model, including: low wind speed. In medium and high wind speed conditions, local wind plays a dominant role, so the influence of swell waves can be ignored, but in low wind speed conditions, the influence of swell waves is very obvious, the problem of inaccurate position of mirror reflection point, a more accurate mirror point calculation method and a higher fidelity sea surface model can be used to calculate the mirror reflection point delay and Doppler, and the inaccurately estimated GPS transmitter EIRP. The accuracy of its estimation will be affected by the change of transmitter power and the calibration of the zenith antenna pattern. The accuracy of the forward model can be improved by improving these aspects.

[0033] The modeling DDM and Jacobian matrix generated by the DDM forward model, where the specific variable names in the CYGNSS Level 1 data are involved, are as follows:

[0034] Based on the GPS bistatic radar scattering model, the DDM (in units of absolute power at the receiving end) is linked to the sea surface wind field.

[0035] The ground truth wind speed uses the ERA5 reanalysis data of the European Center for Medium-Range Weather Forecasts (ECMWF). The time resolution of the ERA5 wind speed product is 1h and the grid resolution is 0.5×0.5. The longitude of the ERA5 data is transformed to the range of 0-360, and an interpolation function is created for the wind field file. Then, the longitude and latitude of CYGNSS are used to interpolate in the ERA5 data. Finally, the obtained wind speed is saved in a new variable of the CYGNSS data file. The whole process ensures that each observation point of CYGNSS can obtain the corresponding ERA5 wind speed through time matching and spatial interpolation. High-quality DDM observation data is essential for the model to correctly and intuitively understand how to map different ocean wind speeds to their corresponding observations. Therefore, it is essential to eliminate low-quality measurements in the data, as follows: The uncertainty of the uncertainty of the BRCS (ddm_brcs_uncert) bistatic radar cross section is less than 1;

[0036] The Nanostar Tracker attitude status is "OK", that is nst_att_status=0, navigation science terminal attitude status; The absolute value of the spacecraft roll is between 1 and 30 degrees, the pitch is between 1 and 10, or the yaw is between 1 and 5, as indicated by the quality_flags quality flags of each DDM Specular reflection point (sp_rx_gain) The receiving antenna gain in the specular reflection point receiving antenna gain direction is greater than 0dBi.

[0037] DDM (prn_fig_of_merit) PRN selection quality assessment range correction gain (RCG) figure of merit (FOM) greater than 0 LES (ddm_les) DDM waveform leading edge slope is greater than 0 Zenith (direct) signal-to-noise ratio (direct_signal_snr) Zenith (direct) signal-to-noise ratio is greater than 0dB In addition, since CYGNSS measurements or ERA5 estimates may also be contaminated by unexpected errors that usually do not affect the main population, samples outside the 95% confidence interval are removed.

[0038] (2) Prepare inversion model data, including CYGNSS L1 GNSS-R data, ERA5 wind speed and direction, extract observation variable parameters from GNSS-R data, perform data quality control, data filtering and model data set construction, and complete spatiotemporal matching; (3) Combined with the Jacobian matrix output by the DDM forward model, an inversion model is constructed, and the inversion results of the model with the Jacobian matrix introduced are evaluated and analyzed.

[0039] The inversion model architecture is: CNN module, auxiliary feature processing module, feature fusion layer and output layer. The CNN module is used to extract the features of a single delay-Doppler correlation power map. The CNN module consists of 2 convolutional layers, 2 pooling layers and 1 flattening layer: the 2 convolutional layers contain 32 3×3 filters and 64 3×3 filters respectively, which are used to perform convolution operation on the single delay-Doppler correlation power map to obtain a first feature map; the 2 pooling layers both use maximum pooling operation to perform maximum pooling operation on the first feature map to obtain a pooled feature map, the pooling window size is 2×2, and the step size is 2; the flattening layer is used to flatten the pooled feature map into a first feature sequence with a length of 128; The auxiliary feature processing module includes three fully connected layers, each of which is composed of 64 neurons, 128 neurons, and 256 neurons, respectively. The auxiliary feature processing module is used to extract features of auxiliary parameters such as incident angle, normalized bistatic radar cross section, and DDM front slope to obtain a second feature sequence. The feature fusion layer is composed of two fully connected layers, each of which contains 512 and 256 neurons, respectively, and is used to concatenate and fuse the first feature sequence and the second feature sequence to obtain a fused feature; The output layer is a fully connected layer, which is used to map the fused features into wind speed prediction values; The inversion model after the introduction of the Jacobian matrix adds: sensitivity weighted convolution layer, custom attention mechanism and dual attention fusion module; Finally, define the loss function.

[0040] The sensitivity weighted convolution layer includes a sensitivity mask generation network and a dynamic feature modulation mechanism: the sensitivity mask generation network consists of two layers of convolutional networks. The first layer is a 1×1 convolutional layer with the same number of input channels as the Jacobian matrix channels and 32 output channels for channel mapping; the second layer is a 3×3 convolutional layer with 32 input channels and the same number of output channels as the original feature channels for capturing spatial correlation; the dynamic feature modulation mechanism includes a learnable scaling factor for dynamically adjusting the generated sensitivity mask and weighting the input features; The sensitivity mask is generated from the Jacobian matrix through a two-layer convolutional network: (8) Where J is the input Jacobian matrix, C1 is the number of channels of the first layer of convolution, which is set to 32, C2 is the number of channels of the second layer of convolution, which is set to the number of input channels, k is the size of the second layer of convolution kernel, which is set to 3. ReLU is the activation function; Dynamic feature modulation mechanism: (9) Where X is the input feature, α is a learnable scaling factor initialized to 1.0, and * represents term-by-term multiplication.

[0041] Preferably, according to the present invention, the custom attention mechanism consists of a multi-head attention structure and Jacobi-guided attention weights: the multi-head attention structure divides the input features into 4 attention heads, and each attention head independently calculates the attention weight to capture the feature correlation of different subspaces; the Jacobi-guided attention weight converts the Jacobi matrix information into attention modulation coefficients through a two-layer perceptron, the first layer output dimension is 64, and the second layer output dimension is the number of attention heads, which is used to dynamically adjust the attention distribution; Jacobian-guided attention weights: Att_weight = σ(MLP(J))(10) Att_out=Attention(Q,K,V) * (1+β* Att_weight) (11) in, is the sigmoid activation function, are scientific parameters. Q (Query), K (Key), and V (Value) are the three key components in the attention mechanism, namely query matrix, key matrix, and value matrix. MLP is a multi-layer perceptron: the first layer parameter amount = input_dim × 64 + 64 (bias term), the second layer parameter amount = 64×num_heads+num_heads (bias term), input_dim is the feature dimension of the Jacobian matrix, and 64 neurons are selected in the middle for feature transformation and dimensionality upgrade. The output is the attention weight of num_heads dimension. The main function of MLP is to convert the information of the Jacobian matrix into the weight coefficient of each head in the attention mechanism, thereby realizing sensitivity-based attention modulation.

[0042] The dual attention fusion module includes channel attention and spatial attention: channel attention calculates the importance weight of each channel through global average pooling and two-layer perceptron. The first layer of perceptron reduces the number of channels by 16 times, and the second layer of perceptron restores the original number of channels. Spatial attention uses a 7×7 convolutional layer to generate an attention map of the spatial dimension, with an output channel of 1. The outputs of the two attentions are applied to the original features through element-by-element multiplication to achieve feature enhancement.

[0043] The loss function consists of a basic loss term, a sensitivity weighted term, and a regularization term: the basic loss term uses mean square error; the sensitivity weighted term calculates the sample weight based on the absolute value of the Jacobian matrix; the regularization term uses L2 regularization to prevent overfitting; The loss function is: (12) in, is the sensitivity weighted loss, is the absolute value of the Jacobian matrix, is the L2 regularization term, is the L2 norm of the weight, that is, the square root of the sum of squares of ownership, W is the weight parameter of the model, is the weight coefficient of sensitivity loss, set to 0.1, is the weight coefficient of L2 regularization, set to 0.001, and MSE is the mean square error: (13) in, is the wind speed value predicted by the model, is the actual wind speed value, is the sample size.

[0044] The evaluation of the inversion model was conducted during the test period from May 1 to July 2017. Since the research focus is on the effect of the introduction of physical constraints, several one-dimensional variable models of DDM variables are selected for effect feedback. The model inversion results are quantified by the following formula: (14) (15) in, is the predicted wind speed, is the reference wind speed from ERA5, and the overall RMSE and bias are calculated based on the total The root mean square error is determined by samples. It is applicable to all inversion architectures and MVE traditional models. RMSE (root mean square error) reflects the average deviation between the model prediction value and the true value. The smaller the value, the higher the model prediction accuracy. Bias (bias) reflects the systematic deviation of the predicted value relative to the true value, which is used to determine whether the model has a tendency to overestimate or underestimate. RMSE reflects the overall accuracy of the prediction, and Bias reflects the systematic error of the prediction.

[0045] The research data mainly comes from the CYGNSS project and is verified by the ERA5 wind speed product. The main comparison is the change in the accuracy of the model inversion before and after the introduction of the Jacobian matrix. The results show that the new model inversion wind shows good consistency with the ERA5 wind speed product.

[0046] Embodiment 2: This embodiment provides a sea surface wind speed inversion system based on a GNSS-R forward model, including: Modeling module: used in the GNSS-R receiver to cross-correlate the reflected signal with the local copy of the transmitted signal within the delay and Doppler frequency range, determine the model input, construct the DDM forward model based on the bistatic scattering model, and calculate the Jacobian matrix; Data preparation module, used to prepare inversion model data, including CYGNSS L1 GNSS-R data, ERA5 wind speed and direction, extract observation variable parameters from GNSS-R data, perform data quality control, data filtering and model data set construction, and complete time-space matching; The inversion construction module is used to combine the Jacobian matrix output by the DDM forward model, construct the inversion model, and evaluate and analyze the inversion results of the model with the Jacobian matrix introduced.

[0047] The above embodiments are only used to help understand the method and core idea of ​​the present invention; at the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

[0048] The above embodiments are used to illustrate the present invention rather than to limit the present invention. Any modification and change made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. The sea surface wind speed inversion method based on the GNSS-R forward model is characterized by: Here are the steps: (1) In the GNSS-R receiver, the reflected signal is first cross-correlated with a local replica of the transmitted signal within the delay and Doppler frequency range to determine the model input, construct a DDM forward model based on the bistatic scattering model, and calculate the Jacobian matrix; (2) Prepare inversion model data, including CYGNSS L1 GNSS-R data, ERA5 wind speed and direction, extract observation variable parameters from GNSS-R data, perform data quality control, data filtering and model data set construction, and complete spatiotemporal matching; (3) Combined with the Jacobian matrix output by the DDM forward model, an inversion model is constructed, and the inversion results of the model with the Jacobian matrix introduced are evaluated and analyzed.

2. The sea surface wind speed inversion method based on the GNSS-R forward model according to claim 1, characterized in that: In step (1), the model inputs include four types, namely: geometry, including the positions and velocities of the transmitter, receiver, and specular reflection points; Metadata, including sampling time, GPS pseudo-random noise code, and specular reflection point index; Power, including the effective isotropic radiated power of the GPS transmitter and the antenna parameters of the receiver; Wind field, a gridded wind field expressed in latitude and longitude coordinates.

3. The sea surface wind speed inversion method based on the GNSS-R forward model as claimed in claim 2, characterized in that: In step (1), the DDM forward model construction process is: The bistatic scattering model formula is: （1） Among them, the integral variable is the two-dimensional position vector defining the point on the sea surface, represents the coherent integration time of the receiver, is the wavelength of the GPS carrier, is The wind speed at is the GPS EIRP (effective isotropic radiated power), which is assumed to remain constant throughout the scattering region. denote the gains of the receiver antenna and transmitter antenna on the surface, respectively, Indicates the transmit power, is the additional rainfall attenuation, Respectively indicate points The path distance from the mean sea surface at to the transmitter and receiver, Surface weighted area for a given delay and Doppler; During the simulation, the semi-empirical Katzberg model is used to convert the wind speed into the surface MSS. In E2ES in the DDMI of CYGNSS, equation (1) is rewritten as a two-dimensional convolution with a delta function to calculate the expected value of the DDM; （2） For fast computation, the Fast Fourier Transform (FFT) method is applied to the two-dimensional convolution: （3） In the fuzziness function and Between is the Dirac delta function, ** denotes the convolution operation, and ,in, is the expected value of the power contribution at each location on the surface, is the mapping of this power from surface coordinates to delay-Doppler coordinates; To calculate the above equation (3), the delay-Doppler space needs to be discretized and calculated using a two-dimensional fast Fourier transform; （4） in, , , is the index in discrete delay-Doppler space, and are the delay and Doppler increment, and for CYGNSS, , , , is the second cycle of the GPSC / A code chip, Hz, is the discretized form of the surface integral.

4. The sea surface wind speed inversion method based on the GNSS-R forward model as claimed in claim 3, characterized in that: In step (1), the Jacobian matrix H is the forward operator with respect to the input vector The partial derivative matrix of each wind speed in is expressed as follows: （5） The Jacobian matrix reflects the sensitivity of each sample of the DDM to the wind speed at each surface grid point.

5. The sea surface wind speed inversion method based on the GNSS-R forward model as claimed in claim 4, characterized in that: In step (1), two metrics are used to quantify the consistency between the modeled DDM and the observed DDM, namely, the relative difference of the valid bins and the correlation coefficient between the two DDM valid bins, as follows: （6） in, is the index of DDM, and is the observed DDM and the modeled DDM, and Represent the indexes of delay and Doppler shift respectively, N is the number of valid bins, It is to observe DDM in the delay and Doppler shift The power value at is to model DDM in the delay and Doppler shift The power value at It is Delayed sampling points, It is Doppler shift sampling points; （7） in, For the The correlation coefficient of the DDM, and is the average of all valid bins of the observed DDM and the modeled DDM, and Represent the indexes of delay and Doppler shift respectively, It is to observe DDM in the delay and Doppler shift The power value at is to model DDM in the delay and Doppler shift The power value at It is Delayed sampling points, It is The more similar the shapes of the observed DDM and the modeled DDM are, the closer the correlation coefficient is to 1.

6. The sea surface wind speed inversion method based on the GNSS-R forward model as claimed in claim 5, characterized in that: In step (3), the inversion model architecture is: CNN module, auxiliary feature processing module, feature fusion layer and output layer. The CNN module is used to extract the features of a single delay-Doppler correlation power map. The CNN module consists of 2 convolutional layers, 2 pooling layers and 1 flattening layer. The auxiliary feature processing module includes three fully connected layers, and is used to extract features from auxiliary parameters to obtain a second feature sequence; The feature fusion layer is composed of two fully connected layers, which are used to concatenate and fuse the first feature sequence and the second feature sequence to obtain a fusion feature; The output layer is a fully connected layer, which is used to map the fused features into wind speed prediction values; The inversion model after the introduction of the Jacobian matrix adds: sensitivity weighted convolution layer, custom attention mechanism and dual attention fusion module; Finally, define the loss function.

7. The sea surface wind speed inversion method based on the GNSS-R forward model as claimed in claim 6, characterized in that: The sensitivity weighted convolution layer includes a sensitivity mask generation network and a dynamic feature modulation mechanism: the sensitivity mask generation network consists of two layers of convolutional networks. The first layer is a 1×1 convolutional layer, the number of input channels is the number of Jacobian matrix channels, and the number of output channels is 32, which is used for channel mapping; the second layer is a 3×3 convolutional layer, the number of input channels is 32, and the number of output channels is the same as the number of original feature channels, which is used to capture spatial correlation; The dynamic feature modulation mechanism includes a scaling factor to dynamically adjust the generated sensitivity mask to weight the input features; The sensitivity mask is generated from the Jacobian matrix through a two-layer convolutional network: （8） Where J is the input Jacobian matrix, C1 is the number of channels of the first layer of convolution, which is set to 32, C2 is the number of channels of the second layer of convolution, which is set to the number of input channels, k is the size of the second layer of convolution kernel, which is set to 3. ReLU is the activation function; Dynamic feature modulation mechanism: （9） Where X is the input feature, α is a learnable scaling factor, initialized to 1.0, and * represents term-by-term multiplication; The custom attention mechanism consists of a multi-head attention structure and Jacobi-guided attention weights: the multi-head attention structure divides the input features into four attention heads, and each attention head independently calculates the attention weight to capture the feature correlation of different subspaces; the Jacobi-guided attention weights convert the Jacobi matrix information into attention modulation coefficients through a two-layer perceptron. The output dimension of the first layer is 64, and the output dimension of the second layer is the number of attention heads, which is used to dynamically adjust the attention distribution; Jacobian-guided attention weights: Att_weight = σ(MLP(J))(10) Att_out=Attention(Q,K,V) * (1+β* Att_weight) (11) in, is the sigmoid activation function, are scientific parameters, Q, K, and V are the three key components in the attention mechanism, namely, query matrix, key matrix, and value matrix, respectively, and MLP is a multi-layer perceptron; The dual attention fusion module includes channel attention and spatial attention: channel attention calculates the importance weight of each channel through global average pooling and two-layer perceptron. The first layer of perceptron reduces the number of channels by 16 times, and the second layer of perceptron restores the original number of channels. Spatial attention uses a 7×7 convolutional layer to generate an attention map of the spatial dimension, with an output channel of 1. The outputs of the two attentions are applied to the original features through element-by-element multiplication to achieve feature enhancement.

8. The sea surface wind speed inversion method based on the GNSS-R forward model as claimed in claim 7, characterized in that: The loss function consists of a basic loss term, a sensitivity weighted term, and a regularization term: the basic loss term uses mean square error; the sensitivity weighted term calculates the sample weight based on the absolute value of the Jacobian matrix; the regularization term uses L2 regularization to prevent overfitting; The loss function is: （12） in, is the sensitivity weighted loss, is the absolute value of the Jacobian matrix, is the L2 regularization term, is the L2 norm of the weight, that is, the square root of the sum of squares of ownership, W is the weight parameter of the model, is the weight coefficient of sensitivity loss, set to 0.1, is the weight coefficient of L2 regularization, set to 0.001, and MSE is the mean square error: （13） in, is the wind speed value predicted by the model, is the actual wind speed value, is the sample size.

9. The sea surface wind speed inversion method based on the GNSS-R forward model as claimed in claim 8, characterized in that: In step (3), the model inversion results are quantified by the following formula: （14） （15） in, is the predicted wind speed, It is the reference wind speed from ERA5. RMSE reflects the average deviation between the model prediction value and the true value. The smaller the value, the higher the model prediction accuracy. Bias reflects the systematic deviation of the predicted value relative to the true value. It is used to determine whether the model has a tendency to overestimate or underestimate. RMSE reflects the overall accuracy of the prediction, and Bias reflects the systematic error of the prediction.

10. The sea surface wind speed inversion system based on the GNSS-R forward model is characterized by: include: Modeling module: used in the GNSS-R receiver to cross-correlate the reflected signal with the local copy of the transmitted signal within the delay and Doppler frequency range, determine the model input, construct the DDM forward model based on the bistatic scattering model, and calculate the Jacobian matrix; Data preparation module, used to prepare inversion model data, including CYGNSS L1 GNSS-R data, ERA5 wind speed and direction, extract observation variable parameters from GNSS-R data, perform data quality control, data filtering and model data set construction, and complete time-space matching; The inversion construction module is used to combine the Jacobian matrix output by the DDM forward model, construct the inversion model, and evaluate and analyze the inversion results of the model with the Jacobian matrix introduced.

Images