A GNSS-R sea surface wind speed inversion method and system based on element-by-element attention fusion

Abstract

This invention discloses a GNSS-R sea surface wind speed inversion method and system based on element-wise attention fusion. By employing gated bottleneck convolution and element-wise attention fusion mechanisms, this invention fully utilizes the complementary information between the two-dimensional spatial structure of the DDM and auxiliary observation parameters, significantly improving the accuracy and stability of sea surface wind speed inversion. Compared with existing methods, this invention exhibits lower root mean square error and smaller system bias across the entire wind speed range, effectively mitigating underestimation problems, especially under high wind speed conditions, and demonstrating higher consistency between the inversion results and the reference wind speed. Simultaneously, this invention significantly reduces the number of network parameters and computational complexity while maintaining accuracy, possessing excellent lightweight characteristics and practical deployment capabilities, better meeting the needs of real-time wind speed inversion on satellites.

Description

Technical Field

[0001] This invention relates to the field of remote sensing technology of Global Navigation Satellite System Reflection Signal (GNSS-R), specifically to a GNSS-R sea surface wind speed inversion method and system based on element-wise attention fusion, which is particularly suitable for spaceborne GNSS-R sea surface wind field detection. Background Technology

[0002] The Global Navigation Satellite System (GNSS) provides positioning, navigation, and timing services globally. GNSS satellites continuously transmit L-band signals back to Earth. These signals have strong penetrating power and are relatively less affected by atmospheric conditions such as clouds, fog, rain, and snow, making them a stable and free source of Earth observation signals. When GNSS signals propagate to the sea surface, a portion is reflected. The intensity, delay, Doppler shift, and spatial distribution of the reflected signal are influenced by the sea surface conditions. Sea surface wind is a significant factor affecting sea surface roughness; different wind speeds result in different sea surface roughnesses, which in turn alter the reflection characteristics of GNSS signals, ultimately manifesting as changes in the reflected signal power and delayed Doppler distribution. Therefore, GNSS-R sea surface wind speed inversion essentially utilizes the physical relationship between "wind field changes—sea surface roughness changes—reflection characteristic changes—signal characteristic changes" to estimate sea surface wind speed from reflected signals.

[0003] Currently, GNSS-R sea surface wind speed inversion methods can be mainly divided into two categories. The first category is traditional inversion methods based on Geophysical Model Functions (GMFs). These methods typically utilize artificially constructed observation features such as NBRCS and LES to establish an empirical mapping relationship between GNSS-R observations and sea surface wind speed. Common forms include exponential functions, polynomial functions, and two-dimensional lookup tables. This type of method has a certain degree of physical interpretability and is relatively mature in application, but it relies heavily on artificial features and struggles to fully utilize the complex two-dimensional spatial structure information in Delay-Doppler Maps (DDMs). The second category is data-driven methods based on machine learning or deep learning. These methods use DDM data and its derived features as the main input, and mine deep features through models such as random forests, backpropagation neural networks, convolutional neural networks, and Transformers to establish a nonlinear mapping relationship between input observations and sea surface wind speed. Compared with traditional empirical functions, deep learning methods have stronger feature representation capabilities and the ability to model complex relationships. Summary of the Invention

[0004] To address the aforementioned technical challenges, this invention utilizes CYGNSS measured DDM data, auxiliary observation parameters, and reference sea surface wind speed data. It employs a deep neural network to learn the complex mapping relationship between DDM spatial structure, signal characteristics, observation geometry, instrument status, and sea surface wind speed. To address the issues of high parameter count in conventional convolution and the difficulty of simple feature fusion in fully characterizing the interaction between DDM and auxiliary information, this embodiment employs gated bottleneck convolution to extract local scattering features of the DDM, enhancing wind speed sensitivity while reducing parameter burden. An auxiliary information encoding module maps one-dimensional auxiliary parameters to a high-dimensional space consistent with DDM features. Furthermore, by combining spatial attention and channel attention, important delay-Doppler regions and key feature channels are highlighted. Element-by-element fusion achieves lightweight interaction between DDM features and auxiliary information, thereby improving the model's ability to express sea surface wind speed changes under complex observation conditions.

[0005] To achieve the above objectives, this invention provides a GNSS-R sea surface wind speed inversion method based on element-wise attention fusion, comprising the following steps: S1. Obtain the GNSS-R sample dataset. Each sample in the GNSS-R sample dataset contains multi-channel delayed Doppler image (DDM) data, auxiliary observation parameters, and the corresponding reference sea surface wind speed. S2. Based on the GNSS-R sample dataset, construct a GNSS-R sea surface wind speed inversion network; the GNSS-R sea surface wind speed inversion network includes a DDM feature extraction branch, an auxiliary information encoding branch, an attention feature fusion module, and a wind speed regression module; S3. Input the DDM data to be inverted and the corresponding auxiliary observation parameters into the GNSS-R sea surface wind speed inversion network to obtain the sea surface wind speed inversion results.

[0006] Preferably, S1 includes: Multi-channel delayed Doppler (DDM) data and auxiliary observation parameters were extracted from CYGNSS L1 level data products. The DDM data included bistatic radar cross section (BRC), effective scattering area (ASA), analog power, and raw counts. The auxiliary observation parameters included normalized bistatic radar cross section (NBRCS), ASA, signal-to-noise ratio (SNR), longitude of the mirror point, latitude of the mirror point, angle of incidence, pseudo-random noise code (PRN), and range correction gain. Based on the latitude and longitude of the mirror point and the observation time, the CYGNSS observation data and the ERA5 reanalysis wind field are spatiotemporally matched, and the reference sea surface wind speed is calculated based on the zonal and meridional wind components at a height of 10m provided by ERA5. The DDM data and auxiliary observation parameters were standardized, and the standardized samples were divided into training set, validation set and test set according to time order.

[0007] Preferably, the DDM feature extraction branch is used to extract DDM features, and the steps include: The input multi-channel DDM data is channel-mapped using 1×1 point convolution to obtain the initial DDM features; Bottleneck convolutional units are used to extract local spatial features from the initial DDM features. The bottleneck convolutional unit consists of a 1×1 point convolution for channel compression, a 3×3 depth convolution for extracting local spatial features, and a 1×1 point convolution for channel recovery. The DDM features are obtained by gating the features output by the bottleneck convolution using a gating unit.

[0008] Preferably, the auxiliary information encoding branch is used to encode auxiliary observation parameters into auxiliary features, and the steps include: A trainable weight matrix is ​​used to linearly project the auxiliary observation parameters, and the projection result is reconstructed into a single-channel two-dimensional feature map. A 1×1 point convolution is used to perform channel mapping on a single-channel two-dimensional feature map to obtain multi-channel two-dimensional auxiliary features; The auxiliary features are obtained by channel recalibration of the multi-channel two-dimensional auxiliary features.

[0009] Preferably, the attention feature fusion module is used to perform element-wise attention fusion of DDM features and auxiliary features to obtain fused features, the steps of which include: Spatial attention enhancement and channel attention enhancement are applied to the DDM features respectively to obtain DDM features with spatial and channel enhancement. The DDM features are processed by depthwise convolution and residual connection to obtain spatially enhanced DDM features; The spatially and channel-enhanced DDM features are interacted with the spatially enhanced DDM features through element-wise addition, and then with the DDM features through element-wise multiplication self-attention interaction to obtain the enhanced DDM features. Spatial attention enhancement and channel attention enhancement are applied to the auxiliary features. The enhanced auxiliary features are then subjected to element-wise multiplication self-attention interaction with the other auxiliary features to obtain the enhanced auxiliary features. The enhanced DDM features are multiplied element-wise with the enhanced auxiliary features to obtain the fused features.

[0010] Preferably, the wind speed regression module is a multilayer perceptron. The wind speed regression module is used to obtain the sea surface wind speed inversion result based on the fused features. The steps include: The fused features are flattened into a one-dimensional vector; the one-dimensional vector is then input into a multilayer perceptron, which contains two nonlinear units. Each nonlinear unit consists of a fully connected layer, a Dropout layer, and a ReLU activation function. The fully connected layer of the first nonlinear unit contains 32 neurons, and the fully connected layer of the second nonlinear unit contains 1 neuron. The output is the predicted sea surface wind speed.

[0011] Preferably, after acquiring the GNSS-R sample dataset, the method further includes quality control of the samples, wherein the quality control includes: Remove samples with wind speeds below 2.5 m / s; Retain samples where the spacecraft attitude quality flags satisfy quality_flags=4 and the nanostar tracker attitude status satisfies nst_att_status=0; Samples with a distance correction gain (RCG) greater than 3 are retained; Samples with a receive antenna gain sp_rx_gain greater than 0dB and a signal-to-noise ratio (SNR) greater than 1dB are retained; Remove samples with a front slope LES less than or equal to 0.

[0012] The present invention also provides a GNSS-R sea surface wind speed inversion system based on element-wise attention fusion. The system is used to implement the above method and includes: an acquisition module, a construction module, and an inversion module. The acquisition module is used to acquire GNSS-R sample datasets. Each sample in the GNSS-R sample dataset contains multi-channel delayed Doppler image (DDM) data, auxiliary observation parameters, and the corresponding reference sea surface wind speed. The construction module is used to construct a GNSS-R sea surface wind speed inversion network based on the GNSS-R sample dataset; wherein, the GNSS-R sea surface wind speed inversion network includes a DDM feature extraction branch, an auxiliary information encoding branch, an attention feature fusion module, and a wind speed regression module; The inversion module is used to input the DDM data to be inverted and the corresponding auxiliary observation parameters into the GNSS-R sea surface wind speed inversion network to obtain the sea surface wind speed inversion results.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention, through a gated bottleneck convolution and element-wise attention fusion mechanism, fully utilizes the complementary information between the two-dimensional spatial structure of the DDM and auxiliary observation parameters, significantly improving the accuracy and stability of sea surface wind speed inversion. Compared with existing methods, this invention exhibits lower root mean square error and smaller systematic bias across the entire wind speed range, effectively mitigating underestimation issues, especially under high wind speed conditions, and demonstrating higher consistency between the inversion results and the reference wind speed. Simultaneously, this invention significantly reduces the number of network parameters and computational complexity while maintaining accuracy, possessing excellent lightweight characteristics and practical deployment capabilities, better meeting the needs of spaceborne real-time wind speed inversion. Attached Figure Description

[0014] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the method flow framework of an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall structure of the GNSS-R sea surface wind speed inversion network according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the gated bottleneck convolutional feature extraction module according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the bottleneck convolution structure in an embodiment of the present invention; Figure 5 This is a schematic diagram of the auxiliary information encoding module structure according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the lightweight attention fusion module structure according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the spatial attention structure according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the channel attention structure according to an embodiment of the present invention; Figure 9 This is a schematic diagram of the depthwise convolution and residual structure according to an embodiment of the present invention; Figure 10 This is a schematic diagram of the multilayer perceptron regression module structure according to an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Example 1 The process framework of this embodiment is as follows: Figure 1 As shown, the specific implementation steps are as follows: S1. Obtain the GNSS-R sample dataset. Each sample in the GNSS-R sample dataset contains multi-channel delayed Doppler image (DDM) data, auxiliary observation parameters, and the corresponding reference sea surface wind speed.

[0019] This embodiment uses the multi-channel Delay-Doppler Map (DDM) and auxiliary observation parameters from the CYGNSSL1 level data product as model inputs, and ERA5 reanalysis wind speed as the reference label to construct a GNSS-R sea surface wind speed inversion sample dataset. First, four types of DDM observation variables are extracted from the CYGNSS data: bistatic radar cross section (BRCS), effective scattering area, simulated power, and raw counts. Each DDM has a size of [size missing]. Therefore, the input DDM data can be represented as: (1) Where 4 represents the number of DDM input channels, and 17 and 11 represent the delay dimension and Doppler dimension, respectively.

[0020] In addition to DDM data, this embodiment also extracts auxiliary parameters related to signal characteristics, observation geometry, and instrument status, including NBRCS, LES, SNR, mirror point longitude, mirror point latitude, incident angle, PRN, and RCG. These auxiliary parameters can be expressed as: (2) Where 8 represents the number of auxiliary parameters.

[0021] To construct wind speed monitoring labels, this embodiment performs spatiotemporal matching between CYGNSS observations and the ERA5 reanalysis wind field based on the latitude and longitude of the CYGNSS mirror point and the observation time. ERA5 provides the zonal wind component at a height of 10m. Meridional wind component The reference wind speed calculation formula is: (3) in, This represents the ERA5 reference wind speed. After matching, each training sample consists of DDM data, auxiliary parameters, and a reference wind speed label.

[0022] Before formally training the network, sample quality control is required. This embodiment removes outlier observation samples and retains valid samples that meet conditions such as satellite attitude quality, receiver gain, signal-to-noise ratio, and LES, in order to reduce the interference of low-quality observations on model training. Specifically: (1) Remove samples with wind speeds below 2.5 m / s to reduce the uncertainty caused by changes in the sea surface scattering mechanism under low wind speed conditions; (2) Retain observation samples with high spacecraft attitude quality, i.e., quality flags satisfy quality_flags=4 and nanostar tracker attitude status satisfies nst_att_status=0; (3) Retain samples with a distance correction gain (RCG) greater than 3; (4) Retain observation samples with a receiving antenna gain (sp_rx_gain) greater than 0dB and a signal-to-noise ratio (SNR) greater than 1dB; (5) Remove abnormal samples with LES less than or equal to 0.

[0023] Since there are significant differences in the dimensions of different input variables, this embodiment standardizes the DDM data and auxiliary parameters. The standardization formula is as follows: (4) in, Represents the original input variables. and Let represent the mean and standard deviation of the training set, respectively. This represents the standardized variable.

[0024] To avoid data leakage, and The calculations are performed only on the training set and uniformly applied to the validation and test sets. Finally, the samples are divided into training, validation, and test sets in chronological order. The training set is used for network parameter learning, the validation set is used for model selection and early stopping detection, and the test set is used for final performance evaluation.

[0025] S2. Based on the GNSS-R sample dataset, construct the GNSS-R sea surface wind speed inversion network.

[0026] The GNSS-R sea surface wind speed inversion network constructed in this embodiment mainly consists of a DDM feature extraction branch, an auxiliary information encoding branch, a lightweight attention fusion module, and a wind speed regression module. The model input is multi-channel DDM data. and auxiliary parameters The output is the estimated sea surface wind speed at the corresponding observation location. The overall structure of the model is as follows: Figure 2 As shown.

[0027] (1) DDM feature extraction branch.

[0028] First, to improve the feature representation capability of DDM, this embodiment adopts... Point convolution Perform channel mapping to project the input DDM from 4 channels to 32 channels to obtain the initial DDM features: (5) in, express Point convolution, with size , The nonlinear transformation function consisting of layer normalization and the ReLU activation function represents the representation. .

[0029] The overall structure of the module is as follows Figure 3 As shown, this embodiment uses a gated bottleneck convolution feature extraction module to extract local features of the DDM. The bottleneck convolution BootConv first... Point convolution Compress the channel dimension, and then utilize depthwise convolution Extract local spatial features, and finally through Point convolution Restore the channel dimension. The bottleneck convolutional BottConv structure is as follows: Figure 4 As shown.

[0030] Single bottleneck convolutional unit It can be represented as: (6) (7) in, Indicates input features, Indicates the use of channel compression Point convolution, with size , express Depthwise convolution, with size , Indicates the use of channel recovery Point convolution, with size , It represents the nonlinear transformation function composed of layer normalization and ReLU activation function.

[0031] To further enhance the effective feature response related to wind speed inversion, this embodiment introduces a gating mechanism based on bottleneck convolution. The gated bottleneck convolution module contains two local feature units and one gating unit, and its calculation process is as follows: (8) (9) in, and Represents local feature units, Indicates a gating unit. Representation layer normalization, This represents the Sigmoid function. This represents element-wise multiplication. This represents the final DDM features. It is achieved through gating weights. It can adaptively adjust the characteristic response of different spatial locations and channels, enhance wind speed sensitive areas, and suppress redundant information.

[0032] (2) Auxiliary information encoding branch.

[0033] Since auxiliary parameters exist as one-dimensional vectors, while DDM features have a two-dimensional spatial structure, they differ in data form and feature dimension. To enable auxiliary information to participate in subsequent element-wise fusion, this embodiment designs an auxiliary information encoding module to map one-dimensional auxiliary parameters into a high-dimensional feature representation consistent with the DDM feature space size and channel dimension. The module structure is as follows: Figure 5 As shown.

[0034] First, the auxiliary parameters are linearly projected using a trainable weight matrix and reconstructed into a single-channel two-dimensional feature map: (10) in, This represents a trainable projection matrix. Indicates the bias term. , Subsequently, through Point convolution ,size Channel mapping is performed and the ReLU activation function is used to introduce nonlinear expressive power: (11) in, , This represents the ReLU function. To further enhance the significant channel responses in the auxiliary features, this embodiment introduces Global Max Pooling (GMP) to generate channel weights and recalibrates the auxiliary features: (12) in, This represents the final auxiliary feature, whose size is the same as the DDM feature. Maintain consistency. This represents the Sigmoid function.

[0035] (3) Lightweight attention feature fusion.

[0036] After obtaining the DDM features and auxiliary features, this embodiment employs a lightweight attention fusion module to achieve feature fusion. Important regions and key feature channels of the DDM features are enhanced by simplifying spatial attention and channel attention. Spatial features are strengthened through convolutional residual structures, and then self-attention-enhanced features of the DDM features are constructed using additive operations and element-wise multiplication. For auxiliary data features, the model enhances their important regions and key feature channels by simplifying spatial attention and channel attention, and then constructs self-attention-enhanced features of the auxiliary features using element-wise multiplication. Finally, lightweight interaction between the enhanced DDM image features and the enhanced auxiliary data features is achieved using element-wise multiplication. The module structure is as follows: Figure 6 As shown.

[0037] This module first enhances the DDM features in both spatial and channel dimensions. In the spatial attention branch, the model performs max pooling along the channel dimension to highlight the significant spatial response of the DDM in the delay-Doppler plane, as shown in the structure below. Figure 7 As shown, it can be represented as: (13) (14) in, This represents max pooling along the channel dimension. This indicates that the process is composed of layer normalization, ReLU activation, and... Point convolution (size is) The spatial attention mapping function is composed of ) Represents spatial attention weights. This represents spatially enhanced features.

[0038] In the channel attention branch, the model uses global average pooling and global max pooling to extract the overall statistical response and salient activation response of the channel, respectively, and generates channel attention weights, with the structure as follows: Figure 8 As shown, it can be represented as: (15) (16) in, and These represent global average pooling and global max pooling, respectively. and This represents the nonlinear mapping function in the corresponding branch, derived from... Point convolution (size is) It consists of the ReLU activation function and the ReLU activation function. Indicates channel attention weights. This represents the features enhanced by spatial attention and channel attention.

[0039] Since the spatial distribution of DDM features is closely related to the energy distribution on the sea surface, this embodiment uses depthwise convolution and residual structures to further enhance the spatial features. The module structure is as follows: Figure 9 As shown. DDM features are achieved through a Depthwise convolution DWConv (size is Spatial information within each channel is extracted, then nonlinear activation is performed using ReLU, and finally, the channels are linked using a residual structure. This process can be represented as: (17) in Represents the ReLU function. To enhance spatial characteristics. In obtaining and Afterwards, it was used as and Then, use element-wise addition to achieve lightweight interaction between the two, and then... As Build self-attention interactions using element-wise multiplication: (18) For auxiliary data features This embodiment designs a simplified structure, using the same spatial attention and channel attention modules as described above to enhance the auxiliary data features, resulting in... Then, element-wise multiplication is performed with the original input to achieve self-attention modeling of the auxiliary data, which can be represented as: (19) Finally, the enhanced DDM features are multiplied element-wise with the enhanced auxiliary data features to obtain the final fused features: (20) in, This represents the final fused feature. This fusion method does not require constructing a large-scale global attention matrix; instead, it achieves lightweight interaction between DDM spatial response, channel correlation, and auxiliary observation information through element-wise multiplication. Compared to simple concatenation or addition, element-wise fusion can more directly characterize the correspondence between DDM features and auxiliary observation conditions, thereby enhancing the model's ability to express wind speed changes under complex observation conditions.

[0040] (4) Multilayer perceptron regression module The network first extracts local scattering features through the DDM branch, generates auxiliary feature representations through the auxiliary information encoding branch, and then obtains fused features through the element-wise attention fusion module. Subsequently, the fused features are flattened into a one-dimensional vector: (twenty one) in, This represents the fused feature vector input to the multilayer perceptron.

[0041] The multilayer perceptron consists of two nonlinear units, each comprising a fully connected layer, a Dropout layer, and a ReLU activation function. Nonlinear unit 1 has a fully connected layer with 32 neurons, while nonlinear unit 2 has a fully connected layer with 1 neuron. These components perform a nonlinear regression mapping from fused features to sea surface wind speed, as shown in the diagram. Figure 10 As shown, it can be represented as: (twenty two) in, This represents the sea surface wind speed predicted by the network.

[0042] This embodiment uses mean squared error as the training loss function to constrain the numerical difference between the predicted wind speed and the reference wind speed: (twenty three) in, This indicates the number of samples in a training batch. Indicates the first Predicted wind speed for each sample This represents the corresponding reference wind speed. During training, the model calculates the predicted wind speed and loss value through forward propagation, and then updates the network parameters through backpropagation, gradually bringing the predicted wind speed closer to the reference wind speed. Training uses the Adam optimizer, and the validation set is used to monitor model performance. Training is stopped early when the validation set loss no longer decreases within several consecutive epochs to avoid overfitting. After training, the model parameters that perform best on the validation set are saved as the final model.

[0043] S3. Input the DDM data to be inverted and the corresponding auxiliary observation parameters into the GNSS-R sea surface wind speed inversion network to obtain the sea surface wind speed inversion results.

[0044] The DDM features are extracted through the DDM feature extraction branch, and the auxiliary observation parameters are encoded into auxiliary features through the auxiliary information encoding branch. The DDM features and auxiliary features are fused element by element through the attention feature fusion module to obtain the fused features. The sea surface wind speed inversion result is obtained through the wind speed regression module based on the fused features.

[0045] Example 2 The advantages of this invention compared to existing technologies will be described in detail below. This invention can enhance the complementary representation of local scattering characteristics, spatial response, channel correlation, and auxiliary observation information with a lower parameter burden, thereby improving the accuracy, stability, and practical application capability of GNSS-R sea surface wind speed inversion.

[0046] (1) Lightweight analysis of gated bottleneck convolution module.

[0047] The model in this invention employs a bottleneck convolution (BottConv) structure in the DDM feature extraction stage to reduce the parameter burden caused by ordinary convolution. For feature maps with both 32 input and 32 output channels, directly using ordinary convolution would be less efficient. Convolution, whose kernel parameters can be expressed as: (twenty four) in, Indicates the kernel size. and These represent the number of input channels and the number of output channels, respectively. For ordinary... convolution, .

[0048] Bottleneck convolution first compresses the input channels to a lower dimension, then performs depthwise convolution and channel restoration. Depthwise convolution performs spatial convolution only within each channel, further reducing the computational burden during feature extraction. Its parameter count can be expressed as: (25) in, This indicates the number of intermediate channels in the bottleneck layer. This invention sets the compression ratio. Therefore when , Sometimes, Under this setting, normal The number of parameters in a convolution (excluding bias terms) is: (26) The number of parameters for the bottleneck convolution in this invention is: (27) Therefore, a single bottleneck convolutional unit is relative to ordinary The parameter reduction ratio of convolution is: (28) This demonstrates that bottleneck convolution can significantly reduce the number of convolution parameters while maintaining the ability to extract local spatial features. Building upon this, the introduction of a gated branch enables the model to not only have a low parameter count but also adaptively enhance wind speed-sensitive features based on the local response of the DDM. Therefore, the gated bottleneck convolution module balances lightweight feature extraction and effective feature enhancement, providing a more efficient DDM feature representation for subsequent attention fusion and wind speed regression.

[0049] (2) Lightweight analysis of feature fusion module.

[0050] If only the spatial correlation modeling process is considered, standard self-attention needs to be constructed. Correlation matrix, N=17 in this invention 11, C=32, its complexity is This invention employs element-wise multiplication for feature interaction, avoiding global pairwise correlation calculations. The complexity of this part is approximately [missing information]. Through the above design, the proposed fusion module can enhance the interaction between DDM spatial features, channel features and auxiliary information at a lower computational cost, thereby improving the model's adaptability to diverse observation conditions.

[0051] (3) Accuracy of sea surface wind speed inversion.

[0052] After the model training is completed, the CYGNSSDDM data from the test set and the auxiliary observation parameters are input into the trained network to obtain the sea surface wind speed estimate at the corresponding mirror point location. To evaluate the model inversion performance, this invention compares the network output wind speed with the ERA5 reference wind speed, using root mean square error, mean deviation, and correlation coefficient as evaluation indicators.

[0053] Root mean square error is used to measure the overall inversion error of the model: (29) The average deviation is used to describe whether the model has a systematic overestimation or underestimation: (30) The correlation coefficient is used to measure the linear consistency between predicted wind speed and reference wind speed. (31) in, and These represent the average values ​​of the predicted wind speed and the reference wind speed, respectively.

[0054] The GNSS-R data used was CYGNSSL1, and the reference wind speed data used was ERA5. The dataset contains approximately 750 samples from August 2023 to March 2024. Approximately 3 million samples from August 2023 to March 2024 were used for model training; approximately 1.6 million samples from April 2024 to July 2024 were used for model validation; and approximately 2.9 million samples from August 2024 to March 2025 were used for model testing.

[0055] To comprehensively evaluate the wind speed inversion performance of different models, this invention divides the test samples into all samples, m / s and Statistics were compiled for three cases: m / s. Among them, m / s corresponds to the low to medium wind speed range where the sample size is relatively concentrated. m / s corresponds to a high wind speed range where samples are relatively scarce. Table 1 presents the RMSE, Bias, and CC results for CyGNSSnet, DDM-former, PA-CNN, STG-DNN, and the method of this invention within different statistical ranges. Table 1 compares the performance of different models across all samples and in the two wind speed ranges.

[0056] Table 1 As shown in Table 1, the method of this invention achieves good overall inversion performance on the entire test set. Compared with CyGNSSnet, DDM-former, PA-CNN, and STG-DNN, the method of this invention has the lowest RMSE (1.385 m / s) and the highest CC (0.845), indicating that it can more effectively reduce the overall wind speed inversion error and improve the consistency between predicted and reference wind speeds. Compared with the best-performing comparative model PA-CNN, the RMSE of the method of this invention decreases from 1.426 m / s to 1.385 m / s, a reduction of approximately 2.88%. Meanwhile, the bias of the method of this invention is -0.022 m / s, close to 0, indicating that the model does not exhibit significant systematic overestimation or underestimation on the entire test set.

[0057] In the low-to-medium wind speed range of 0–15 m / s, the method of this invention also achieved the lowest RMSE and the highest CC, at 1.332 m / s and 0.841, respectively. Compared with PA-CNN, the RMSE is reduced by approximately 2.70%; compared with CyGNSSnet, DDM-former, and STG-DNN, the method of this invention also exhibits lower error and higher correlation. This indicates that under the condition of sufficient sample size in low-to-medium wind speeds, the proposed model can more stably learn the mapping relationship between GNSS-R observation features and sea surface wind speed. The low Bias (0.007 m / s) also indicates that the method of this invention has almost no systematic bias in this wind speed range.

[0058] In the high wind speed range (>15 m / s), the RMSE of all models increases significantly, and the bias is negative, indicating that high wind speed inversion is difficult and generally underestimates the wind speed. The method of this invention achieves the lowest RMSE in this range (4.555 m / s) and a bias of -3.858 m / s, with absolute values ​​smaller than other models, indicating that the method of this invention can alleviate the underestimation problem in high wind speeds to some extent. Although the CC of the method of this invention is 0.253, slightly lower than PA-CNN and STG-DNN, its RMSE and bias performance are superior, indicating that the model has better error control capabilities under high wind speed conditions.

[0059] Overall, the method of this invention achieves optimal RMSE and CC on the entire test set and in the 0–15 m / s range, and achieves the lowest RMSE and minimum absolute bias in the high wind speed range >15 m / s. The results show that the proposed model can effectively improve the accuracy of GNSS-R sea surface wind speed inversion, especially demonstrating good comprehensive performance in error control and bias suppression.

[0060] Example 3 The present invention also provides a GNSS-R sea surface wind speed inversion system based on element-wise attention fusion, comprising: an acquisition module, a construction module, and an inversion module; The acquisition module is used to acquire GNSS-R sample datasets. Each sample in the GNSS-R sample dataset contains multi-channel delayed Doppler image (DDM) data, auxiliary observation parameters, and the corresponding reference sea surface wind speed. The construction module is used to build a GNSS-R sea surface wind speed inversion network based on the GNSS-R sample dataset; the GNSS-R sea surface wind speed inversion network includes a DDM feature extraction branch, an auxiliary information encoding branch, an attention feature fusion module, and a wind speed regression module; The inversion module is used to input the DDM data to be inverted and the corresponding auxiliary observation parameters into the GNSS-R sea surface wind speed inversion network to obtain the sea surface wind speed inversion results.

[0061] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A GNSS-R sea surface wind speed inversion method based on element-wise attention fusion, characterized in that the steps are as follows: include: S1. Obtain the GNSS-R sample dataset. Each sample in the GNSS-R sample dataset contains multi-channel delayed Doppler image (DDM) data, auxiliary observation parameters, and the corresponding reference sea surface wind speed. S2. Based on the GNSS-R sample dataset, construct a GNSS-R sea surface wind speed inversion network; the GNSS-R sea surface wind speed inversion network includes a DDM feature extraction branch, an auxiliary information encoding branch, an attention feature fusion module, and a wind speed regression module; S3. Input the DDM data to be inverted and the corresponding auxiliary observation parameters into the GNSS-R sea surface wind speed inversion network to obtain the sea surface wind speed inversion results.

2. The GNSS-R sea surface wind speed inversion method based on element-wise attention fusion according to claim 1, characterized in that, S1 includes: Multi-channel delayed Doppler (DDM) data and auxiliary observation parameters were extracted from CYGNSS L1 level data products. The DDM data included bistatic radar cross section (BRC), effective scattering area (ASA), analog power, and raw counts. The auxiliary observation parameters included normalized bistatic radar cross section (NBRCS), ASA, signal-to-noise ratio (SNR), longitude of the mirror point, latitude of the mirror point, angle of incidence, pseudo-random noise code (PRN), and range correction gain. Based on the latitude and longitude of the mirror point and the observation time, the CYGNSS observation data and the ERA5 reanalysis wind field are spatiotemporally matched, and the reference sea surface wind speed is calculated based on the zonal and meridional wind components at a height of 10m provided by ERA5. The DDM data and auxiliary observation parameters were standardized, and the standardized samples were divided into training set, validation set and test set according to time order.

3. The GNSS-R sea surface wind speed inversion method based on element-wise attention fusion according to claim 1, characterized in that, The DDM feature extraction branch is used to extract DDM features, and the steps include: The input multi-channel DDM data is channel-mapped using 1×1 point convolution to obtain the initial DDM features; Bottleneck convolutional units are used to extract local spatial features from the initial DDM features. The bottleneck convolutional unit consists of a 1×1 point convolution for channel compression, a 3×3 depth convolution for extracting local spatial features, and a 1×1 point convolution for channel recovery. The DDM features are obtained by gating the features output by the bottleneck convolution using a gating unit.

4. The GNSS-R sea surface wind speed inversion method based on element-wise attention fusion according to claim 1, characterized in that, The auxiliary information encoding branch is used to encode auxiliary observation parameters into auxiliary features, and the steps include: A trainable weight matrix is ​​used to linearly project the auxiliary observation parameters, and the projection result is reconstructed into a single-channel two-dimensional feature map. A 1×1 point convolution is used to perform channel mapping on a single-channel two-dimensional feature map to obtain multi-channel two-dimensional auxiliary features; The auxiliary features are obtained by channel recalibration of the multi-channel two-dimensional auxiliary features.

5. The GNSS-R sea surface wind speed inversion method based on element-wise attention fusion according to claim 1, characterized in that, The attention feature fusion module is used to perform element-wise attention fusion of DDM features and auxiliary features to obtain fused features. The steps include: Spatial attention enhancement and channel attention enhancement are applied to the DDM features respectively to obtain DDM features with spatial and channel enhancement. The DDM features are processed by depthwise convolution and residual connection to obtain spatially enhanced DDM features; The spatially and channel-enhanced DDM features are interacted with the spatially enhanced DDM features through element-wise addition, and then with the DDM features through element-wise multiplication self-attention interaction to obtain the enhanced DDM features. Spatial attention enhancement and channel attention enhancement are applied to the auxiliary features. The enhanced auxiliary features are then subjected to element-wise multiplication self-attention interaction with the other auxiliary features to obtain the enhanced auxiliary features. The enhanced DDM features are multiplied element-wise with the enhanced auxiliary features to obtain the fused features.

6. The GNSS-R sea surface wind speed inversion method based on element-wise attention fusion according to claim 1, characterized in that, The wind speed regression module is a multilayer perceptron. It is used to obtain sea surface wind speed inversion results based on fused features. The steps include: The fused features are flattened into a one-dimensional vector; the one-dimensional vector is then input into a multilayer perceptron, which contains two nonlinear units. Each nonlinear unit consists of a fully connected layer, a Dropout layer, and a ReLU activation function. The fully connected layer of the first nonlinear unit contains 32 neurons, and the fully connected layer of the second nonlinear unit contains 1 neuron. The output is the predicted sea surface wind speed.

7. The GNSS-R sea surface wind speed inversion method based on element-wise attention fusion according to claim 2, characterized in that, After acquiring the GNSS-R sample dataset, the process also includes quality control of the samples, which includes: Remove samples with wind speeds below 2.5 m / s; Retain samples where the spacecraft attitude quality flags satisfy quality_flags=4 and the nanostar tracker attitude status satisfies nst_att_status=0; Samples with a distance correction gain (RCG) greater than 3 are retained; Samples with a receive antenna gain sp_rx_gain greater than 0dB and a signal-to-noise ratio (SNR) greater than 1dB are retained; Remove samples with a front slope LES less than or equal to 0.

8. A GNSS-R sea surface wind speed inversion system based on element-wise attention fusion, the system being used to implement the method described in any one of claims 1-7, characterized in that, include: Acquisition module, construction module, inversion module; The acquisition module is used to acquire GNSS-R sample datasets. Each sample in the GNSS-R sample dataset contains multi-channel delayed Doppler image (DDM) data, auxiliary observation parameters, and the corresponding reference sea surface wind speed. The construction module is used to construct a GNSS-R sea surface wind speed inversion network based on the GNSS-R sample dataset; wherein, the GNSS-R sea surface wind speed inversion network includes a DDM feature extraction branch, an auxiliary information encoding branch, an attention feature fusion module, and a wind speed regression module; The inversion module is used to input the DDM data to be inverted and the corresponding auxiliary observation parameters into the GNSS-R sea surface wind speed inversion network to obtain the sea surface wind speed inversion results.

Images