A multi-modal deep learning method for space-borne GNSS-R sea surface wind speed retrieval based on CNN

Abstract

The application discloses a kind of multi-modal deep learning methods for satellite-borne GNSS-R sea surface wind speed inversion based on CNN, GloWS-Net multi-modal deep learning model is constructed using the DDM features and multiple auxiliary parameters extracted by CNN fusion, comprising the following steps: obtaining satellite-borne GNSS-R observation data, ERA5 wind speed, ERA5 swell height, ERA5 wave direction data and IMERG rainfall data;All the data sets obtained are preprocessed and spatio-temporal matched;Data quality control and data division;GloWS-Net model construction and training;the test data set is input into the trained GloWS-Net model, the inversion wind speed value is obtained, and the result is evaluated.GloWS-Net model includes convolutional layer for extracting effective features from the combination of BRCSDDM and corresponding effective scattering area (effective scattering area), and fully connected layer for processing auxiliary parameters and higher level input parameters.Using the technical scheme of the application, high-precision and high-spatial and temporal resolution global ocean wind speed inversion can be realized.

Description

Technical Field

[0001] This invention belongs to the field of GNSS marine remote sensing monitoring technology, and in particular relates to a multimodal deep learning method for sea surface wind speed inversion based on CNN in spaceborne GNSS-R. Background Technology

[0002] Sea surface wind speed is a key factor influencing ocean circulation and global climate. As one of the most severe natural disasters, tropical cyclones, typhoons, and hurricanes generated by high wind speeds severely damage infrastructure and endanger lives. Furthermore, sea surface wind speed is also a crucial element in meteorological observation of sea surface information, playing a vital role in weather forecasting, climate pattern research, and aviation safety. Global Navigation Satellite System-Reflectometry (GNSS-R) technology uses a receiver mounted on a satellite to receive direct signals from GNSS satellites and echo signals reflected from reflective surfaces. Through processing, a time-delay Doppler image is obtained, along with the corresponding two-dimensional correlation power of the reflected signal. Using certain inversion methods, the physical parameters of the Earth's surface scattering surface can be obtained. Spaceborne GNSS-R has significant advantages such as short revisit periods, low observation costs, and high spatiotemporal resolution, enabling all-weather, all-time, and wide-coverage global sea surface wind speed inversion.

[0003] Currently, research on sea surface wind speed inversion methods for spaceborne GNSS-R mainly includes three categories: (1) Empirical model method. Some features related to sea surface wind speed are extracted from the DDM, such as leading edge slope (LES) and normalized bistatic radar cross section (NBRCS). Then, by fitting an empirical function that links DDM observations with wind speed, linear regression is applied to construct the GMF, and then the minimum variance estimator (MVE) is used to combine the inversion results of individual observations and eliminate residuals. However, the MVE method still has limited improvement in accuracy compared to the inversion of individual observations. It is very difficult to consider parameters such as incident angle, reflection geometry parameters (such as the latitude and longitude of the specular reflection point) and sea state (such as significant wave height) as model input parameters, making it difficult to develop a joint model. (2) Machine learning method. Fully connected network (FCN) is applied to develop multi-feature wind speed inversion models. However, the improvement is mainly in wind speeds of 5-10 m / s. (3) Deep learning method. Existing deep learning-based wind speed retrieval models outperform empirical models in accuracy; however, the uneven distribution of wind speed samples results in poor performance at high wind speeds. Furthermore, existing methods neglect the impact of rainfall and ocean waves on reflected signals.

[0004] In summary, this invention proposes a multimodal deep learning method for sea surface wind speed inversion based on CNN in spaceborne GNSS-R. Summary of the Invention

[0005] The technical problem to be solved by this invention is to overcome the shortcomings of the existing technology and provide a multimodal deep learning method for sea surface wind speed inversion based on CNN in spaceborne GNSS-R. The method uses CNN to extract effective features from BRCS DDM and the corresponding effective scattering area, and integrates basic GNSS-R auxiliary parameters and higher-level input parameters (such as ocean swell height, sea surface rainfall and wave direction information) to construct an end-to-end multimodal deep learning wind speed inversion model, which can effectively improve the accuracy of global sea surface wind speed inversion in spaceborne GNSS-R.

[0006] To solve the above-mentioned technical problems, the basic concept of the technical solution adopted by the present invention is as follows:

[0007] A multimodal deep learning method for sea surface wind speed inversion based on CNN in spaceborne GNSS-R includes the following steps:

[0008] Step S1: Acquire spaceborne GNSS-R observation data, ERA5 wind speed, ERA5 surge height, ERA5 wave direction data, and IMERG precipitation data;

[0009] Step S2: Preprocess and perform spatiotemporal matching on all acquired datasets;

[0010] Step S3: Data quality control and data partitioning;

[0011] Step S4 involves building and training the GloWS-Net model. The GloWS-Net model has three input lines: the first for the Normalized Bistatic Radar Cross Section (NBRCS) DDM, the second for the effective scattering area, and the third for auxiliary parameters. Specifically, the third line inputs auxiliary variables such as the Normalized Bistatic Radar Cross Section (NBRCS), leading edge slope (LES), trailing edge slope (TES), signal-to-noise ratio (SNR), receiver antenna gain (sp_rx_gain), latitude and longitude of the specular reflection point, incident angle, surge height, wave direction, and rainfall intensity.

[0012] Step S5: Input the test dataset into the trained GloWS-Net model to obtain the inverted wind speed value and evaluate the result.

[0013] Preferably, the spaceborne GNSS-R observation data includes L1b BRCS DDM, effective scattering area, normalized bistatic radar cross section (NBRCS), leading edge slope (LES), trailing edge slope (TES), signal-to-noise ratio (SNR), receiver antenna gain (sp_rx_gain), latitude and longitude of the specular reflection point, and angle of incidence; the ERA5 wind speed is the combined wind speed of the U and V components 10m above the sea surface; the ERA5 swell height is the SWH (significant height of total swell); the ERA5 wave direction data is the mean wave direction provided in the ERA5 product; and the IMERG rainfall data is the IMERG-F rainfall intensity.

[0014] Furthermore, step S2 includes the following sub-steps,

[0015] Step S2.1: Downsample the CYGNSS GNSS-R data;

[0016] Step S2.2: Spatiotemporal matching of CYGNSS GNSS-R downsampled data with ECMWF ERA5 wind speed, ERA5 wave direction, ERA5 swell SWH and IMERG precipitation data to obtain the matched dataset.

[0017] Preferably, the post-matching dataset quality control described in step S3 includes deleting all observations containing NaN values; discarding all observations less than 0; ensuring the RCG value is greater than 3; discarding observations if the gain (sp_rx_gain) of the receiving antenna in the direction of the reflection point is less than 0 dBi; and discarding observations if the uncertainty of BRCS (ddm_BRCS_uncert) is greater than 1. Data filtering in CYGNSS L1B data uses the Basic Quality Control (QC) flag. Data filtering is applied using a single QC flag bit, rather than the entire QC flag bit (“Overall Quality Poor”, least significant bit or bit 0). Specifically, invalid or abnormal DDM data (bits 4, 7, 8, 9, 10, 15, 17-20) is deleted; data from other instruments with data transmission and calibration problems are deleted (bits 1, 5, 6, 13, 14, 21-26); data with uncertain transmitter power or antenna gain height must be deleted (bits 16 and 27); spacecraft data with large attitude errors or abnormal attitudes are deleted (bits 2, 3, and 28); L1B observations containing GPS IIF satellites are deleted (“low_quality_GPS_ant_knowledge”, bit 27); the quality control flags “sp_near_land” (bit 12) and “sp_very_near_land” (bit 11) are used to delete data collected when signals reflect on land; data with specular reflection points more than 25 km from land are selected to reduce modeling errors. The filtered dataset described in step S3 is randomly divided into a training set, a validation set, and a test set, accounting for 30%, 15%, and 55% of the filtered dataset, respectively.

[0018] Preferably, the GloWS-Net model described in step S4 consists of three input lines: the first line is used to input BRCSDDM, the second line is used to input the effective scattering area, and the third line is used to input auxiliary parameters.

[0019] As a preferred option, the training process of the GloWS-Net global sea surface wind speed inversion model is specifically implemented as follows.

[0020] After preprocessing all the acquired datasets, the filtered datasets are divided into training set, validation set and test set;

[0021] The filtered dataset was normalized to zero mean and unit variance based on features. A validation set was used to avoid overfitting, employing early stopping with six epochs. The learning rate was set to 0.0005. To further prevent overfitting, dropout was set to 0.1.

[0022] The rectified linear unit (ReLU) is used as the activation function, Adam is selected as the optimizer, and mean squared error (MSE) is used as the loss function.

[0023] The training set is used to train the network model, and the validation set is used to supervise the model training. The batch size is set to 32, meaning that each time the network updates its weight coefficients, it uses only samples from one batch to train the model. To ensure performance quality, the number of epochs is set to 100.

[0024] Once training is complete, the test dataset is input into the trained GloWS-Net model to obtain the inverted wind speed values, and the results are evaluated.

[0025] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art.

[0026] This invention employs a CNN-based multimodal deep learning model (GloWS-Net) to achieve sea surface wind speed inversion for spaceborne GNSS-R. The GloWS-Net model includes convolutional layers for extracting effective features from a combination of the BRCS DDM and the corresponding effective scattering area, and fully connected layers for processing auxiliary parameters and higher-level input parameters. The combination of CNN and handcrafted features in the GloWS-Net multimodal deep learning model further improves inversion performance. The optimal architecture was determined on a validation set and evaluated on the training dataset to verify the model's generality. During training the GloWS-Net model, several traditional machine learning strategies were used to prevent overfitting, including batch normalization and early stopping. The GloWS-Net model achieves high inversion accuracy for low, medium, and high wind speeds, producing high-precision and high spatiotemporal resolution wind speed inversion results for the global ocean.

[0027] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0028] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention, but do not constitute an undue limitation of the invention. Obviously, the drawings described below are merely some embodiments, and those skilled in the art can obtain other drawings based on these drawings without creative effort. In the drawings:

[0029] Figure 1 This is a flowchart of a multimodal deep learning method for sea surface wind speed inversion based on CNN in an embodiment of the present invention.

[0030] Figure 2 This is a structural diagram of the multimodal deep learning model (GloWS-Net) for sea surface wind speed inversion based on CNN in an embodiment of the present invention.

[0031] Figure 3 This is a scatter plot of the wind speed inverted by the GloWS-Net model and the ERA5 wind speed in this embodiment of the invention.

[0032] Figure 4 This is a histogram showing the deviation distribution between the wind speed retrieved from the GloWS-Net model and the ERA5 wind speed in an embodiment of the present invention. It should be noted that these figures and descriptions are not intended to limit the scope of the invention in any way, but rather to illustrate the concept of the invention to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0034] Example 1

[0035] To verify the advantages of the method proposed in this invention, CYGNSS GNSS-R data, ERA5 wind speed, ERA5 surge height, ERA5 wave direction data, and IMERG rainfall data for all days of 2021 were obtained for experiments. The experimental results of this invention were compared with the wind speed inversion results of empirical models and some current network models (such as FCN and CNN models). The basic configuration of the experimental platform in this embodiment is shown in Table 1:

[0036] Table 1. Configuration of the experimental platform

[0037]

[0038] A multimodal deep learning method for sea surface wind speed inversion based on CNN in spaceborne GNSS-R is presented, with the technical solution and implementation process as attached. Figure 1 As shown, it includes the following steps:

[0039] Step S1: Acquire spaceborne GNSS-R observation data, ERA5 wind speed, ERA5 surge height, ERA5 wave direction data, and IMERG precipitation data;

[0040] Step S2: Preprocess and perform spatiotemporal matching on all acquired datasets;

[0041] Step S3: Data quality control and data partitioning;

[0042] Step S4: GloWS-Net model construction and training. The constructed multimodal deep learning GloWS-Net model is shown in the attached figure. Figure 2 As shown, the GloWS-Net model has three input lines: the first for the Normalized Bistatic Radar Cross Section (NBRCS) DDM, the second for the effective scattering area, and the third for auxiliary parameters. Specifically, the third line inputs auxiliary variables including the Normalized Bistatic Radar Cross Section (NBRCS), Leading Edge Slope (LES), Trailing Edge Slope (TES), Signal-to-Noise Ratio (SNR), Receiver Antenna Gain (sp_rx_gain), Latitude and Longitude of the specular reflection point, Incident Angle, Surge Height, Wave Direction, and Rainfall Intensity.

[0043] Step S5: Input the test dataset into the trained GloWS-Net model to obtain the inverted wind speed value and evaluate the result.

[0044] In step S1, the spaceborne GNSS-R observation data includes L1b BRCS DDM, effective scattering area, normalized bistatic radar cross section (NBRCS), leading edge slope (LES), trailing edge slope (TES), signal-to-noise ratio (SNR), receiver antenna gain (sp_rx_gain), latitude and longitude of the specular reflection point, and angle of incidence; the ERA5 wind speed is the combined wind speed of the U and V components 10m above the sea surface; the ERA5 swell height is the SWH (significant height of total swell); the ERA5 wave direction data is the mean wave direction provided in the ERA5 product; and the IMERG rainfall data is the IMERG-F rainfall intensity.

[0045] Step S2 includes the following sub-steps,

[0046] CYGNSS GNSS-R data is downsampled;

[0047] The CYGNSS GNSS-R downsampled data was spatiotemporally matched with ECMWF ERA5 wind speed, ERA5 wave direction, ERA5 swell SWH and IMERG precipitation data to obtain the matched dataset.

[0048] The post-matching dataset quality control described in step S3 includes deleting all observations containing NaN values; discarding all observations less than 0; ensuring the RCG value is greater than 3; discarding observations if the gain (sp_rx_gain) of the receiving antenna in the direction of the reflection point is less than 0 dBi; and discarding observations if the uncertainty of BRCS (ddm_BRCS_uncert) is greater than 1. Data filtering in CYGNSS L1B data uses the Basic Quality Control (QC) flag. Data filtering is applied using a single QC flag bit, rather than the entire QC flag bit (“Overall Quality Poor”, least significant bit or bit 0). Specifically, invalid or abnormal DDM data (bits 4, 7, 8, 9, 10, 15, 17-20) is deleted; data from other instruments with data transmission and calibration problems are deleted (bits 1, 5, 6, 13, 14, 21-26); data with uncertain transmitter power or antenna gain height must be deleted (bits 16 and 27); spacecraft data with large attitude errors or abnormal attitudes are deleted (bits 2, 3, and 28); L1B observations containing GPS IIF satellites are deleted (“low_quality_GPS_ant_knowledge”, bit 27); the quality control flags “sp_near_land” (bit 12) and “sp_very_near_land” (bit 11) are used to delete data collected when signals reflect on land; data with specular reflection points more than 25 km from land are selected to reduce modeling errors. The filtered dataset described in step S3 is randomly divided into a training set, a validation set, and a test set, accounting for 30%, 15%, and 55% of the filtered dataset, respectively.

[0049] The training process of the GloWS-Net global sea surface wind speed inversion model described in step S4 specifically includes:

[0050] After preprocessing all the acquired datasets, the filtered datasets are divided into training set, validation set and test set;

[0051] The filtered dataset was normalized to zero mean and unit variance based on features. A validation set was used to avoid overfitting, employing early stopping with six epochs. The learning rate was set to 0.0005. To further prevent overfitting, dropout was set to 0.1.

[0052] The rectified linear unit (ReLU) is used as the activation function (as shown in Equation (1)), the optimizer is Adam, and the mean squared error (MSE) is used as the loss function (as shown in Equation (2)).

[0053]

[0054] In the formula, x is the input value of the previous layer of neurons.

[0055]

[0056] In the formula, N is the number of samples. To predict wind speed, u i For reference wind speed.

[0057] The training set is used to train the network model, and the validation set is used to supervise the model training. The batch size is set to 32, meaning that each time the network updates its weight coefficients, it uses only samples from one batch to train the model. To ensure performance quality, the number of epochs is set to 100.

[0058] After training is complete, the test dataset is input into the trained GloWS-Net model to obtain the retrieved wind speed values, and the results are evaluated. Root mean square error (RMSE), mean absolute error (MAE), correlation coefficient (CC), and mean absolute percentage error (MAPE) are used as metrics to evaluate the model's retrieval performance, and their calculation formulas are as follows:

[0059]

[0060]

[0061]

[0062]

[0063] In the above formula, m is the sample size, and y i,ws and y i,T These are the model-inverted wind speed and the ERA5 reference wind speed, respectively. and These are the inverted average wind speed and the reference average wind speed, respectively.

[0064] The trained GloWS-Net model was used to retrieve global sea surface wind speed. The experimental results of this invention were compared with those of empirical models and some current network models (such as FCN and CNN models). The final results are shown in Table 2, which presents the statistical analysis of the retrieval accuracy of different methods for wind speeds less than 10 m / s, wind speeds between 10 and 15 m / s, wind speeds greater than 15 m / s, and wind speeds between 0 and 30 m / s. The following conclusions can be drawn from the table:

[0065] (1) In terms of accuracy obtained from the wind speed range of 0-30m / s, although the GloWS-Net and FCN models are comparable in terms of RMSE, the GloWS-Net model method proposed in this embodiment outperforms the FCN model in terms of MAE and MAPE, especially in terms of MAPE (16.72%).

[0066] (2) In terms of the four metrics (MRSE, MAE, CC, and MAPE), the GloWS-Net model method proposed in this embodiment significantly outperforms the MVE and CNN models. Compared with the MVE method, the accuracy is improved by 10.31%, 14.37%, 5.88%, and 16.19%, respectively; compared with the CNN model method, the accuracy is improved by 21.25%, 23.58%, 5.88%, and 20.09%, respectively.

[0067] Table 2 shows the accuracy of different models in retrieving wind speed on the test dataset.

[0068]

[0069] Furthermore, to compare the significant improvement in correlation between the wind speed inverted by the GloWS-Net model proposed in this embodiment and other inversion methods and ERA5 wind speed products, as shown in the appendix... Figure 3 As shown, the GloWS-Net model proposed in this embodiment exhibits a better correlation between the inverted wind speed and the ERA5 wind speed, outperforming existing MVE, FCN, and CNN model methods. Clearly, the GloWS-Net model inverts more data points symmetrically clustered along the y=x line, and fewer data points scattered around this line. The CNN model shows the worst correlation between the inverted wind speed and the ERA5 wind speed, clearly indicating that more auxiliary parameters are needed in the CNN model to obtain better results. However, the GloWS-Net model proposed in this embodiment integrates the DDM and effective scattering area features extracted by the CNN, along with more important auxiliary parameters, further enhancing the model's robustness, stability, and versatility.

[0070] The GloWS-Net model proposed in this embodiment achieves high accuracy in retrieving global sea surface wind speeds, which is largely consistent with the global wind speed distribution in ERA5. (See attached...) Figure 4As shown in the histogram of the deviation distribution between the wind speed inverted by the GloWS-Net model and the ERA5 wind speed (the figure shows the mean deviation (μ), standard deviation (σ), mean absolute error (MAE), and 80th percentile (Qua) ​​of the deviation; the blue bars represent the error distribution, the red dashed line represents the probability density function fitting curve of the error, and the green dashed line represents the wind speed deviation of 0 m / s), it can be seen that the deviation between the wind speed inverted by the GloWS-Net model and the ERA5 wind speed is very concentrated (80% of the wind speed deviation is less than 2.39 m / s) and is near the wind speed deviation of 0 m / s, indicating that the multimodal deep learning method proposed in this invention has significant advantages in inverting global sea surface wind speed.

[0071] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-described technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A multimodal deep learning method for sea surface wind speed inversion based on CNN in spaceborne GNSS-R, characterized in that, Includes the following steps: Step S1: Acquire spaceborne GNSS-R observation data, ERA5 wind speed, ERA5 surge height, ERA5 wave direction data, and IMERG precipitation data; Step S2: Preprocess and perform spatiotemporal matching on all acquired datasets; Step S3: Data quality control and data partitioning; Step S4: Model construction and training. The model has three input lines: the first is used to input the BRCS DDM, the second is used to input the effective scattering area, and the third is used to input auxiliary variables such as normalized bistatic radar cross section, leading edge slope, trailing edge slope, signal-to-noise ratio, receiver antenna gain, latitude and longitude of the specular reflection point, incident angle, surge height, wave direction, and rainfall intensity. CNN is used to extract effective features from the BRCS DDM and the corresponding effective scattering area, and the GNSS-R normalized bistatic radar cross section, leading edge slope, trailing edge slope, signal-to-noise ratio, receiver antenna gain, latitude and longitude of the specular reflection point, incident angle, surge height, wave direction, and rainfall intensity are fused to construct an end-to-end multimodal deep learning wind speed inversion model. Step S5: Input the test dataset into the trained model to obtain the inverted wind speed value and evaluate the result.

2. The multimodal deep learning method for sea surface wind speed inversion based on CNN in spaceborne GNSS-R according to claim 1, characterized in that, The spaceborne GNSS-R observation data includes L1b BRCS DDM, effective scattering area, normalized bistatic radar cross section, leading edge slope, trailing edge slope, signal-to-noise ratio, receiver antenna gain, latitude and longitude of the specular reflection point, and angle of incidence. The ERA5 wind speed is the combined wind speed of the U and V components 10 m above the sea surface. The ERA5 surge height is the SWH of the surge. The ERA5 wave direction data is the average wave direction provided in the ERA5 product; The IMERG rainfall data refers to the IMERG-F rainfall intensity.

3. The multimodal deep learning method for sea surface wind speed inversion based on CNN for spaceborne GNSS-R according to claim 1, characterized in that, Step S2 includes the following sub-steps: Step S2.1: Downsample the CYGNSS GNSS-R data; Step S2.2: Spatiotemporal matching of CYGNSS GNSS-R downsampled data with ECMWF ERA5 wind speed, ERA5 wave direction, ERA5 swell SWH and IMERG precipitation data to obtain the matched dataset.

4. The multimodal deep learning method for sea surface wind speed inversion based on CNN for spaceborne GNSS-R according to claim 3, characterized in that: In step S3, data quality control is performed on the matched dataset obtained in step S2.

2. The quality control of the matched dataset includes deleting all observations containing NaN values; discarding all observations less than 0; ensuring that the RCG value is greater than 3; discarding observations if the gain of the receiving antenna in the direction of the reflection point is less than 0 dBi; and discarding observations if the uncertainty of the BRCS is greater than 1. Data filtering uses basic quality control (QC) markers in CYGNSS L1B data; Instead of using the entire QC flag, data filtering is applied using a single QC flag. This involves deleting invalid or abnormal data from the DDM; deleting data from other instruments with data transmission and calibration issues; deleting data with uncertain transmitter power or antenna gain; deleting spacecraft data with large attitude errors or abnormal attitudes; deleting L1B observations containing GPS IIF satellites; using the quality control flags "sp_near_land" and "sp_very_near_land" to delete data collected when signals reflect off land; and selecting data where specular reflection points are more than 25 km from land to reduce modeling errors. The filtered dataset is randomly divided into training, validation, and test sets, accounting for 30%, 15%, and 55% of the filtered dataset, respectively.

5. The multimodal deep learning method for sea surface wind speed inversion based on CNN for spaceborne GNSS-R according to claim 1, characterized in that, The model training process in step 4 is implemented as follows: After preprocessing all the acquired datasets, the filtered datasets are divided into training set, validation set and test set; After filtering, the dataset is normalized to zero mean and unit variance according to features. The validation set is used to avoid overfitting. An early stopping condition with six epochs is adopted, and the learning rate is set to 0.0005. The random deactivation is set to 0.

1. The rectified linear unit is used as the activation function, the Adam optimizer is selected, and the mean squared error is used as the loss function. The training set is used to train the network model, the validation set is used to supervise the model training, the batch training size is set to 32, and when the network updates the weight coefficients each time, only samples from one batch are used to train the model; the number of epochs is set to 100.

6. The multimodal deep learning method for sea surface wind speed inversion based on CNN for spaceborne GNSS-R according to claim 5, characterized in that, The activation function is ； In the formula, x is the input value of the previous layer of neurons; The loss function is ； In the formula, N is the number of samples. To predict wind speed, For reference wind speed.

7. The multimodal deep learning method for sea surface wind speed inversion based on CNN in spaceborne GNSS-R according to claim 1, characterized in that, In step 5, the root mean square error (RMSE), mean absolute error (MAE), correlation coefficient (CC), and mean absolute percentage error (MAPE) are used as indicators to evaluate the model inversion performance. The calculation formulas are as follows: ； ； ； ； In the above formula, m is the sample size. and These are the model-inverted wind speed and the ERA5 reference wind speed, respectively. and These are the inverted average wind speed and the reference average wind speed, respectively.

Images