Same-orbit time sequence sea surface wind speed inversion method and system combining GNSS-R and scatterometer

Abstract

The invention belongs to the technical field of ocean data processing, and discloses a GNSS-R and scatterometer combined same-orbit time sequence sea surface wind speed inversion method and system, and the method comprises the following steps: obtaining historical satellite data of ocean surface wind, carrying out the preprocessing of the historical satellite data, and obtaining the preprocessed data, constructing a training sample set in combination with the reference wind speed and the surge auxiliary data; constructing an initial wind speed inversion model by using a plurality of single-moment multi-frequency information fusion modules and multi-moment interactive attention enhancement modules, and training the initial wind speed inversion model by using the training sample set to obtain a same-orbit time sequence sea surface wind speed inversion model; and obtaining a time delay-Doppler correlation power diagram, GNSS-R observation parameters, scatterometer parameters and effective wave height of storm and surge, and completing sea surface wind speed inversion by using the same-orbit time sequence sea surface wind speed inversion model to obtain sea surface wind speed prediction data.

Description

Technical Field

[0001] The present invention belongs to the technical field of marine data processing, and particularly relates to a method and system for retrieving the sea surface wind speed in the same orbit and in time series by combining GNSS-R and a scatterometer. Background Art

[0002] The L-band signal used by GNSS satellites has strong penetration ability and is less affected by clouds, fog, rain and snow, and can be used as a free signal source for earth observation. When the GNSS signal is reflected, it is affected by the earth's surface, resulting in changes in aspects such as the signal intensity, time delay, and frequency. The sea surface wind field is one of the important factors affecting the sea surface roughness. The sea surface roughness is different under different wind speed conditions. The difference in roughness will cause a change in the reflectivity, which will in turn cause a change in the reflected signal power and frequency, and finally be captured by the receiver. The GNSS-R wind speed inversion is achieved by using the relationship of "wind field -> roughness -> reflectivity -> signal characteristics".

[0003] The current GNSS-R wind speed inversion can be divided into three categories: (1) Wind speed inversion method based on waveform matching: Waveform matching is one of the common methods for GNSS-R wind speed inversion. The implementation process of this method is to first obtain a series of DDMs through field measurement or numerical simulation and record the corresponding wind speed values to construct a "DDM - wind speed" reference library for matching. Then, the DDM with unknown wind speed is compared (matched) with each DDM in the reference library, and the wind speed corresponding to the reference DDM with the smallest difference is used as the inverted wind speed. However, this method requires the establishment of a fine theoretical waveform library, with a large amount of calculation, long time consumption, and a cumbersome inversion process. (2) Wind speed inversion method based on a geographical model function: The wind speed inversion method based on an empirical function mainly performs wind speed inversion by establishing a mapping relationship between GNSS-R observables and wind speed. This method is also one of the common methods for GNSS-R wind speed inversion. The empirical function used here is a geographical model function (GMF), including forms such as exponential functions, polynomial functions, and 2D lookup tables (LUTs). However, this method mainly relies on a single observation value in the DDM and it is difficult to comprehensively consider the influence of reflection geometric parameters. (3) Wind speed inversion method based on machine learning (deep learning): Compared with the empirical function model, the machine learning model has a stronger non-linear mapping ability. This type of method mainly uses DDM data and its derived features as the main inputs, and mines deep features through models such as random forests, BP neural networks, and convolutional neural networks to establish a mapping relationship between the input and the wind speed. Although this method has strong non-linear fitting ability, it has problems such as insufficient single-source feature information, failure to fuse multi-source remote sensing data, and mostly based on single-moment modeling, ignoring the time series characteristics of multi-moment observations along the orbit, which limits the inversion accuracy and the model generalization ability.

[0004] To solve the above problems, the present invention extracts multi-frequency information based on DDM characteristics, fuses multi-source observation data, and performs time-series modeling in combination with an interactive attention mechanism, aiming to improve the accuracy and robustness of sea surface wind speed inversion. Summary of the Invention

[0005] The present invention aims to solve the deficiencies of the prior art and provides the following solutions:

[0006] A co-orbit time-series sea surface wind speed inversion method combining GNSS-R and scatterometer includes the following steps:

[0007] Obtain historical satellite data of the ocean surface wind, preprocess the historical satellite data to obtain preprocessed data, and construct a training sample set in combination with reference wind speed and swell auxiliary data;

[0008] Use a number of single-moment multi-frequency information fusion modules and multi-moment interactive attention enhancement modules to construct an initial wind speed inversion model, and use the training sample set to train the initial wind speed inversion model to obtain a co-orbit time-series sea surface wind speed inversion model;

[0009] Obtain the delay-Doppler correlation power map, GNSS-R observation parameters, scatterometer parameters, and the effective wave height of wind waves and swells, and use the co-orbit time-series sea surface wind speed inversion model to complete the sea surface wind speed inversion to obtain sea surface wind speed prediction data.

[0010] Preferably, the method for preprocessing the historical satellite data includes:

[0011] Calculate the range gain correction through the antenna gain at the specular reflection point, the distance from the receiver to the specular reflection point, and the distance from the navigation satellite to the specular reflection point, and correct the historical satellite data for range gain correction:

[0012] ,

[0013] where RCG represents the range gain correction, G SP represents the antenna gain in the direction of the specular emission point, represents the distance from the receiver to the specular reflection point, R SP represents the distance from the navigation satellite to the specular reflection point, 10 27 represents the scaling factor;

[0014] Perform incidence angle correction on the corrected data, and screen out the data with an incidence angle less than a preset threshold in the corrected data;

[0015] Perform specular reflection point type correction on the filtered data. In the specular reflection point types: 0 represents open sea area, 0.5 represents coastal sea area within 25 kilometers from land, 1 represents land, and 2 represents sea ice. Filter out the data with specular reflection point type less than 1 to obtain the data after type correction;

[0016] Perform multi-look average operation on the wind field measurement radar data, extract the backscattering coefficients of the wind field measurement radar data in C-band HH polarization and VV polarization, and integrate the data after type correction and the scattering coefficients to obtain the preprocessed data.

[0017] Preferably, the method for constructing the training sample set includes:

[0018] Select the U component and V component of the 10-meter wind field measured by ECMWF, and calculate the reference wind speed based on the U component and the V component:

[0019] ,

[0020] where wspd represents the reference wind speed, u 10 represents the U component of the 10-meter wind field, and v 10 represents the V component of the 10-meter wind field;

[0021] Select the significant wave height of wind waves and swell waves as the swell auxiliary data, and perform bilinear interpolation on the reference wind speed and the swell auxiliary data to obtain the training sample set.

[0022] Preferably, the single-moment multi-frequency information fusion module includes: a wavelet transform unit, a global enhancement unit, a local enhancement unit, a BP unit, and a feature fusion unit;

[0023] The wavelet transform unit consists of a low-pass filter and a high-pass filter; the wavelet transform unit is used to perform two-dimensional wavelet decomposition on the time-delay Doppler correlation power map, divide the image into 1 low-frequency approximation sub-band and 3 high-frequency detail sub-bands, use the low-frequency approximation sub-band as the low-frequency DDM image, and perform weighted fusion on the 3 high-frequency detail sub-bands to obtain the high-frequency DDM image;

[0024] The global enhancement unit consists of patch embedding, positional encoding, layer normalization, multi-head self-attention, and a multi-layer perceptron; the patch embedding is used to convert the low-frequency DDM image and the high-frequency DDM image into a sequence of feature vectors; the positional encoding is used to introduce delay-Doppler coordinates consistent with the size of the feature vector sequence, and integrate the feature vector sequence with the delay-Doppler coordinates to obtain an input vector; the layer normalization is used to standardize the input vector to obtain a normalized vector; the multi-head self-attention processes the normalized vector based on the self-attention mechanism to obtain a number of input features; the multi-layer perceptron is used to perform linear transformation and non-linear activation on each input feature to obtain a global feature sequence;

[0025] The local enhancement unit includes a CNN network and a bidirectional feature enhancement sub-unit; the CNN is used to initially extract the local texture features of the low-frequency DDM image and the high-frequency DDM image, and the bidirectional feature enhancement sub-unit is used to perform global average pooling operation and weighting on the local texture features to obtain a local feature sequence;

[0026] The BP unit is used to extract features from GNSS-R observation parameters, scatterometer parameters, and the significant wave height of wind waves and swell waves, and perform weighted summation to obtain a multi-source parameter fusion feature;

[0027] The feature fusion unit is used to fuse the global feature sequence, the local feature sequence, and the multi-source parameter fusion feature to obtain a single-moment fusion feature.

[0028] Preferably, the multi-moment interactive attention enhancement module includes: a first bidirectional long short-term memory network, a second bidirectional long short-term memory network, and a fully connected layer;

[0029] The first bidirectional long short-term memory network is used to extract the time-dependent features of the single-moment fusion feature, and combine the multi-head self-attention mechanism to enhance the internal correlation between different time steps;

[0030] The second bidirectional long short-term memory network is used to further extract the deep temporal features of a number of the single-moment fusion features, and strengthen the interactive modeling between features of different time steps through the multi-head cross-attention mechanism;

[0031] The fully connected layer is used to fuse the time-dependent features and the deep temporal features to obtain a sea surface wind speed prediction result.

[0032] Preferably, the method for training the initial wind speed inversion model includes:

[0033] Initialize the parameters in the initial wind speed inversion model, and input the training sample set into the initial wind speed inversion model;

[0034] The training sample set obtains a predicted value through forward propagation and calculates a loss function;

[0035] The parameters in the initial wind speed inversion model are updated using the backpropagation algorithm and the Adam optimizer;

[0036] When the change in the continuously iterated loss function is less than 0.0001 or the set number of iterations is reached, the training ends, and the along-track multi-temporal sea surface wind speed inversion model is obtained.

[0037] Preferably, the method for calculating the loss function includes: constructing the loss function using the mean squared error loss function, direction consistency loss, and amplitude consistency loss:

[0038] ,

[0039] ,

[0040] ,

[0041] ,

[0042] where L Total represents the loss function, L MSE represents the mean squared error loss function, L Direction represents the direction consistency loss, L Trend represents the amplitude consistency loss, α, β, and γ represent weight coefficients, softplus represents the softplus function, and w1, w2, and w3 represent learnable parameters;

[0043] The calculation method of the mean squared error loss function includes:

[0044] ,

[0045] where n represents the number of samples, y i represents the true value of the i-th sample, and f i represents the inversion value of the i-th sample;

[0046] The calculation method of the direction consistency loss includes:

[0047] ,

[0048] ,

[0049] ,

[0050] where Δy i represents the difference in the true values of the samples, and Δf iDenote the difference of the sample inversion values, y i+1 Denote the true value of the (i + 1)-th sample, f i+1 Denote the inversion value of the (i + 1)-th sample;

[0051] The calculation method of the amplitude consistency loss includes:

[0052] ,

[0053] ,

[0054] ,

[0055] wherein, Var(Δy) denotes the variance of the sample true values, Var(Δf) denotes the variance of the sample inversion values, Denote the mean value of the sample true value differences, Denote the mean value of the sample inversion value differences.

[0056] Preferably, the update rule of the Adam optimizer includes:

[0057] ,

[0058] wherein, t denotes the number of iterations, g t Denote the current gradient, Denote the gradient of the loss function L with respect to the model parameter 𝜔, m t Denote the first moment estimate of the gradient, β1 denotes the decay rate controlling the first moment estimate, m t-1 Denote the first moment estimate at the previous iteration, v t Denote the second moment estimate of the gradient, β2 denotes the decay rate controlling the second moment estimate, Denote the result after correcting the estimate value of the first moment estimate, Denote the t-th power of the decay rate β1, Denote the result after correcting the estimate value of the second moment estimate, Denote the t-th power of the decay rate β1, Δω t Denote the update value of the model parameter 𝜔 at the t-th iteration, α1 denotes the learning rate, Denote the smoothing term.

[0059] The present invention also provides a co-orbit time-series sea surface wind speed inversion system combining GNSS-R and scatterometer. The inversion system applies the inversion method described in any one of the above, and includes: a data set construction module, a model construction module, and a prediction module;

[0060] The dataset construction module is used to obtain historical satellite data of ocean surface wind, preprocess the historical satellite data to obtain preprocessed data, and construct a training sample set in combination with reference wind speed and swell auxiliary data;

[0061] The model construction module uses a number of single - moment multi - frequency information fusion modules and multi - moment interactive attention enhancement modules to construct an initial wind speed inversion model, and trains the initial wind speed inversion model using the training sample set to obtain a same - orbit time - series sea surface wind speed inversion model;

[0062] The prediction module is used to obtain the time - delay - Doppler correlation power map, GNSS - R observation parameters, scatterometer parameters, and the significant wave height of wind waves and swells, and uses the same - orbit time - series sea surface wind speed inversion model to complete the inversion of sea surface wind speed to obtain sea surface wind speed prediction data.

[0063] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0064] In view of the structural characteristics of GNSS - R DDM images, the present invention designs a single - moment multi - frequency information fusion module (MIFM) to extract complementary features between high - and low - frequency DDM images and multi - source physical parameters. In addition, considering that space - borne observation data has spatio - temporal continuity within the orbit, to enhance the model's ability to model the time - evolution law of wind speed, on the basis of the traditional BiLSTM structure, the present invention further introduces self - attention and cross - attention mechanisms to construct a multi - moment interactive attention enhancement module (MIAE) to strengthen the interactive modeling and dynamic association between time - series features. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the technical solutions of the present invention, the following briefly introduces the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0066] Figure 1 It is a schematic flowchart of the method according to the embodiment of the present invention;

[0067] Figure 2 It is a schematic structural diagram of the initial wind speed inversion model according to the embodiment of the present invention;

[0068] Figure 3 It is the single - moment multi - frequency information fusion module according to the embodiment of the present invention;

[0069] Figure 4 It is a schematic structural diagram of the bidirectional feature enhancement sub - unit according to the embodiment of the present invention;

[0070] Figure 5Schematic diagram of the BP unit structure according to an embodiment of the present invention;

[0071] Figure 6 Multi-moment interaction attention enhancement module according to an embodiment of the present invention. Detailed implementation manners

[0072] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0073] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific implementation manners.

[0074] Embodiment 1

[0075] In this embodiment, as Figure 1 shown, the along-track sequential sea surface wind speed inversion method combining GNSS-R and scatterometer includes the following steps:

[0076] S1. Obtain historical satellite data of the ocean surface wind, preprocess the historical satellite data to obtain preprocessed data, and construct a training sample set in combination with reference wind speed and swell auxiliary data.

[0077] In this embodiment, first obtain the historical satellite data of the ocean surface wind of Tianmu-1 (TM), and preprocess the historical satellite data. The method for preprocessing the historical satellite data includes: calculating range gain correction through the antenna gain at the specular reflection point, the distance from the receiver to the specular reflection point, and the distance from the navigation satellite to the specular reflection point, and correcting the historical satellite data for range gain correction:

[0078] ,

[0079] where RCG represents range gain correction, G SP represents the antenna gain in the direction of the specular emission point, represents the distance from the receiver to the specular reflection point, R SP represents the distance from the navigation satellite to the specular reflection point, 10 27It represents the scaling factor, which is used to adjust the RCG to the range of 1 - 100 order of magnitude. To eliminate the data with low signal gain, in this paper, the data with RCG > 10 is selected for subsequent wind speed inversion to reduce the interference of low-quality data on the inversion result. Incidence angle correction is performed on the corrected data, and the data with incidence angle less than the preset threshold in the corrected data is screened out. In this embodiment, since the antenna gain decreases with the increase of the incidence angle and the quality of the reflected signal decreases accordingly, the data with incidence angle less than 60° is selected in data processing to ensure the signal quality. The data screened out is corrected for the type of specular reflection point. In the type of specular reflection point: 0 represents the open sea area, 0.5 represents the coastal sea area within 25 kilometers from the land, 1 represents the land, and 2 represents the sea ice. The data with the type of specular reflection point less than 1 is screened out to obtain the data after type correction; multi-look average operation is performed on the wind field measurement radar data, and the backscattering coefficients of the wind field measurement radar data in the C-band HH polarization and VV polarization are extracted. The data after type correction and the backscattering coefficients are integrated to obtain the preprocessed data.

[0080] To ensure the consistency and matching accuracy of multi-source data fusion, 15 minutes is selected as the time window for synchronization processing in this embodiment. The spatio-temporal matching conditions between the TM data and the FY data of the Fengyun-3 E star wind field measurement radar are: the difference in longitude and latitude is less than 0.1 degree, and the time difference is less than 15 minutes. After matching, the backscattering coefficient provided by FY is interpolated to the corresponding TM data points. Given that one TM data point may correspond to multiple FY data points and the closer observation values have higher weights for the interpolation result, in this embodiment, the inverse distance weighted interpolation method (IDW) is used to weight the FY data, and finally the estimated value of the backscattering coefficient of the wind field measurement radar at the specular reflection point of each TM data is obtained.

[0081] The method for constructing the training sample set includes: the reference wind speed is calculated from the east-west wind speed component U 10 (10m u-component of wind) and the north-south wind speed component V 10 (10m v-component of wind) at a height of 10 meters provided by ECMWF. The wind speed formula is as follows:

[0082] ,

[0083] where wspd represents the reference wind speed, u 10 represents the U component of the 10-meter wind field, and v 10 represents the V component of the 10-meter wind field.

[0084] The significant wave height of wind waves and swell waves is selected as the auxiliary data for swell waves, and the reference wind speed and the auxiliary data for swell waves are bilinearly interpolated to obtain a training sample set. Considering that both the reference wind speed and the SWH are global grid data, in this embodiment, the bilinear interpolation method is used to map their values to the position of the specular reflection point of the TM data. The specific calculation process is as follows: To obtain the estimated value of the unknown function f at P(x, y), given Q 11 (x1, y1), Q 12 (x1, y2), Q 21 (x2, y1), Q 22 (x2, y2) and their corresponding pixel values. First, interpolation is performed in the x direction to obtain R1 and R2:

[0085] ,

[0086] ,

[0087] where R1 represents the point obtained by horizontal interpolation between Q 11 and Q 21 , R2 represents the point obtained by horizontal interpolation between Q 12 and Q 22 , f(R1) represents the function value at point R1, f(R2) represents the function value at point R2, f(Q 11 ) represents the function value at point Q 11 , f(Q 12 ) represents the function value at point Q 12 , f(Q 21 ) represents the function value at point Q 21 , f(Q 22 ) represents the function value at point Q 22 . Then, based on R1 and R2, interpolation is performed in the y direction to obtain the function value f(P) of the target point P to be interpolated:

[0088] ,

[0089] In this embodiment, the target point P is the TM longitude and latitude coordinate data point of the specular reflection point of the TM; the reference wind speed data and the SWH data are bilinearly interpolated to the TM data point to complete the dataset production.

[0090] S2. Use several single - moment multi - frequency information fusion modules and multi - moment interactive attention enhancement modules to construct an initial wind speed inversion model, as Figure 2 shown, and use the training sample set to train the initial wind speed inversion model to obtain a cross - track time - series sea surface wind speed inversion model.

[0091] The single - moment multi - frequency information fusion module is as Figure 3As shown, it includes: a wavelet transform unit, a global enhancement unit, a local enhancement unit, a BP unit, and a feature fusion unit.

[0092] The wavelet transform unit is composed of a low-pass filter and a high-pass filter; the wavelet transform unit is used to perform two-dimensional wavelet decomposition on the time-delay Doppler correlation power map, divide the image into 1 low-frequency approximation sub-band and 3 high-frequency detail sub-bands, use the low-frequency approximation sub-band as the low-frequency DDM image, and perform weighted fusion on the 3 high-frequency detail sub-bands to obtain a high-frequency DDM image.

[0093] In this embodiment, in order to extract different scale features in the DDM image, wavelet transform is performed on the DDM image in this embodiment, and it is decomposed into high-frequency and low-frequency components by using two-dimensional discrete wavelet transform. Specifically, Haar wavelet is selected in this embodiment, and two-dimensional wavelet decomposition is performed once on the input DDM image, and the image is divided into 1 low-frequency approximation sub-band and 3 high-frequency detail sub-bands, which are as follows: Low-frequency approximation sub-band:

[0094] ,

[0095] Horizontal direction high-frequency detail sub-band:

[0096] ,

[0097] Vertical direction high-frequency detail sub-band:

[0098] ,

[0099] Diagonal direction high-frequency detail sub-band:

[0100] ,

[0101] Among them, X[k, l] is the pixel value of the k-th row and l-th column of the original image, g[·] is the Haar wavelet low-pass filter coefficient, h[·] is the Haar wavelet high-pass filter coefficient, and [m, n] is the position index of the output sub-image; 2m and 2n represent the downsampling factor of 2. The filter coefficients of the Haar wavelet are defined as follows: Low-pass filter:

[0102] ,

[0103] High-pass filter:

[0104] ,

[0105] The low-frequency approximation sub-band cA obtained by wavelet transform is the low-frequency DDM image; the detail information in the three high-frequency sub-bands is weighted and fused to form a high-frequency DDM image, and its fusion calculation formula is as follows:

[0106] ,

[0107] Taking the DDM diagram with an input size of 20×9 as an example, after two-dimensional wavelet transform processing, the low-frequency DDM and high-frequency DDM with sizes of 10×4 are obtained respectively.

[0108] The global enhancement unit consists of patch embedding, positional encoding, layer normalization, multi-head self-attention, and a multi-layer perceptron; patch embedding is used to convert the low-frequency DDM image and high-frequency DDM image into a sequence of feature vectors; positional encoding is used to introduce delay Doppler coordinates consistent with the size of the feature vector sequence and integrate the feature vector sequence with the delay Doppler coordinates to obtain an input vector; layer normalization is used to standardize the input vector to obtain a normalized vector; multi-head self-attention processes the normalized vector based on the self-attention mechanism to obtain several input features; the multi-layer perceptron is used to perform linear transformation and non-linear activation on each input feature to obtain a global feature sequence.

[0109] In this embodiment, the global enhancement unit establishes information connections between different positions of the image through the self-attention mechanism, so as to better extract the global features of the high- and low-frequency DDM images. The global enhancement module designed in this embodiment consists of patch embedding, positional encoding, layer normalization, multi-head self-attention, and a multi-layer perceptron. The details of each part will be introduced in detail below. The purpose of patch embedding is to convert image features into sequence inputs. In this module, the input image is first divided into blocks. To accurately capture the information of each pixel, the patch size is set to 1×1, and each pixel is encoded, so as to ensure that the local features of all pixels are completely retained. For the high-frequency or low-frequency DDM image after wavelet transformation, its size is 10×4, and the total number of corresponding patches is N = H×W = 10×4 = 40. Each patch is denoted as patch i , where i = 1, 2, …, N. After flattening each patch, a vector x i is formed:

[0110] ,

[0111] All the flattened vectors are combined into a sequence of feature vectors X:

[0112] ,

[0113] The main function of positional encoding is to provide the position information of the elements in the sequence for the model. To enhance the model's perception ability of global information, delay Doppler coordinates P consistent with the size of X are introduced for position marking:

[0114] ,

[0115] Finally, the input vector Z of the obtained encoder is the sum of the feature vector X and the position marker P:

[0116] .

[0117] Layer normalization is used to normalize the eigenvalues to a relatively stable range to accelerate the convergence process of model training.

[0118] The multi-head self-attention mechanism aims to fully model the similarity between input vectors, thereby capturing global dependencies and enhancing the network's perception of image context information. The specific working principle of multi-head self-attention is as follows: First, use the linear transformation matrices W Q 、W K and W V Multiply with the input vector Z respectively to calculate Q (Query), K (Key), and V (Value):

[0119] ,

[0120] ,

[0121] ,

[0122] Among them, Q represents the query vector, which is used to determine the information that should be focused on in the attention mechanism; K represents the key vector, which is used to calculate the correlation between Q and the features at other positions in Z and provide additional context information. V represents the value vector, which contains the position information of each element in Z. W Q 、W K and W V Are randomly generated at the beginning of network training and updated during training:

[0123] ,

[0124] Among them, learning_rate represents the learning rate, Represents the derivative of the loss function with respect to the weight matrix; in the global enhancement unit, Q, K, and V are the key elements to implement the self-attention mechanism. The output result of the self-attention mechanism is:

[0125] ,

[0126] Among them, Attention(Q,K,V) represents the attention weight, d k Represents the dimension size of the input vector. The essence of the multi-head self-attention mechanism is a linear transformation after splicing the calculation results of multiple attentions. This mechanism allows the model to use different feature information obtained at different positions, thereby increasing the diversity of features. The calculation method of multi-head self-attention is as follows:

[0127] ,

[0128] ,

[0129] where r represents the total number of heads. In this embodiment, the number of heads is 4, head j represents the feature of the j-th head, W represents the corresponding weight matrix when obtaining the output result, and Concat represents the concatenation operation of vectors.

[0130] To further improve the non-linear expression ability of the module, a multi-layer perceptron (MLP) is introduced after the attention output in this embodiment. The perceptron consists of two fully connected layers with the number of neurons being 128 and 64 respectively, which are used for linear transformation and non-linear activation of each input feature, so as to enhance the expression ability of the model. Finally, after being processed by the global enhancement module, the high-frequency and low-frequency DDM images respectively output a global feature sequence with a feature length of 64 as the input of the subsequent module.

[0131] The local enhancement unit includes a CNN network and a bidirectional feature enhancement sub-unit; the CNN is used to initially extract the local texture features of the low-frequency DDM image and the high-frequency DDM image, and the bidirectional feature enhancement sub-unit is used to perform global average pooling operation and weighting on the local texture features to obtain a local feature sequence.

[0132] In this embodiment, the local enhancement unit extracts the basic local texture features through the CNN, and on this basis, a bidirectional feature enhancement module is introduced. By performing global average pooling in the row and column directions, the attention weights in the horizontal and vertical directions are constructed, so as to enhance the directional feature expression ability of the local area. The bidirectional feature enhancement module designed in this embodiment consists of global average pooling in the row and column directions, an attention weighting mechanism, and a feature flattening structure.

[0133] The CNN is used to initially extract the local texture features of the high-frequency or low-frequency DDM image. The convolutional layer uses a 3×3 filter for feature extraction. This filter can be regarded as a feature extractor for local structure perception, and generates a feature map through convolution operation with the input image as the input of the bidirectional feature enhancement module. Taking the input high-frequency or low-frequency DDM image (4×10) as an example, a feature map of the same size (4×10×1) is output after convolution processing. The output result of the filter can be expressed by the following formula:

[0134] ,

[0135] Among them, z represents the output result of the filter, T represents the input high-frequency or low-frequency DDM image, M and b respectively represent the weights and biases of the convolution kernel, * represents the convolution operation, and f(·) represents the activation function. In this embodiment, the activation function is ReLU.

[0136] The bidirectional feature enhancer unit is as Figure 4 shown, aiming to enhance the expression ability of the significantly direction-sensitive regions in the convolution feature map. Its input is the feature map after the above convolution processing, denoted as , where H = 4 represents the height (number of rows) of the feature map F, and L = 10 represents the width (number of columns) of the feature map F. First, global average pooling operations are respectively applied in the vertical (row) direction and the horizontal (column) direction to obtain two one-dimensional attention features, which are respectively expressed as:

[0137] ,

[0138] ,

[0139] The two one-dimensional attention features are multiplied through the outer product to construct a two-dimensional attention weight map:

[0140] ,

[0141] where σ(·) represents the Sigmoid activation function, which is used to normalize the attention weights. Then, the attention weights are applied to the input feature map to obtain a weighted feature map:

[0142] ,

[0143] Finally, the weighted feature map is flattened into a local feature sequence with a feature length of 40 as the output of the local enhancement module.

[0144] The BP unit is used to extract features from the GNSS-R observation parameters, scatterometer parameters, and the significant wave height of wind waves and swell waves and perform weighted summation to obtain the multi-source parameter fusion feature.

[0145] In this embodiment, the BP unit is as Figure 5As shown in the figure, the feature extraction for multi-source parameter information includes three fully connected layers. The input of this unit includes 9 types of GNSS-R observation parameters, microwave scatterometer parameters, and swell auxiliary information, namely the longitude and latitude of the specular reflection point, normalized bistatic radar cross-section (NBRCS), DDM front slope (LES), signal-to-noise ratio, incident angle, range correction gain (RCG), HH polarization backscattering coefficient, VV polarization backscattering coefficient, and the significant wave height of wind waves and swells (SWH). The three fully connected layers are composed of 32, 64, and 128 neurons respectively. Each neuron performs a weighted sum of multiple output features from the previous layer and is processed by an activation function to generate an output feature. Therefore, the feature output by the i-th neuron in the first hidden layer can be expressed as:

[0146] ,

[0147] where x j represents the j-th input feature of this neuron, represents the weight value of the connection line between this neuron and the j-th input feature, represents the bias value of this neuron, and f(·) represents the activation function. In this embodiment, the activation function is ReLU.

[0148] The ReLU function is one of the most commonly used non-linear functions in neural networks due to its advantages such as simple calculation, fast convergence speed, and effectively alleviating the vanishing gradient problem. This function can truncate all negative parts of the input data to 0 and output the positive parts unchanged. Its mathematical expression is:

[0149] ,

[0150] where x represents the input data and f(x) represents the output result of the activation function.

[0151] The feature fusion unit is used to fuse the global feature sequence, local feature sequence, and multi-source parameter fusion feature to obtain the single-moment fusion feature.

[0152] In this embodiment, the feature fusion unit performs feature concatenation (Concat) on the high-frequency and low-frequency DDM features extracted by the global enhancement module and the local enhancement module respectively to form a complete DDM spatial feature; subsequently, it is concatenated again with the multi-source parameter fusion feature extracted by the BP module to construct the single-moment fusion feature.

[0153] The multi-moment interactive attention enhancement module is shown in Figure 6 the figure and includes: the first bidirectional long short-term memory network, the second bidirectional long short-term memory network, and the fully connected layer.

[0154] The first bidirectional long short-term memory network is used to extract the time-dependent features of several single-moment fusion features, and the multi-head self-attention mechanism is combined to enhance the internal correlation between different time steps. The second bidirectional long short-term memory network is used to further extract the deep temporal features of the single-moment fusion features, and the multi-head cross-attention mechanism is used to strengthen the interaction modeling between the features of different time steps. The fully connected layer is used to fuse the time-dependent features and the deep temporal features to obtain the sea surface wind speed prediction result.

[0155] In this embodiment, the core idea of BiLSTM is to introduce a backpropagation path on the basis of the traditional LSTM, enabling the model to simultaneously utilize sequence information from both the past and the future for modeling, thereby enhancing the modeling ability for context dependence. This bidirectional structure is particularly suitable for temporal tasks that require comprehensive consideration of historical and future context information. Similar to the traditional LSTM, BiLSTM introduces the cell state as the main carrier of the information flow, and controls the information update process through three types of gating mechanisms: the input gate, the forget gate, and the output gate. These gating units generate weight values between 0 and 1 through the sigmoid function to regulate the acceptance degree of new information, the retention degree of historical information, and the influence degree of the current state on the output. With the synergistic effect of the bidirectional information flow and the gating structure, BiLSTM can effectively capture the sequence dependence characteristics within a long time span and improve the performance of the model in temporal modeling tasks.

[0156] (a) Forget Gate: The forget gate is used to control the historical information components to be discarded in the cell state. It generates a weight value between 0 and 1 through a sigmoid function, representing the retention ratio of each state unit. Its mathematical expression is:

[0157] ,

[0158] where, f t represents the forget gate, σ represents the sigmoid function, W f represents the weight matrix of the forget gate, h t-1 represents the hidden layer state of the previous time step, x t represents the input of the current time step, t represents the time step, and b f represents the bias vector of the forget gate.

[0159] (b)Input Gate: The input gate consists of two parts: one is the sigmoid layer, which is used to determine the information components to be updated in the state; the other is the tanh layer, which is used to generate the candidate state vector at the current time step. The outputs of the two are multiplied element-wise to form the candidate information finally written into the cell state. The mathematical expression of the sigmoid layer is:

[0160] ,

[0161] The mathematical expression of the tanh layer is:

[0162] ,

[0163] where, i t represents the first sigmoid layer, W f represents the weight matrix of the first sigmoid layer, b f represents the bias vector of the first sigmoid layer, represents the first tanh layer, W C represents the weight matrix of the first tanh layer, b C represents the bias vector of the first tanh layer.

[0164] (c)Cell State: The cell state is the core structure of the LSTM, which is used to carry and store long-term dependency information across time steps. The update of the cell state is obtained by adding the output of the forget gate and the output of the input gate, and its mathematical expression is:

[0165] ,

[0166] where, C t represents the updated cell state, C t-1 represents the cell state of the previous time step.

[0167] (d)Output Gate: The output gate is used to control the hidden layer output at the current time step. It determines the state information to be included in the output through a sigmoid layer, then forms the candidate output through the tanh layer, and finally combines these two parts to form the final output. The mathematical expression of the sigmoid layer of the output gate is:

[0168] ,

[0169] The mathematical expression of the final hidden layer state is:

[0170] ,

[0171] where, o t represents the second sigmoid layer, Wo represents the weight matrix of the second sigmoid layer, b o represents the bias vector of the second sigmoid layer, h t represents the final hidden layer.

[0172] The Multi - Moment Interaction Attention Enhancement Module (MIAE) models the temporal dependence relationship of the multi - moment feature sequence through a double - layer BiLSTM network, and fuses the self - attention and cross - attention mechanisms to strengthen the dynamic association and information interaction between different time steps. This module first receives the fused features of multiple time steps as sequence inputs, which are input into the first - layer BiLSTM to extract the bidirectional hidden state vectors (128 - dimensional) of each time step, constructing preliminary temporal perception features. Subsequently, the self - attention mechanism is introduced to dynamically adjust the response weights of the features of each time step, highlighting the leading role of critical moments in the wind speed change. The enhanced feature sequence is passed to the second - layer BiLSTM to extract deeper - level temporal feature representations (64 - dimensional). Then, the module adopts the cross - attention mechanism, using the output of the second - layer BiLSTM as the query vector (Query), and the temporal perception features of the first layer as the key (Key) and value (Value) to achieve cross - modeling and feature fusion between different levels. Finally, the fused features are output through a fully - connected layer to obtain the sea surface wind speed inversion results corresponding to consecutive time steps.

[0173] After the neural network model is constructed, it needs to be trained with a large number of historical samples to fit the non - linear mapping relationship between the input and output. The entire training process can be divided into two stages: forward propagation and backward propagation. Among them, forward propagation means that the input data is processed by the weighted and activation functions of each layer and passed layer by layer to the output layer to complete the calculation of the predicted value; backward propagation is based on the gradient descent and optimization algorithm, and the prediction error is propagated backward along the network structure and used to update the parameter weights to minimize the overall loss.

[0174] During the training process, the loss function is used to measure the difference between the model inversion result and the true observation value. The commonly used evaluation index in the wind speed inversion model is the Mean Square Error (MSE) loss function. However, in this embodiment for the multi - moment sea surface wind speed inversion task, the model output is a wind speed sequence of consecutive multiple time steps. If only the MSE index is used, it is difficult to comprehensively describe the dynamic trend of the wind speed evolution over time.

[0175] In this embodiment, the method for training the initial wind speed inversion model includes: initializing the parameters in the initial wind speed inversion model and inputting the training sample set into the initial wind speed inversion model; obtaining the predicted value through forward propagation of the training sample set and calculating the loss function; using the backpropagation algorithm and the Adam optimizer to update the parameters in the initial wind speed inversion model; when the change in the continuously iterated loss function is less than 0.0001 or the set number of iterations is reached, the training ends, and the same-track multi-temporal sea surface wind speed inversion model is obtained.

[0176] The method for calculating the loss function includes: constructing the loss function by using the mean square error loss function, the direction consistency loss, and the amplitude consistency loss:

[0177] ,

[0178] ,

[0179] ,

[0180] ,

[0181] where L Total represents the loss function, L MSE represents the mean square error loss function, L Direction represents the direction consistency loss, L Trend represents the amplitude consistency loss, α, β, and γ represent the weight coefficients, softplus represents the softplus function, w1, w2, and w3 represent the learnable parameters; the calculation method of the mean square error loss function includes:

[0182] ,

[0183] where n represents the number of samples, y i represents the true value of the i-th sample, f i represents the inversion value of the i-th sample; the calculation method of the direction consistency loss includes, first, taking the difference of the predicted wind speed sequence and the true wind speed sequence respectively, calculating the gradient change between adjacent time steps to form the predicted gradient vector and the true gradient vector, and then using the cosine similarity to quantify the similarity degree of the two sets of gradient vectors in direction. The calculation formula is as follows:

[0184] ,

[0185] ,

[0186] ,

[0187] where Δy iDenote the difference in the true value of the sample, Δf i Denote the difference in the inversion value of the sample, y i+1 Denote the true value of the (i + 1)-th sample, f i+1 Denote the inversion value of the (i + 1)-th sample; The calculation method of the amplitude consistency loss includes calculating the variance of the predicted gradient and the true gradient, and using the absolute difference between the two as the amplitude consistency loss, so as to prompt the model to be consistent with the actual observation in the local fluctuation amplitude. The calculation formula is as follows:

[0188] ,

[0189] ,

[0190] ,

[0191] where, Var(Δy) represents the variance of the true value of the sample, and Var(Δf) represents the variance of the inversion value of the sample, Denote the mean value of the difference in the true value of the sample, Denote the mean value of the difference in the inversion value of the sample.

[0192] In this embodiment, the adaptive moment estimation (Adam) optimization algorithm is adopted. During the error backpropagation process, this optimizer is used to iteratively update the parameters in the neural network to accelerate convergence and minimize the loss function. The update rule of the Adam optimizer is as follows:

[0193] ,

[0194] where, t represents the number of iterations, g t Denote the current gradient, Denote the gradient of the loss function L with respect to the model parameter 𝜔, m t Denote the first moment estimation of the gradient, β1 represents the decay rate controlling the first moment estimation, m t-1 Denote the first moment estimation at the previous iteration, v t Denote the second moment estimation of the gradient, β2 represents the decay rate controlling the second moment estimation, Denote the result after correcting the estimated value of the first moment estimation, Denote the t-th power of the decay rate β1, Denote the result after correcting the estimated value of the second moment estimation, Denote the t-th power of the decay rate β1, Δω t Denote the updated value of the model parameter 𝜔 at the t-th iteration, α1 represents the learning rate, Denote the smoothing term.

[0195] S3. Obtain the delay-Doppler correlation power map, GNSS-R observation parameters, scatterometer parameters, and the significant wave height of wind waves and swell waves, and use the along-track time-series sea surface wind speed inversion model to complete the inversion of the sea surface wind speed, obtaining the predicted sea surface wind speed data.

[0196] In this embodiment, after completing the prediction of the sea surface wind speed, in order to evaluate the accuracy and fitting effect of the model inversion results, appropriate regression evaluation indicators need to be introduced for quantification. This embodiment uses the Root Mean Squared Error (RMSE) as the main evaluation indicator. RMSE can be used to measure the overall deviation between the model prediction value and the reference value, and its mathematical expression is as follows:

[0197] ,

[0198] where y i represents the true value of the i-th sample, and f i represents the inversion value of the i-th sample; n represents the number of samples. The smaller the RMSE value, the smaller the difference between the predicted data and the reference data, and the better the fitting effect; conversely, it indicates that the difference between the predicted data and the reference data is larger, and the fitting effect is worse.

[0199] Based on the data in September 2023, the training set and the test set are divided according to the ratio of 7:3, and the time-series length of a single input is set to N = 10. The experimental results show that the Root Mean Squared Error (RMSE) of the present invention is 1.13 m / s, demonstrating excellent performance in both prediction accuracy and time-series modeling ability. The experimental results are shown in Table 1:

[0200] Table 1

[0201] .

[0202] Embodiment 2

[0203] In this embodiment, the along-track time-series sea surface wind speed inversion system combining GNSS-R and scatterometer includes: a dataset construction module, a model construction module, and a prediction module.

[0204] The dataset construction module is used to obtain the historical satellite data of the ocean surface wind, preprocess the historical satellite data to obtain the preprocessed data, and construct a training sample set in combination with the reference wind speed and the swell auxiliary data.

[0205] The model construction module uses a number of single-moment multi-frequency information fusion modules and multi-moment interactive attention enhancement modules to construct an initial wind speed inversion model, and uses the training sample set to train the initial wind speed inversion model to obtain the along-track time-series sea surface wind speed inversion model.

[0206] The prediction module is used to obtain the delay-Doppler correlation power map, GNSS-R observation parameters, scatterometer parameters, and the significant wave height of wind waves and swell waves, and complete the inversion of the sea surface wind speed using the along-track sequential sea surface wind speed inversion model to obtain the predicted sea surface wind speed data.

[0207] The embodiments described above are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for retrieving sea surface wind speed in the same orbit in time series by combining GNSS-R and scatterometer, characterized in that, It includes the following steps: Obtain historical satellite data of ocean surface wind, preprocess the historical satellite data to obtain preprocessed data, and construct a training sample set by combining reference wind speed and swell auxiliary data; Construct an initial wind speed inversion model using several single - moment multi - frequency information fusion modules and multi - moment interactive attention enhancement modules, and train the initial wind speed inversion model using the training sample set to obtain a same - orbit time - series sea surface wind speed inversion model; Obtain a delay - Doppler correlation power map, GNSS - R observation parameters, scatterometer parameters, and significant wave height of wind and swell, and use the same - orbit time - series sea surface wind speed inversion model to complete sea surface wind speed inversion to obtain sea surface wind speed prediction data.

2. The GNSS-R and scatterometer combined along-track sequential sea surface wind speed inversion method according to claim 1, wherein The method for preprocessing the historical satellite data includes: Calculate distance gain correction through the antenna gain at the specular reflection point, the distance from the receiver to the specular reflection point, and the distance from the navigation satellite to the specular reflection point, and correct the historical satellite data for distance gain; ， Among them, RCG represents range gain correction, and G SP represents the antenna gain in the direction of the specular emission point, represents the distance from the receiver to the specular reflection point, R SP represents the distance from the navigation satellite to the specular reflection point, 10 27 represents the scaling factor; Perform incidence angle correction on the corrected data, and screen out the data with an incidence angle less than a preset threshold in the corrected data; Perform specular reflection point type correction on the screened - out data. In the specular reflection point type: 0 represents open sea, 0.5 represents coastal waters within 25 kilometers of land, 1 represents land, 2 represents sea ice. Screen out the data with a specular reflection point type less than 1 to obtain type - corrected data; Perform multi - look number mean operation on the wind field measurement radar data, extract the backscattering coefficients of the wind field measurement radar data in the C - band HH polarization and VV polarization, and integrate the type - corrected data and the backscattering coefficients to obtain the preprocessed data.

3. The co-orbit sequential sea surface wind speed inversion method by combining GNSS-R and scatterometer according to claim 1, wherein The method for constructing the training sample set includes: Select the U - component and V - component of the 10 - meter wind field measured by ECMWF, and calculate the reference wind speed based on the U - component and the V - component; ， where wspd represents the reference wind speed, and u 10 represents the U component of the wind field at 10 meters, and v 10 represents the V component of the wind field at 10 meters; Select the significant wave height of wind and swell as the swell auxiliary data, and perform bilinear interpolation on the reference wind speed and the swell auxiliary data to obtain the training sample set.

4. The GNSS-R and scatterometer combined along-track sequential sea surface wind speed inversion method according to claim 1, characterized in that The single - moment multi - frequency information fusion module includes: a wavelet transform unit, a global enhancement unit, a local enhancement unit, a BP unit, and a feature fusion unit; The wavelet transform unit is composed of a low - pass filter and a high - pass filter; the wavelet transform unit is used to perform two - dimensional wavelet decomposition on the delay - Doppler correlation power map, divide the image into 1 low - frequency approximation sub - band and 3 high - frequency detail sub - bands, use the low - frequency approximation sub - band as the low - frequency DDM image, and perform weighted fusion on the 3 high - frequency detail sub - bands to obtain a high - frequency DDM image; The global enhancement unit consists of patch embedding, positional encoding, layer normalization, multi-head self-attention, and a multi-layer perceptron; the patch embedding is used to convert the low-frequency DDM image and the high-frequency DDM image into a sequence of feature vectors; the positional encoding is used to introduce delay-Doppler coordinates consistent with the size of the sequence of feature vectors, and integrate the sequence of feature vectors with the delay-Doppler coordinates to obtain an input vector; the layer normalization is used to normalize the input vector to obtain a normalized vector; the multi-head self-attention processes the normalized vector based on the self-attention mechanism to obtain a number of input features; the multi-layer perceptron is used to perform linear transformation and non-linear activation on each input feature to obtain a global feature sequence; The local enhancement unit includes a CNN network and a bidirectional feature enhancement sub-unit; the CNN is used to initially extract the local texture features of the low-frequency DDM image and the high-frequency DDM image, and the bidirectional feature enhancement sub-unit is used to perform global average pooling operation and weighting on the local texture features to obtain a local feature sequence; The BP unit is used to extract features from GNSS-R observation parameters, scatterometer parameters, and the significant wave height of wind waves and swell waves, and perform weighted summation to obtain a multi-source parameter fusion feature; The feature fusion unit is used to fuse the global feature sequence, the local feature sequence, and the multi-source parameter fusion feature to obtain a single-moment fusion feature.

5. The GNSS-R and scatterometer combined along-track sequential sea surface wind speed inversion method according to claim 4, characterized in that, The multi-moment interactive attention enhancement module includes: a first bidirectional long short-term memory network, a second bidirectional long short-term memory network, and a fully connected layer; The first bidirectional long short-term memory network is used to extract the time-dependent features of a number of the single-moment fusion features, and enhance the internal correlation between different time steps by combining the multi-head self-attention mechanism; The second bidirectional long short-term memory network is used to further extract the deep temporal features of the single-moment fusion features, and strengthen the interactive modeling between the features of different time steps through the multi-head cross-attention mechanism; The fully connected layer is used to fuse the time-dependent features and the deep temporal features to obtain the sea surface wind speed prediction result.

6. The GNSS-R and scatterometer combined along-track sequential sea surface wind speed inversion method according to claim 1, characterized in that, The method for training the initial wind speed inversion model includes: Initializing the parameters in the initial wind speed inversion model, and inputting the training sample set into the initial wind speed inversion model; The training sample set obtains a predicted value through forward propagation, and calculates the loss function; Using the backpropagation algorithm and the Adam optimizer to update the parameters in the initial wind speed inversion model; When the change in the continuously iterated loss function is less than 0.0001 or the set number of iterations is reached, the training ends, and the same-orbit multi-temporal sea surface wind speed inversion model is obtained.

7. The method for retrieving sea surface wind speed in the same orbit and time sequence by combining GNSS-R and scatterometer according to claim 6, wherein The method for calculating the loss function includes: constructing the loss function using the mean square error loss function, the direction consistency loss, and the amplitude consistency loss: ， ， ， ， Among them, L Total represents the loss function, and L MSE represents the mean squared error loss function, and L Direction represents the direction consistency loss, and L Trend represents the amplitude consistency loss. α, β, and γ represent weight coefficients, softplus represents the softplus function, and w1, w2, and w3 represent learnable parameters; The calculation method of the mean square error loss function includes: ， Among them, n represents the number of samples, and y i represents the true value of the i-th sample, and f i represents the inversion value of the i-th sample; The calculation method of the direction consistency loss includes: ， ， ， Among them, Δy i represents the difference in the true value of the sample, and Δf i represents the difference in the inversion value of the sample, y i+1 represents the true value of the (i + 1)-th sample, and f i+1 represents the inversion value of the (i + 1)-th sample; The calculation method of the amplitude consistency loss includes: ， ， ， Among them, Var(Δy) represents the variance of the true values of the samples, and Var(Δf) represents the variance of the inversion values of the samples. represents the mean of the differences in the true values of the samples, represents the mean of the differences in the inversion values of the samples.

8. The GNSS-R and scatterometer combined along-track sequential sea surface wind speed inversion method according to claim 6, characterized in that, The update rule of the Adam optimizer includes: ， Among them, t represents the number of iterations, and g t represents the current gradient, represents the gradient of the loss function L with respect to the model parameter 𝜔, m t represents the first-order moment estimate of the gradient, and β1 represents the decay rate controlling the first-order moment estimate, m t-1 represents the first-order moment estimate at the previous iteration, v t represents the second-order moment estimate of the gradient, and β2 represents the decay rate controlling the second-order moment estimate, represents the result after correcting the estimated value of the first-order moment estimate, represents the t-th power of the decay rate β1, represents the result after correcting the estimated value of the second-order moment estimate, represents the t-th power of the decay rate β1, Δω t represents the updated value of the model parameter 𝜔 at the t-th iteration, and α1 represents the learning rate, represents the smoothing term.

9. A co - orbiting sequential sea surface wind speed inversion system combining GNSS - R and scatterometer, the inversion system applying the inversion method according to any one of claims 1 - 8, characterized in that, Includes: A dataset construction module, a model construction module, and a prediction module; The dataset construction module is used to obtain historical satellite data of ocean surface wind, preprocess the historical satellite data to obtain preprocessed data, and construct a training sample set in combination with reference wind speed and swell auxiliary data; The model construction module constructs an initial wind speed inversion model by using a number of single-moment multi-frequency information fusion modules and multi-moment interactive attention enhancement modules, and trains the initial wind speed inversion model by using the training sample set to obtain a same-orbit time-series sea surface wind speed inversion model; The prediction module is used to obtain a delay-Doppler correlation power map, GNSS-R observation parameters, scatterometer parameters, and significant wave height of wind waves and swells, and complete sea surface wind speed inversion by using the same-orbit time-series sea surface wind speed inversion model to obtain sea surface wind speed prediction data.

Images