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

Through the same-track multi-time sequence model and a double-layer LSTM network, combined with surge assist data, the problems of insufficient GNSS-R wind speed inversion accuracy and surge interference are solved, and higher wind speed prediction accuracy and stability are achieved.

CN119001767AInactive Publication Date: 2024-11-22CHINA UNIV OF PETROLEUM (EAST CHINA)

Patent Information

Application Number
CN202411488736.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-24
Publication Date
2024-11-22
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing GNSS-R wind speed inversion method has insufficient accuracy and sensitivity to sea surface surge interference when dealing with sea surface wind speed, resulting in inaccurate wind speed prediction.

Method used

Using the same-orbit multi-time sequence model, combining the same-orbit neighborhood information fusion module and the multi-time double-layer LSTM module, the training sample set is constructed using historical satellite data, reference wind speed and surge assisted data, and features are extracted through the CNN and BP modules, and through The double-layer LSTM network establishes a wind speed inversion model to reduce surge interference and improve wind speed prediction accuracy.

Benefits of technology

The accuracy and stability of wind speed inversion were significantly improved, with the RMSE reduced from 1.88 m/s to 1.02 m/s and the MAE reduced from 1.41 m/s to 0.63 m/s, enhancing the robustness and predictive performance of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119001767A_ABST
    Figure CN119001767A_ABST
Patent Text Reader

Abstract

The invention discloses a same-orbit multi-time-sequence GNSS-R sea surface wind speed inversion method and system, and belongs to the field of ocean data processing, and the method comprises the following steps: obtaining historical satellite data of ocean surface wind, preprocessing the historical satellite data to obtain preprocessed data, and constructing a training sample set in combination with a reference wind speed and surge auxiliary data; constructing an initial wind speed inversion model by using the same-orbit neighborhood information fusion module and the multi-moment double-layer LSTM module, and training the initial wind speed inversion model by using the training sample set to obtain a same-orbit multi-time-sequence sea surface wind speed inversion model; and a time delay-Doppler correlation power diagram, GNSS-R auxiliary parameters and the effective wave height of the storm and surge are obtained, and sea surface wind speed inversion is completed by using the same-orbit multi-time-sequence sea surface wind speed inversion model to obtain sea surface wind speed prediction data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of ocean data processing, and in particular relates to a same-track multi-time series GNSS-R sea surface wind speed inversion method and system. 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. GNSS signals are affected by the earth's surface during the reflection process, causing changes in signal strength, delay, frequency, etc. The sea surface wind field is one of the important factors affecting the roughness of the sea surface. The roughness of the sea surface is different under different wind speed conditions. The difference in roughness will cause changes in reflectivity, which will cause changes in the power and frequency of the reflected signal, and eventually be captured by the receiver. 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 commonly used methods for GNSS-R wind speed inversion. The implementation process of this method is to first obtain a series of DDMs through field measurements or numerical simulations and record the corresponding wind speed values ​​to build a "DDM-wind speed" benchmark library for matching. Then, the DDM of the unknown wind speed is compared (matched) with each DDM in the benchmark library, and the wind speed corresponding to the benchmark DDM with the smallest difference is used as the inverted wind speed; (2) Wind speed inversion method based on geographic model function: The wind speed inversion method based on empirical function mainly inverts wind speed by establishing a mapping relationship between GNSS-R observations and wind speed. This method is also one of the commonly used methods for GNSS-R wind speed inversion. The empirical function used here is the Geography Model Function. ,GMF), including exponential function, polynomial function and 2D lookup table (LUT) and other forms; (3) Wind speed inversion method based on machine learning (deep learning): Compared with the empirical function model, the machine learning model has a stronger nonlinear mapping ability. This type of method uses DDM data and its derived features as the main input, and mines deep features through models such as random forest, BP neural network, and convolutional neural network to establish a mapping relationship between input and wind speed. Summary of the invention

[0004] The present invention aims to solve the deficiencies of the prior art and provides the following solutions:

[0005] A method for inverting sea surface wind speed using a same-track multi-time series GNSS-R system, comprising the following steps:

[0006] Acquire 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 surge auxiliary data;

[0007] An initial wind speed inversion model is constructed by using a same-track neighborhood information fusion module and a multi-time double-layer LSTM module, and the initial wind speed inversion model is trained by using the training sample set to obtain a same-track multi-time series sea surface wind speed inversion model;

[0008] The delay-Doppler correlation power diagram, GNSS-R auxiliary parameters and effective wave height of wind and wave surge are obtained, and the sea surface wind speed inversion is completed by using the same-track multi-time series sea surface wind speed inversion model to obtain the sea surface wind speed prediction data.

[0009] Preferably, the method for preprocessing the historical satellite data includes:

[0010] The range gain correction is calculated by the antenna gain at the mirror reflection point, the distance from the receiver to the mirror reflection point, and the distance from the navigation satellite to the mirror reflection point, and the range gain correction is performed on the historical satellite data: ,

[0011] Where RCG represents the range gain correction, G SP represents the antenna gain in the direction of the mirror emission point, Indicates the distance from the receiver to the mirror reflection point, R SP Indicates the distance from the navigation satellite to the mirror reflection point, 10 27 represents the scaling factor;

[0012] Performing incident angle correction on the corrected data, and screening out data whose incident angle is less than a preset threshold value in the corrected data;

[0013] The filtered data are corrected for the mirror reflection point type, in which: 0 represents open sea area, 0.5 represents coastal sea area within 25 kilometers from the land, 1 represents land, and 2 represents sea ice. The data with the mirror reflection point type less than 1 are filtered out to obtain the preprocessed data.

[0014] Preferably, the method for constructing the training sample set includes:

[0015] The U component and the V component of the 10-meter wind field measured by ECMWF are selected, and the reference wind speed is calculated based on the U component and the V component: ,

[0016] Where wspd represents the reference wind speed, u 10 represents the U component of the 10-meter wind field, v 10represents the V component of the 10-meter wind field;

[0017] The effective wave heights of wind waves and swell waves are selected as the swell wave auxiliary data, and the reference wind speed and the swell wave auxiliary data are bilinearly interpolated to obtain the training sample set.

[0018] Preferably, the same-track neighborhood information fusion module includes: a number of CNN modules and a number of BP modules;

[0019] The CNN module is used to extract 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 are respectively composed of 32 3×3 filters and 16 3×3 filters, and 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 a maximum pooling operation, and are used to perform a maximum pooling operation on the first feature map to obtain a pooled feature map, with a pooling window size of 2×2 and a step size of 2; the flattening layer is used to flatten the pooled feature map into a first feature sequence with a length of 160;

[0020] The BP module includes three fully connected layers, each of which is composed of 32 neurons, 64 neurons, and 128 neurons, respectively. The BP module is used to extract features of the GNSS-R auxiliary parameters and the effective wave height of the wind wave and surge to obtain a second feature sequence;

[0021] The on-track neighborhood information fusion module connects the first feature sequence and the second feature sequence to obtain a fused feature sequence with a length of 288.

[0022] Preferably, the multi-time double-layer LSTM module is composed of 2 layers of LSTM networks, and each layer of the LSTM network includes: a forget gate, an input gate, a unit state update and an output gate;

[0023] The forget gate comprises: ,

[0024] Among them, 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 state of the previous time step, x t represents the input of the current time step, b f Represents the bias vector of the forget gate;

[0025] The input gate is composed of the first sigmoid layer and the first tanh layer, including:

[0026] ,

[0027] ,

[0028] Among them, i t represents the first sigmoid layer, W i represents the weight matrix of the first sigmoid layer, b i 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;

[0029] The unit status update includes:

[0030] ,

[0031] Among them, C t Represents the updated unit state, C t-1 represents the cell state at the previous time step;

[0032] The output gate is composed of the second sigmoid layer and the final hidden layer, including:

[0033] ,

[0034] ,

[0035] Among them, t represents the second sigmoid layer, W o 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.

[0036] Preferably, the method for training the initial wind speed inversion model includes:

[0037] Initializing parameters in the initial wind speed inversion model, and transferring the training sample set into the initial wind speed inversion model;

[0038] The training sample set obtains a predicted value through forward propagation and calculates a loss function;

[0039] Using a back propagation algorithm and an Adam optimizer to update the parameters in the initial wind speed inversion model;

[0040] When the change of the continuous iteration loss function is less than 0.0001 or reaches the set number of iterations, the training ends and the same-track multi-time series sea surface wind speed inversion model is obtained.

[0041] Preferably, the method for calculating the loss function includes: constructing the loss function using each wind speed error, the gradient change error between two wind speeds and the overall wind speed error:

[0042] ,

[0043] Among them, L represents the loss function, L MSE Represents each wind speed error, L D Represents the gradient change error between two wind speeds, L T represents the overall wind speed error;

[0044] The calculation method of each wind speed error includes:

[0045] ,

[0046] Where n represents the number of samples, y p represents the true value of the pth sample, f p represents the inversion value of the pth sample;

[0047] The calculation method of the gradient change error between the two wind speeds includes:

[0048] ,

[0049] ,

[0050] ,

[0051] Among them, Δy p Represents the difference between the true value of the sample, Δf p Represents the difference of sample inversion values, y p+1 represents the true value of the p+1th sample, f p+1 represents the inversion value of the p+1th sample;

[0052] The calculation method of the overall wind speed error includes:

[0053] .

[0054] Preferably, the update rule of the Adam optimizer includes:

[0055] ,

[0056] Among them, s represents the number of iterations, g s represents the current gradient, Represents the loss function L relative to the model parameters The gradient of L represents the loss function, m srepresents the first-order moment estimate of the gradient, β 1 represents the decay rate of the control first-order moment estimate, m s-1 represents the first-order moment estimate at the previous iteration, v s represents the second-order moment estimate of the gradient, β 2 represents the decay rate of the control second-order moment estimate, It represents the result after correcting the estimated value of the first-order moment estimate. represents the decay rate β 1 The sth power, It represents the result after correcting the estimated value of the second-order moment estimate. represents the decay rate β 1 The sth power, Represents the model parameters at the sth iteration Update value, α represents the learning rate, represents the smoothing term.

[0057] The present invention also provides a same-track multi-time series GNSS-R sea surface wind speed inversion system, the inversion system applying any of the above-mentioned inversion methods, including: a data acquisition module, a model building module and an inversion module;

[0058] The data acquisition module is used to acquire 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 surge auxiliary data;

[0059] The model building module uses the same-track neighborhood information fusion module and the multi-time double-layer LSTM module to build an initial wind speed inversion model, and uses the training sample set to train the initial wind speed inversion model to obtain a same-track multi-time series sea surface wind speed inversion model;

[0060] The inversion module is used to obtain the delay-Doppler correlation power diagram, GNSS-R auxiliary parameters and effective wave height of wind wave and surge, and use the same-track multi-time series sea surface wind speed inversion model to complete the sea surface wind speed inversion to obtain sea surface wind speed prediction data.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] The present invention adopts the same-track multi-time series model to process 10 DDMs of the same track at one time, fully considering the spatiotemporal correlation between adjacent data points, which is more effective than the existing single DDM processing method. Through this design, the accuracy of wind speed inversion is greatly improved; the RMSE of the inversion result is increased from 1.88 m / s of single DDM processing to 1.67 m / s, and the MAE is increased from 1.41 m / s to 1.06 m / s, which significantly improves the accuracy and stability of wind speed inversion.

[0063] The existing wind speed inversion model is often disturbed by sea swells, resulting in an overestimation of the wind speed prediction. The present invention effectively reduces the influence of swells on the inversion model by introducing the effective wave height of wind and wave swells as auxiliary information, thereby improving the accuracy of wind speed prediction. After adding the effective wave height of wind and wave swells, the accuracy of the inversion model is further improved, and the RMSE is reduced from 1.67 m / s to 1.02 m / s, and the MAE is reduced from 1.06 m / s to 0.63 m / s, which significantly enhances the robustness and prediction performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0065] Figure 1 A schematic diagram of a method flow chart of an embodiment of the present invention;

[0066] Figure 2 It is a schematic diagram of a same-track multi-time series sea surface wind speed inversion model according to an embodiment of the present invention;

[0067] Figure 3 Schematic diagram of a CNN module according to an embodiment of the present invention;

[0068] Figure 4 Schematic diagram of a BP module according to an embodiment of the present invention;

[0069] Figure 5 Schematic diagram of a multi-time double-layer LSTM module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0070] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0071] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0072] Embodiment 1

[0073] In this embodiment, if Figure 1 As shown, a same-track multi-time series GNSS-R sea surface wind speed inversion method includes the following steps:

[0074] S1. 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 surge auxiliary data.

[0075] In this embodiment, the historical satellite data of ocean surface wind of Tianmu-1 (TM) is first obtained, and the historical satellite data is preprocessed. The preprocessing method includes: calculating the distance gain correction RCG through the antenna gain at the mirror reflection point, the distance from the receiver to the mirror reflection point, and the distance from the navigation satellite to the mirror reflection point (these data are all provided by Tianmu-1, and their function is to reflect the actual gain of the signal at this location. The larger the RCG, the more suitable the data quality is for wind speed inversion), and performing distance gain correction on the historical satellite data:

[0076] ,

[0077] Where RCG represents the range gain correction, G SP represents the antenna gain in the direction of the mirror emission point, Indicates the distance from the receiver to the mirror reflection point, R SP Indicates the distance from the navigation satellite to the mirror reflection point, 10 27 Represents the scaling factor, the purpose is to scale the RCG to the order of 1-100, and select data with RCG>10, the purpose is to delete data with insufficient gain to avoid interfering with wind speed inversion. The corrected data is corrected for the incident angle, and the data with an incident angle less than a preset threshold in the corrected data is screened out. In this embodiment, the incident angle is provided by Tianmu-1. As the incident angle increases, the corresponding antenna gain continues to decrease, which directly affects the quality of the received reflected signal. Therefore, data with an incident angle less than 60° is selected during data processing. The filtered data is corrected for the mirror reflection point type. The mirror reflection point type is provided by Tianmu-1. In the mirror reflection point type: 0 represents open waters, 0.5 represents coastal waters within 25 kilometers of land, 1 represents land, and 2 represents sea ice. Data with a mirror reflection point type less than 1 is screened out to eliminate the influence of land and sea ice to obtain pre-processed data.

[0078] In this embodiment,

[0079] The method for constructing the training sample set includes: selecting the U component and V component of the 10-meter wind field measured by ECMWF, and calculating the reference wind speed based on the U component and the V component:

[0080] ,

[0081] Where, wspd represents the reference wind speed, u 10 represents the U component of the 10-meter wind field, v 10Represents the V component of the 10-meter wind field.

[0082] The effective wave height of wind waves and swells is selected as auxiliary swell data, and the reference wind speed and swell auxiliary data are bilinearly interpolated to obtain a training sample set. Since swells increase the roughness of the sea surface, it is easy to cause the wind speed inversion model to overestimate the wind speed. Therefore, the interference of swells on wind speed inversion can be eliminated by using the effective wave height SWH (Significant Height of Combined Wind Waves and Swell) of wind waves and swells as the swell auxiliary data input. Because the reference wind speed and SWH are grid data covering the world, there are four adjacent grid data points for each TM longitude and latitude coordinate data point, so bilinear interpolation is used. Bilinear interpolation is a method of interpolating data points in two-dimensional space. This method takes into account the four data points closest to the interpolation point and obtains the result through linear interpolation calculation. Specifically, bilinear interpolation first performs linear interpolation on two groups of points in one direction, and then performs linear interpolation in the other direction to estimate the value of the target point. To obtain the pixel value of the unknown function f in P(x,y), it is known that Q 11 (x 1 ,y 1 ), Q 12 (x 1 ,y 2 ), Q 21 (x 2 ,y 1 ), Q 22 (x 2 ,y 2 ) and the corresponding pixel values. First, interpolate in the x direction to get R 1 and R 2 :

[0083] ,

[0084] ,

[0085] Among them, R 1 Indicates Q 11 and Q 21 The point obtained by horizontal interpolation, R 2 Indicates Q 12 and Q 22 The point obtained by horizontal interpolation, f(R 1 ) means at point R 1 The function value at, f(R 2 ) means at point R 2 The function value at, f(Q 11 ) means at point Q 11The function value at 12 ) means at point Q 12 The function value at 21 ) means at point Q 21 The function value at 22 ) means at point Q 22 The function value at , and then based on R 1 and R 2 Interpolate in the y direction to obtain the function value f(P) of the target point P to be interpolated:

[0086] ,

[0087] In this embodiment, the target point P is the TM longitude and latitude coordinate data point of the mirror 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 data set production.

[0088] S2. The initial wind speed inversion model is constructed using the same-track neighborhood information fusion module and the multi-time double-layer LSTM module, and the initial wind speed inversion model is trained using the training sample set to obtain the same-track multi-time series sea surface wind speed inversion model, such as Figure 2 shown.

[0089] The same-track neighborhood information fusion module includes: several CNN modules and several BP modules;

[0090] The CNN module is used to extract the features of a single delay-Doppler correlation power map (DDM). The CNN module consists of 2 convolutional layers, 2 pooling layers and 1 flattening layer, such as Figure 3 As shown in the figure: the two convolutional layers are composed of 32 3×3 filters and 16 3×3 filters respectively, which are used to perform convolution operation on the single delay-Doppler correlation power map to obtain the first feature map; the two pooling layers both use the maximum pooling operation to perform the maximum pooling operation on the first feature map to obtain the 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 160.

[0091] In this embodiment, the CNN module is used to extract the features of a single DDM, and is composed of two convolutional layers, two pooling layers, and a flattening layer. The two convolutional layers in this module are composed of 32 and 16 3×3 filters, respectively. Each filter can be regarded as a feature extractor, which performs convolution operations with the input single DDM to generate feature maps, which are the output results of the convolutional layer. Taking the input single DDM effective scattering area image (9×20) as an example, after processing by the first convolutional layer, a 9×20×32 feature map is obtained, where the output result of the kth (k=1,…,32) filter can be expressed as follows:

[0092]

[0093] Among them, h k represents the output result of the kth filter, φ represents the activation function, W k represents the weight of the kth convolution kernel, b k Represents the bias of the kth convolution kernel, and * represents the convolution operation. Both pooling layers in the CNN module use the maximum pooling operation, with a pooling window size of 2×2 and a step size of 2. The pooling layer can downsample the feature map output by the convolution layer to reduce information redundancy and retain key features. Each input DDM (9×20) is passed through two convolutional layers and two pooling layers to obtain a 2×5×16 feature map. After that, the flattening layer flattens it into a feature sequence of length 160. In order to take into account the neighborhood information between adjacent DDMs, the features extracted from adjacent DDMs by CNN are added. For example, for DDM (2), the output of its CNN module is CNN (1) + 2CNN (2) + CNN (3). Among them, adding CNN (2) twice is to pay more attention to the original features of the DDM at this moment while taking into account the neighborhood information. After the addition process, the output is still a feature sequence of length 160.

[0094] The BP module includes three fully connected layers, such as Figure 4 As shown, the three fully connected layers are composed of 32 neurons, 64 neurons and 128 neurons respectively. The BP module is used to extract the features of the GNSS-R auxiliary parameters and the effective wave height of wind and swell waves to obtain the second feature sequence. In this embodiment, the BP module is used for feature extraction of auxiliary information and includes three fully connected layers. The input of the BP module includes 7 categories of GNSS-R parameters and swell auxiliary information, namely the longitude and latitude of the mirror reflection point, the normalized bistatic radar cross section (NBRCS), the DDM leading slope (LES), the signal-to-noise ratio, the angle of incidence, the distance correction gain (RCG), and the effective wave height (SWH) of wind and swell waves. The three fully connected layers are composed of 32, 64 and 128 neurons respectively. Each neuron performs cumulative summation operation with multiple output features of the previous layer, and then is processed by an activation function (non-linear layer) to generate an output feature. Therefore, there are as many feature outputs as there are neurons in the fully connected layer. Taking the input features as an example, the features output by the i-th neuron of the first hidden layer can be expressed as follows:

[0095]

[0096] Among them, x j represents the jth input feature of the neuron, Represents the weight value of the connection line between the neuron and the jth input feature, Represents the bias value of the neuron, and g represents the ReLU activation function. The ReLU function is a commonly used nonlinear activation function that can truncate all negative parts of the input data to 0 and output the positive parts intact. Its mathematical expression is:

[0097]

[0098] Among them, z represents the input data, and max represents the maximum value between 0 and z. Since the derivative of the ReLU activation function is a constant 1 in the positive part and 0 in the negative part, the gradient calculation is simple and does not cause the problem of gradient disappearance. In addition, the ReLU function also has the advantages of fast calculation speed and strong expression ability, so it is widely used in neural networks.

[0099] The on-track neighborhood information fusion module connects the first feature sequence and the second feature sequence to obtain a fused feature sequence with a length of 288. In this embodiment, the model inputs 10 DDMs at a time, connects each set of features extracted by the CNN module and the BP module, and forms ten feature sequences with a length of 288. These 288×10 features are then input into the LSTM module for wind speed inversion.

[0100] The multi-time double-layer LSTM module consists of two layers of LSTM networks, each of which includes a forget gate, an input gate, a unit state update, and an output gate.

[0101] Although traditional recurrent neural networks (RNNs) can theoretically capture long-term dependencies of sequence data, in practice they are often plagued by the problem of gradient vanishing or exploding, which limits their performance in sequence data modeling. To solve these problems, the Long Short-Term Memory (LSTM) network model came into being. As a special variant of RNN, LSTM not only successfully solves the problem that RNN is difficult to capture long-term dependencies by introducing a unique gating mechanism (including forget gate, input gate, and output gate), but also greatly improves the learning and expression capabilities of time series data. The core idea of ​​LSTM is to design a cell state that can flexibly control the flow of information. This cell state runs through the entire sequence, allowing information to be stored or forgotten for a long time. LSTM consists of three key gating units: input gate, forget gate, and output gate, which together determine the update of the cell state and the output of the final hidden state of each cell. These gating units generate values ​​between 0 and 1 through the sigmoid function (where 0 means completely rejected and 1 means completely passed. Specifically, the closer the value is to 0, the more likely the information is to be forgotten during the transmission process; the closer the value is to 1, the more likely the information is to be retained during the transmission process.), representing the degree of acceptance of new information (input gate), the degree of forgetting old information (forget gate), and the degree of exposure of the cell state to the output (output gate).

[0102] This embodiment designs a two-layer LSTM, such as Figure 5 As shown in Figure 2, compared with single-layer LSTM, multi-layer LSTM has stronger expression and generalization capabilities. In the two-layer LSTM structure, the first layer LSTM can capture the primary features of sequence data, while the second layer can further refine these features and capture higher-level abstract features.

[0103] The following introduces a single-layer LSTM and its structure, and then introduces a multi-time double-layer LSTM. A single-layer LSTM network includes: forget gate, input gate, cell state update, and output gate.

[0104] The forget gate determines which information is discarded from the cell state. It generates a value between 0 and 1 through a sigmoid function, indicating the degree of retention of each state value. The forget gate includes:

[0105] ,

[0106] Among them, f t represents the forget gate, σ represents the sigmoid function, W f represents the weight matrix of the forget gate, ht-1 represents the hidden state of the previous time step, x t represents the input of the current time step, t represents the time step, b f Represents the bias vector of the forget gate.

[0107] The input gate consists of the first sigmoid layer and the first tanh layer. A sigmoid layer determines which values ​​will be updated, and a tanh layer generates a new candidate value vector. The output of the sigmoid layer and the tanh layer of the input gate are multiplied to obtain the updated candidate values, including:

[0108] ,

[0109] ,

[0110] Among them, 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.

[0111] The cell state update is the core of LSTM, which carries the information of the previous time step. The cell state update is obtained by adding the output of the forget gate and the output of the input gate, including:

[0112] ,

[0113] Among them, C t Represents the updated unit state, C t-1 Represents the cell state at the previous time step.

[0114] The output gate determines the value of the next hidden state. It consists of the second sigmoid layer and the final hidden layer. A sigmoid layer determines which unit states will be output, and then a tanh layer generates candidate values ​​for the output state. Finally, these two parts are combined to form the final output, including:

[0115] ,

[0116] ,

[0117] Among them, t represents the second sigmoid layer, W o represents the weight matrix of the second sigmoid layer, b orepresents the bias vector of the second sigmoid layer, h t represents the final hidden layer.

[0118] The two-layer LSTM module processes sequence data by setting up two layers of LSTM in parallel. The input layer receives time series features at different times in parallel and passes them to the first layer of LSTM. The first layer of LSTM uses mechanisms such as forget gate, input gate, and output gate to control the flow of information at each time step, establishes connections between time steps through cell states, and generates hidden state outputs. The second layer of LSTM receives the hidden state of the first layer corresponding to the time step as the input of the second layer (that is, the hidden layer output of the first layer of LSTM at the Xth moment is used as the input of the second layer of LSTM at the Xth moment), and further refines the GNSS-R spatiotemporal correlation characteristics of the sea surface wind speed. Finally, the hidden state of the second layer is mapped through the fully connected layer to obtain the sea surface wind speed prediction results at each moment. This structure can better capture and express the complex time dependency between GNSS-R features, thereby improving the performance of the model in the inversion of sea surface wind speed from sequence data.

[0119] Specifically, the multi-time double-layer LSTM module of this embodiment uses the 10 sets of features (10 moments) extracted by the same-track neighborhood information fusion module for wind speed inversion that takes into account temporal and spatial correlations, and includes two LSTM layers and one fully connected layer. The two LSTM layers are composed of 64 and 32 neurons respectively (i.e., the output of each unit of the first layer LSTM (h 1-1 —h 1-10 ) contains 64 features, and the output of each unit in the second layer LSTM (h 2-1 —h 2-10 ) contains 32 features). The fully connected layer contains 10 neurons, which are used to map the 32 features output by each LSTM unit in the second layer into a single wind speed value for each time step (i.e., the model inputs 10 DDMs and their auxiliary parameters, and finally outputs 10 corresponding wind speed inversion values). These 10 sets of features are regarded as a 10-length (time T 1 -T 10 ) is input into LSTM, and the time series processing capability of LSTM is used to capture the spatiotemporal correlation between features, so that the feature information at previous moments can be taken into account when making wind speed predictions at the current moment, thereby improving the accuracy and reliability of wind speed inversion.

[0120] The method for training the initial wind speed inversion model includes: initializing the parameters in the initial wind speed inversion model, and passing the training sample set into the initial wind speed inversion model; obtaining the predicted value of the training sample set through forward propagation, and calculating the loss function; using the back propagation algorithm and the Adam optimizer to update the parameters in the initial wind speed inversion model; when the change of the continuous iterative loss function is less than 0.0001 or reaches the set number of iterations, the training ends and the same-track multi-time series sea surface wind speed inversion model is obtained.

[0121] In this embodiment, after the neural network model is built, a large amount of historical data training is required to fit the nonlinear relationship between the input data and the output data. And this training process can be divided into two stages: forward propagation of information and back propagation of errors. Among them, forward propagation refers to the process of transferring data from the input layer to the output layer, which is realized by weighted calculation. The back propagation of errors is a process based on the gradient descent algorithm and the optimization algorithm to update the parameter weights in the neural network by back propagating the prediction error. In the neural network training process, the loss function is used to measure the difference between the inversion value and the true value. The wind speed inversion model often uses the mean square error loss function (MSE) to characterize the error between the predicted wind speed and the actual observed wind speed. In the inversion of the sea surface wind speed of the same track multi-series GNSS-R in the present invention, because 10 wind speed values ​​are output, if MSE is used as the loss function alone, the change trend of wind speed over time cannot be fully captured. Therefore, an improved loss function is proposed in this embodiment, and this composite loss function is intended to simultaneously optimize the accuracy of the wind speed value, the wind speed change trend, and the prediction accuracy of the overall wind speed. The method for calculating the loss function includes: constructing the loss function using each wind speed error, the gradient change error between two wind speeds, and the overall wind speed error:

[0122] ,

[0123] Among them, L represents the loss function, L MSE Represents each wind speed error, L D Represents the gradient change error between two wind speeds, L T Represents the overall wind speed error.

[0124] The mean square error loss function (MSE) is a loss function commonly used in regression problems. In this embodiment, the calculation method of each wind speed error includes:

[0125] ,

[0126] Where n represents the number of samples, y p represents the true value of the pth sample, f pRepresents the inversion value of the pth sample. This loss function measures the prediction accuracy of the model by the average of the square of the difference between the inversion value and the true value. The smaller the loss value, the smaller the difference between the inversion value and the true value, which also means that the model fitting effect is better.

[0127] In this embodiment, the difference between two adjacent wind speeds is used to represent the gradient change between the two wind speeds, and the gradient change is used to represent the error. That is, the calculation method of the gradient change error between the two wind speeds includes:

[0128] ,

[0129] ,

[0130] ,

[0131] Among them, Δy p Represents the difference between the true value of the sample, Δf p Represents the difference of sample inversion values, y p+1 represents the true value of the p+1th sample, f p+1 Represents the inversion value of the p+1th sample. This loss function measures the prediction accuracy of the model by the gradient change between the inversion value and the true value.

[0132] In this embodiment, the overall value of 10 wind speeds is used to represent the overall wind speed error, that is, the calculation method of the overall wind speed error includes:

[0133] ,

[0134] The loss function measures the prediction accuracy of the model by the change between the overall inversion value of 10 wind speeds and the overall true value.

[0135] The update rules of the Adam optimizer include:

[0136] ,

[0137] Among them, s represents the number of iterations, g s represents the current gradient, Represents the loss function L relative to the model parameters The gradient of L represents the loss function, m s represents the first-order moment estimate of the gradient, β 1 represents the decay rate of the control first-order moment estimate, m s-1 represents the first-order moment estimate at the previous iteration, v s represents the second-order moment estimate of the gradient, β 2 represents the decay rate of the control second-order moment estimate, It represents the result after correcting the estimated value of the first-order moment estimate. represents the decay rate β 1 The sth power, It represents the result after correcting the estimated value of the second-order moment estimate. represents the decay rate β 1 The sth power, Represents the model parameters at the sth iteration Update value, α represents the learning rate, Represents a smoothing term to prevent the denominator from being zero.

[0138] This embodiment also provides a method for model testing:

[0139] After the sea surface wind speed is predicted, some indicators are needed to evaluate the accuracy of the inversion results. For regression problems, common evaluation indicators are root mean square error (RMSE) and mean absolute error (MAE). Root mean square error (RMSE) is a commonly used indicator to measure the size of the error between the predicted data and the reference data. Its mathematical formula is as follows:

[0140] ,

[0141] The smaller the RMSE value, the smaller the difference between the predicted data and the reference data, and the better the fitting effect; conversely, the larger the difference between the predicted data and the reference data, and the worse the fitting effect.

[0142] Mean Absolute Error (MAE) is the average of the absolute values ​​of the deviations of all individual observations from the arithmetic mean. The mean absolute error can avoid the problem of errors canceling each other out, and thus can accurately reflect the size of the actual prediction error. The calculation formula is as follows:

[0143] ,

[0144] The smaller the MAE value, the smaller the difference between the predicted data and the reference data, and the better the fitting effect; conversely, the larger the difference between the predicted data and the reference data, and the worse the fitting effect.

[0145] S3. Obtain the delay-Doppler correlation power diagram (DDM), GNSS-R auxiliary parameters and effective wave height of wind and wave surge, and use the same-track multi-time series sea surface wind speed inversion model to complete the sea surface wind speed inversion and obtain the sea surface wind speed prediction data.

[0146] In this embodiment, the GNSS-R auxiliary parameters include: longitude and latitude of the mirror reflection point, normalized bistatic radar cross section (NBRCS), DDM leading edge slope (LES), signal-to-noise ratio, incident angle, and range correction gain (RCG).

[0147] Embodiment 2

[0148] In this embodiment, a same-track multi-time series GNSS-R sea surface wind speed inversion system includes: a data acquisition module, a model building module and an inversion module.

[0149] The data acquisition 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 surge auxiliary data.

[0150] The model construction module uses the same-track neighborhood information fusion module and the multi-time double-layer LSTM module to build an initial wind speed inversion model, and uses the training sample set to train the initial wind speed inversion model to obtain the same-track multi-time series sea surface wind speed inversion model.

[0151] The inversion module is used to obtain the delay-Doppler correlation power diagram, GNSS-R auxiliary parameters and effective wave height of wind and wave surge, and use the same-track multi-time series sea surface wind speed inversion model to complete the sea surface wind speed inversion and obtain the sea surface wind speed prediction data.

[0152] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for inverting sea surface wind speed using a same-track multi-time series GNSS-R, characterized in that: The following steps are involved: Acquire 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 surge auxiliary data; An initial wind speed inversion model is constructed by using a same-track neighborhood information fusion module and a multi-time double-layer LSTM module, and the initial wind speed inversion model is trained by using the training sample set to obtain a same-track multi-time series sea surface wind speed inversion model; The delay-Doppler correlation power diagram, GNSS-R auxiliary parameters and effective wave height of wind and wave surge are obtained, and the sea surface wind speed inversion is completed by using the same-track multi-time series sea surface wind speed inversion model to obtain the sea surface wind speed prediction data.

2. According to claim 1, a same-track multi-time series GNSS-R sea surface wind speed inversion method is characterized in that: The method for preprocessing the historical satellite data includes: The range gain correction is calculated by the antenna gain at the mirror reflection point, the distance from the receiver to the mirror reflection point, and the distance from the navigation satellite to the mirror reflection point, and the range gain correction is performed on the historical satellite data: , Where RCG represents the range gain correction, G SP represents the antenna gain in the direction of the mirror emission point, Indicates the distance from the receiver to the mirror reflection point, R SP Indicates the distance from the navigation satellite to the mirror reflection point, 10 27 represents the scaling factor; Performing incident angle correction on the corrected data, and screening out data whose incident angle is less than a preset threshold value in the corrected data; The filtered data are corrected for the mirror reflection point type, in which: 0 represents open sea area, 0.5 represents coastal sea area within 25 kilometers from the land, 1 represents land, and 2 represents sea ice. The data with the mirror reflection point type less than 1 are filtered out to obtain the preprocessed data.

3. According to claim 1, a method for inverting sea surface wind speed using a same-track multi-time series GNSS-R is characterized in that: The method for constructing the training sample set includes: The U component and the V component of the 10-meter wind field measured by ECMWF are selected, and the reference wind speed is calculated based on the U component and the V component: , Where wspd represents the reference wind speed, u 10 represents the U component of the 10-meter wind field, v 10 represents the V component of the 10-meter wind field; The effective wave heights of wind waves and swell waves are selected as the swell wave auxiliary data, and the reference wind speed and the swell wave auxiliary data are bilinearly interpolated to obtain the training sample set.

4. According to claim 1, a method for inverting sea surface wind speed using a same-track multi-time series GNSS-R is characterized in that: The same track neighborhood information fusion module includes: a number of CNN modules and a number of BP modules; The CNN module is used to extract 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 are respectively composed of 32 3×3 filters and 16 3×3 filters, and 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 a maximum pooling operation, and are used to perform a maximum pooling operation on the first feature map to obtain a pooled feature map, with a pooling window size of 2×2 and a step size of 2; the flattening layer is used to flatten the pooled feature map into a first feature sequence with a length of 160; The BP module includes three fully connected layers, each of which is composed of 32 neurons, 64 neurons, and 128 neurons, respectively. The BP module is used to extract features of the GNSS-R auxiliary parameters and the effective wave height of the wind wave and surge to obtain a second feature sequence; The on-track neighborhood information fusion module connects the first feature sequence and the second feature sequence to obtain a fused feature sequence with a length of 288.

5. According to claim 1, a method for inverting sea surface wind speed using a same-track multi-time series GNSS-R is characterized in that: The multi-time double-layer LSTM module is composed of a 2-layer LSTM network, and each layer of the LSTM network includes: a forget gate, an input gate, a unit state update and an output gate; The forget gate comprises: , Among them, 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 state of the previous time step, x t represents the input of the current time step, b f Represents the bias vector of the forget gate; The input gate is composed of the first sigmoid layer and the first tanh layer, including: , , Among them, i t represents the first sigmoid layer, W i represents the weight matrix of the first sigmoid layer, b i 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; The unit status update includes: , Among them, C t Represents the updated unit state, C t-1 represents the cell state at the previous time step; The output gate is composed of the second sigmoid layer and the final hidden layer, including: , , Among them, t represents the second sigmoid layer, W o 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.

6. According to claim 1, a method for inverting sea surface wind speed using a same-track multi-time series GNSS-R is characterized in that: The method for training the initial wind speed inversion model includes: Initializing parameters in the initial wind speed inversion model, and transferring the training sample set into the initial wind speed inversion model; The training sample set obtains a predicted value through forward propagation and calculates a loss function; Using a back propagation algorithm and an Adam optimizer to update the parameters in the initial wind speed inversion model; When the change of the continuous iteration loss function is less than 0.0001 or reaches the set number of iterations, the training ends and the same-track multi-time series sea surface wind speed inversion model is obtained.

7. According to claim 6, a same-track multi-time series GNSS-R sea surface wind speed inversion method is characterized in that: The method for calculating the loss function includes: constructing the loss function using each wind speed error, the gradient change error between two wind speeds, and the overall wind speed error: , Among them, L represents the loss function, L MSE Represents each wind speed error, L D Represents the gradient change error between two wind speeds, L T represents the overall wind speed error; The calculation method of each wind speed error includes: , Where n represents the number of samples, y p represents the true value of the pth sample, f p represents the inversion value of the pth sample; The calculation method of the gradient change error between the two wind speeds includes: , , , Among them, Δy p Represents the difference between the true value of the sample, Δf p Represents the difference of sample inversion values, y p+1 represents the true value of the p+1th sample, f p+1 represents the inversion value of the p+1th sample; The calculation method of the overall wind speed error includes: 。 8. According to claim 6, a method for inverting sea surface wind speed using a same-track multi-time series GNSS-R is characterized in that: The update rules of the Adam optimizer include: , Among them, s represents the number of iterations, g s represents the current gradient, Represents the loss function L relative to the model parameters The gradient of L represents the loss function, m s represents the first-order moment estimate of the gradient, β1 represents the decay rate of the control first-order moment estimate, m s-1 represents the first-order moment estimate at the previous iteration, v s represents the second-order moment estimate of the gradient, β2 represents the decay rate of the control second-order moment estimate, It represents the result after correcting the estimated value of the first-order moment estimate. represents the attenuation rate β1 to the sth power, It represents the result after correcting the estimated value of the second-order moment estimate. represents the attenuation rate β1 to the sth power, Δω s Represents the model parameters at the sth iteration Update value, α represents the learning rate, represents the smoothing term.

9. A same-track multi-time series GNSS-R sea surface wind speed inversion system, the inversion system applying the inversion method according to any one of claims 1 to 8, characterized in that: include: Data acquisition module, model building module and inversion module; The data acquisition module is used to acquire 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 surge auxiliary data; The model building module uses the same-track neighborhood information fusion module and the multi-time double-layer LSTM module to build an initial wind speed inversion model, and uses the training sample set to train the initial wind speed inversion model to obtain a same-track multi-time series sea surface wind speed inversion model; The inversion module is used to obtain the delay-Doppler correlation power diagram, GNSS-R auxiliary parameters and effective wave height of wind wave swell, and use the same track multi-time series sea surface wind speed inversion model to complete the sea surface wind speed inversion to obtain sea surface wind speed prediction data.

Citation Information

Patent Citations

  • CNN-based satellite-borne GNSS-R sea surface wind speed inversion multi-modal deep learning method

    CN116148863A

  • Offshore wind field inversion method based on improved CNN-LSTM network

    CN118690669A

Cited By

  • METHOD AND SYSTEM FOR INVERSING WIND VELOCITY AT THE SEA SURFACE WITH TIME SERIES ON THE SAME ORBAL USING A COMBINATION OF GLOBAL SATELLITE NAVIGATION SYSTEM REFLECTOMETRY (GNSS-R) AND A SPREAD RADIATION METER

    BE1033022B1

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

    CN119647292A

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

    CN120277618A

  • Co-track time-series sea surface wind speed retrieval method and system based on GNSS-R and scatterometer

    CN120277618B

  • Multi-modal information fusion sea surface wind speed inversion method, device and system

    CN120337728A