Multi-channel multi-mode space-time 3D convolution and Transform fused satellite-borne GNSS-R wave height inversion model

Abstract

The invention relates to a multi-channel multi-mode space-time 3D convolution and Transform fused satellite-borne GNSS-R wave height inversion model, and aims to solve the problems that a traditional empirical model is poor in adaptability to complex oceans and existing deep learning cannot fully excavate space-time correlation. According to the method, three modules are innovatively fused for cooperative processing: firstly, a 3D CNN-ConvLSTM is utilized to extract local spatial-temporal characteristics of multichannel GNSS-R data and capture a dynamic time sequence; secondly, performing deep coding on the sea surface environment parameters through a multi-head attention mechanism of Transform, and establishing global dependence between features; and finally, realizing cross-modal fusion by adopting a weighted summation and feature splicing strategy. And through training of an Adam optimizer and control of an early stop strategy, optimization is carried out by taking a mean square error as a loss function. Experiments show that compared with a traditional model and a machine learning model, the method has the advantages that the significant wave height estimation error is reduced by 40-53%, the correlation coefficient reaches 0.84-0.91, the precision and generalization ability under the complex sea condition are remarkably improved, and reliable support is provided for ocean remote sensing monitoring and disaster early warning.

Description

Technical Field

[0001] This invention relates to the fields of marine remote sensing and artificial intelligence technology, and in particular to a spaceborne GNSS-R wave height inversion model that fuses multi-channel, multi-modal spatiotemporal 3D convolution with Transformer. Background Technology

[0002] Significant wave height, as a core parameter reflecting sea surface fluctuations, is of great significance for accurate estimation in marine scientific research and human maritime activities. In oceanographic research, it is a crucial basis for exploring the mechanisms of wave generation, propagation, and evolution. In meteorological forecasting, significant wave height is closely related to weather systems such as sea breezes and storm surges, directly affecting the prediction accuracy of meteorological models. Furthermore, in the field of maritime operational safety, significant wave height data is crucial for ship navigation safety, offshore platform stability, and offshore engineering construction. In short, the accuracy of significant wave height estimation directly impacts multiple important areas, including marine resource development, disaster prevention and mitigation, and global climate change research.

[0003] In the field of spaceborne global navigation satellite system reflectometry (GNSS-R) technology, accurate estimation of significant wave height is crucial for oceanographic research, weather forecasting, and maritime operational safety. However, as research progresses, traditional methods for estimating significant wave height have gradually revealed their limitations, failing to meet the ever-increasing demand for high precision.

[0004] Spaceborne GNSS-R data possesses multi-channel characteristics, with different channels containing diverse information about sea surface conditions. For example, channels with different frequencies and polarizations can reflect sea surface roughness and reflection characteristics from different perspectives. However, many previous methods have failed to fully exploit the potential of multi-channel data, utilizing only single or a few channels. This has resulted in a large amount of useful information being ignored, making it difficult to comprehensively and accurately describe sea surface conditions and consequently affecting the accuracy of significant wave height estimation.

[0005] Early methods for estimating significant wave height based on traditional algorithms relied heavily on simple physical models and empirical formulas, which proved poorly adaptable to the complex and ever-changing marine environment. When faced with the complex scattering characteristics of the sea surface and GNSS-R signals affected by various factors, these methods failed to accurately extract features related to significant wave height, resulting in large estimation errors.

[0006] In recent years, deep learning methods have been introduced into this field, and Convolutional Neural Networks (CNNs) have been applied to process GNSS-R data due to their advantages in image feature extraction. However, ordinary CNNs have shortcomings in processing multi-channel, multi-modal spatiotemporal data. Their convolution operations can usually only capture local spatiotemporal features, making it difficult to establish long-range dependencies. They do not fully grasp the complex spatiotemporal variation information contained in spaceborne GNSS-R data, resulting in limitations in the model's understanding of the data. In summary, although deep learning methods are widely used, they have not fully integrated multi-channel, multi-modal spatiotemporal information and the complex relationships between features.

[0007] The proposed spaceborne GNSS-R wave height inversion model, which integrates multi-channel, multi-modal spatiotemporal 3D convolution with Transformer, offers significant advantages. The 3D CNN module enables deep convolution of multi-channel data in both temporal and spatial dimensions, fully leveraging the features of each channel at different spatiotemporal locations. Compared to methods utilizing only single-channel data, this greatly improves the efficiency of utilizing multi-source information and comprehensively captures sea surface state information. The image processing module within the Transformer module globally models the image features corresponding to the multi-channel data, breaking the limitations of traditional methods on local features and capturing long-distance dependencies, effectively integrating long-distance spatiotemporal correlations in multi-channel data. Its feature processing module further optimizes the extracted features, enhancing their expressive power and enabling the model to better understand the complex information behind the multi-channel data. The Convolutional Long Short-Term Memory (ConvLSTM) module utilizes its advantages in processing time-series data, combining the features extracted by 3D CNN and Transformer to further analyze the temporal trends of the data, thereby more accurately estimating the significant wave height. Compared with other methods, the model of this invention, through the collaborative work of multiple modules, fully leverages the advantages of multi-channel and multi-modal spatiotemporal data, significantly improving the accuracy and stability of effective wave height estimation, and providing more reliable data support for ocean-related research and applications.

[0008] In view of the above, this invention proposes a spaceborne GNSS-R wave height inversion model that integrates multi-channel, multi-modal spatiotemporal 3D convolution with Transformer. Summary of the Invention

[0009] To address the shortcomings of traditional effective wave height estimation methods, such as poor adaptability to complex marine environments and difficulty in meeting high-precision requirements, as well as the limitations of existing deep learning methods in processing multi-channel and multi-modal spatiotemporal data of spaceborne GNSS-R, such as the inability to fully exploit the potential of multiple channels, the inability to capture only local spatiotemporal features, and the difficulty in establishing long-distance dependencies and complex relationships between features, this invention proposes a spaceborne GNSS-R wave height inversion model that integrates multi-channel and multi-modal spatiotemporal 3D convolution with Transformer.

[0010] To achieve the above-mentioned technical objectives and effects, the present invention provides the following technical solution:

[0011] A spaceborne GNSS-R wave height inversion model fused with multi-channel, multi-modal spatiotemporal 3D convolution and Transformer, characterized by the following steps:

[0012] Step S1: Preprocess the spaceborne GNSS-R multi-channel data to unify the feature dimensions of data from different channels;

[0013] Step S2: Extract spatiotemporal features of multi-channel data using 3D convolution and ConvLSTM2D;

[0014] Step S3: Use the Transformer encoder to perform deep encoding on the multimodal feature data;

[0015] Step S4: Fuse the image branching processing result from step S2 with the feature input processing result from step S3;

[0016] Step S5: Construct a fully connected network based on the fused feature representation to predict the effective wave height;

[0017] Step S6: Use the trained model to estimate the effective wave height.

[0018] Furthermore, the preprocessing in step S1 above includes:

[0019] Data standardization, normalization, and dimensional adjustment;

[0020] Extract onboard GNSS-R variable data, including bistatic scattering cross section (BRCS), effective scattering area, raw count, power delay Doppler map, peak-to-average delay Doppler map (DDMA), leading edge slope (LES), trailing edge slope (TES), sum of leading edge waveforms (LEWS), sum of trailing edge waveforms (TEWS), equivalent isotropic radiated power (gps_eirp), receiver antenna gain (sp_rx_gain), incident angle (sp_inc_angle), latitude and longitude of reflection point (sp_lat, sp_lon), and range correction gain (RCG);

[0021] Extract environmental variables from the sea surface, including wind speed, wind direction, water depth, and rainfall.

[0022] Furthermore, the 3D convolutional structure in step S2 above includes a combination of multiple convolutional layers, pooling layers, and ConvLSTM2D layers, used to extract spatiotemporal features from the following image data:

[0023] Bistatic scattering cross section (BRCS), effective scattering area, raw counts, and power-time delay Doppler plot.

[0024] Furthermore, the Transformer encoder in the appeal step S3 includes: a multi-head attention mechanism module, a feedforward neural network module, and a layer normalization module.

[0025] Furthermore, the feature fusion strategy in step S4 above includes operations such as weighted summation, feature concatenation, and element-wise multiplication.

[0026] Furthermore, the model training in step S5 above employs: the Adam optimization algorithm; early stopping to prevent overfitting; and mean squared error as the damage function.

[0027] A spaceborne GNSS-R effective wave height estimation model system is used to execute the aforementioned spaceborne GNSS-R wave height inversion model fused with multi-channel, multi-modal spatiotemporal 3D convolution and Transformer, including:

[0028] The data preprocessing module is used to perform the above step S1: preprocessing the spaceborne GNSS-R multichannel data;

[0029] The image branch processing module is used to perform the above step S2: extracting spatiotemporal features of multi-channel data using 3D convolution and ConvLSTM2D;

[0030] The feature input processing module is used to perform the above step S3: using a Transformer encoder to perform deep encoding on the multimodal feature data;

[0031] The feature fusion module is used to perform the above step S4: fusing the image branch processing result with the feature input processing result;

[0032] The prediction module is used to perform step S5 above: constructing a fully connected network based on the fused feature representation to predict the effective wave height;

[0033] The training module is used to train the model parameters;

[0034] The estimation module is used to perform step S6 above: estimating the effective wave height using the trained model.

[0035] Furthermore, the aforementioned data preprocessing module includes: a standardization unit, a normalization unit, and a dimension adjustment unit.

[0036] Furthermore, the aforementioned image branch processing module includes: a 3D convolutional neural network submodule and a ConvLSTM2D submodule.

[0037] Furthermore, the aforementioned feature input processing module includes: a Transformer encoder submodule, a multi-head attention mechanism submodule, and a feedforward neural network submodule.

[0038] Furthermore, the aforementioned feature fusion module includes: a weighted summation unit, a feature splicing unit, and an element-wise multiplication unit.

[0039] Furthermore, the optimization algorithms used in the above training module include the Adam optimization algorithm, early stopping, and mean square error method.

[0040] The training module, also known as the effective wave height estimation model construction module, constructs a spaceborne GNSS-R effective wave height estimation model based on 3D CNN-Trans-ConvLSTM. This model includes a 3D convolutional neural network (3D CNN) module, a Transformer module, and a convolutional long short-term memory network (ConvolutionalLSTM, ConvLSTM) module. The Transformer module includes modules for image processing and feature processing.

[0041] The estimation module, namely the significant wave height estimation, takes the preprocessed GNSS-R variable data and sea surface environmental variable parameters, such as wind speed, wind direction, water depth, and rainfall, and inputs them into the constructed significant wave height estimation model. Through the model's feature extraction, analysis, and calculation process, the significant wave height is accurately estimated.

[0042] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention discloses a spaceborne GNSS-R wave height inversion model that fuses multi-channel, multi-modal spatiotemporal 3D convolution with Transformer. Compared with traditional methods and existing deep learning methods, this invention, through the fusion of multi-channel, multi-modal spatiotemporal 3D convolution and Transformer networks, more comprehensively utilizes the spatiotemporal information and relationships between features in the data, improving the accuracy and reliability of effective wave height estimation, and exhibiting good generalization ability under different sea states. Attached Figure Description

[0043] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the specific embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0044] Figure 1This is a flowchart of the spaceborne GNSS-R wave height inversion model fused with multi-channel multimodal spatiotemporal 3D convolution and Transformer provided in this embodiment of the invention;

[0045] Figure 2 This is a structural diagram of the effective wave height estimation model provided in the embodiments of the present invention;

[0046] Figure 3 This is a comparison between the effective wave height estimate of the model channel 1 proposed in this embodiment and the ERA5 data;

[0047] Figure 4 This is a comparison between the effective wave height estimate of channel 2 output by the proposed model in this embodiment and the ERA5 data;

[0048] Figure 5 This is a comparison between the effective wave height estimate of channel 3 of the model proposed in this embodiment and the ERA5 data;

[0049] Figure 6 This is a comparison between the effective wave height estimate of channel 4 of the model proposed in this embodiment and the ERA5 data. Detailed Implementation

[0050] 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.

[0051] This invention provides a spaceborne GNSS-R wave height inversion model that fuses multi-channel, multi-modal spatiotemporal 3D convolution with Transformer, which will be described in detail below with reference to the accompanying drawings:

[0052] Example 1

[0053] To verify the feasibility and reliability of the spaceborne GNSS-R wave height inversion model fused with multi-channel multimodal spatiotemporal 3D convolution and Transformer, CYGNSS GNSS-R L1 observation data were downloaded from the official website. Preprocessing was performed to extract the following: bistatic scattering cross section (BRCS), effective scattering area, raw counts, power-delay Doppler plot, peak-to-average delay Doppler plot (DDMA), leading edge slope (LES), trailing edge slope (TES), sum of leading edge waveforms (LEWS), sum of trailing edge waveforms (TEWS), equivalent isotropic radiated power (gps_eirp), receiver antenna gain (sp_rx_gain), incident angle (sp_inc_angle), latitude and longitude of the reflection point (sp_lat, sp_lon), and range correction gain (RCG). Download sea surface environmental variable parameters data from the official website: wind speed, wind direction, water depth, and rainfall. Specifically, download sea surface wind speed, wind direction, and water depth data from the European Centre for Medium-Range Weather Forecasts (ECMWF), and rainfall data from the IMERG website. A multi-channel, multi-modal spatiotemporal 3D convolutional and Transformer fusion spaceborne GNSS-R wave height inversion model is described in the attached technical solution implementation process. Figure 1 As shown, it includes the following steps:

[0054] Step S1: Preprocess the spaceborne GNSS-R multi-channel data to unify the feature dimensions of the data from different channels;

[0055] Step S2: Use 3D convolution and ConvLSTM2D to extract the spatiotemporal features of multi-channel data;

[0056] Step S3: Use the Transformer encoder to perform deep encoding on the multimodal feature data;

[0057] Step S4: Fuse the image branching processing result with the feature input processing result;

[0058] Step S5: Construct a fully connected network based on the fused feature representation to predict the effective wave height;

[0059] Step S6: Estimate the significant wave height value based on the significant wave height estimation model.

[0060] As one implementation of this embodiment, the preprocessing step S1 includes data standardization, normalization, and dimensionality adjustment. Furthermore, the data preprocessing also includes: downloading CYGNSS GNSS-R L1 observation data from the official website, and preprocessing to extract the following: bistatic scattering cross section (BRCS), effective scattering area, raw counts, power-time delay Doppler plot (power_analog), peak-to-average delay Doppler plot (DDMA), leading edge slope (LES), trailing edge slope (TES), leading edge waveform summation (LEWS), trailing edge waveform summation (TEWS), equivalent isotropic radiated power (gps_eirp), receiver antenna gain (sp_rx_gain), incident angle (sp_inc_angle), latitude and longitude of the reflection point (sp_lat, sp_lon), and range correction gain (RCG). Download marine environmental variable parameter data from the official website: wind speed, wind direction, water depth, and rainfall. Specifically, download sea surface wind speed, wind direction, and water depth data from the European Centre for Medium-Range Weather Forecasts (ECMWF), and download rainfall data from the IMERG website. Then, perform data feature extraction and data quality control preprocessing.

[0061] As one implementation of this embodiment, the 3D convolutional structure used in the image branching process in step S2 includes a combination of multiple convolutional layers, pooling layers, and ConvLSTM2D layers. The input image includes: bistatic scattering cross section (BRCS), effective scattering area, raw counts, and power-time Doppler image.

[0062] As one implementation of this embodiment, the Transformer encoder in step S3 includes a multi-head attention mechanism, a feedforward neural network, and a layer normalization module.

[0063] As one implementation method of this embodiment, the multi-branch feature fusion strategy in step S4 includes operations such as weighted summation, concatenation, and element-wise multiplication.

[0064] As one implementation of this embodiment, the optimization algorithms used in step S5 for model training include the Adam optimization algorithm and early stopping. The effective wave height estimation model includes: a 3D convolutional neural network (3D CNN) module, a Transformer module, and a convolutional long short-term memory network (Convolutional LSTM, ConvLSTM) module, wherein the Transformer module includes a module for image processing and a module for feature processing. (Appendix) Figure 2This invention demonstrates the framework of a spaceborne GNSS-R effective wave height estimation model based on 3DCNN-Trans-ConvLSTM.

[0065] The model mainly consists of an image feature extraction module based on a three-dimensional convolutional neural network (3D CNN), a dual-path feature processing module based on Transformer, a temporal feature analysis module based on a convolutional long short-term memory network (ConvLSTM), and a multi-source feature fusion output module. The 3D CNN module extracts spatiotemporal features from four types of image inputs—BRCS, effective scattering area, power delay Doppler image, and raw count—through multi-layer 3D convolution and pooling operations, capturing local spatiotemporal information of multi-channel data in 17×11×4 dimensions. A dual-path feature processing module based on Transformer processes image features and parametric features separately. The image processing module globally models the features output by the 3D CNN to capture long-distance dependencies, while the feature processing module optimizes parametric features such as GNSS-R variables and sea surface environmental variables to enhance feature representation. The ConvLSTM module performs temporal analysis on the image features processed by the 3D CNN to uncover dynamic trends in the data over time. After each module completes feature extraction and processing, the multi-source feature fusion output module first sums and fuses the features from the four image branches, then multiplies and fuses them with the parametric features. A multi-layer fully connected network enables deep feature interaction, and Dropout technology is applied to prevent overfitting, ultimately outputting a 4-channel effective wave height estimation result. The model uses mean squared error as the loss function and employs the Adam optimizer for parameter optimization to ensure the stability of model training and estimation accuracy.

[0066] As one implementation method of this embodiment, the effective wave height estimation in step S6 involves inputting the preprocessed GNSS-R variable data and sea surface environmental variable parameter data, such as wind speed, wind direction, water depth, and rainfall, into the constructed effective wave height estimation model. Through the feature extraction, analysis, and calculation process of the model, the accurate estimation of the effective wave height is completed.

[0067] Appendix Figure 3 -Appendix Figure 6The results are as follows: the effective wave height estimates of channel 1-channel 4 of the model proposed in this embodiment are compared with the ERA5 data. The experimental results show that: Compared with ERA5 data, the root mean square error (RMSE), bias, correlation coefficient, and mean absolute percentage error of the estimated significant wave height from channel 1 data are 0.410m, 0.003m, 0.83, and 17.13%, respectively; the RMSE, bias, correlation coefficient, and mean absolute percentage error of the estimated significant wave height from channel 2 data are 0.449m, 0.002m, 0.79, and 19.50%, respectively; the RMSE, bias, correlation coefficient, and mean absolute percentage error of the estimated significant wave height from channel 3 data are 0.451m, 0.0009m, 0.78, and 19.29%, respectively; and the RMSE, bias, correlation coefficient, and mean absolute percentage error of the estimated significant wave height from channel 4 data are 0.452m, 0.002m, 0.78, and 19.57%, respectively. Therefore:

[0068] (1) Overall performance

[0069] Root Mean Square Error (RMSE): Channel 1 has the lowest RMSE at 0.410m, while the RMSEs of other channels are slightly higher, but all are around 0.45m. This indicates that the model's estimate of the significant wave height differs little from the ERA5 data, and the overall accuracy is high.

[0070] Bias: The biases of all channels are very small, with absolute values ​​not exceeding 0.003m, indicating that there is almost no systematic bias between the model estimates and the ERA5 data, and the model estimates are unbiased.

[0071] Correlation coefficient (R): Channel 1 has the highest correlation coefficient at 0.83, while channels 2 and 3 both have correlation coefficients of around 0.78, and channel 4 has a correlation coefficient of 0.78. This indicates a strong linear correlation between the model estimates and the ERA5 data, and the model is able to capture the trend of significant wave height changes well.

[0072] Mean Absolute Percentage Error (MAPE): Channel 1 has the lowest MAPE at 17.13%, while the MAPE for other channels is around 19%. This indicates that the relative error between the model estimates and the ERA5 data is small, and the model's estimation accuracy is high.

[0073] (2) Comparative Analysis

[0074] Compared with other channels: Channel 1 performed best in all four indicators of RMSE, bias, correlation coefficient and MAPE, indicating that the data of Channel 1 may be the most sensitive to the estimation of significant wave height and contains richer information on significant wave height.

[0075] Comparison of Channels 2-4: The metrics for Channels 2-4 are relatively similar, with RMSE between 0.449 and 0.452 m, deviation between 0.0009 and 0.002 m, correlation coefficients between 0.78 and 0.79, and MAPE between 19.29% and 19.57%. This indicates that these three channels perform similarly in estimating significant wave height and all provide relatively reliable results.

[0076] Overall, the proposed model demonstrates high accuracy and stability in estimation results across all channels, effectively utilizing multi-channel data for significant wave height estimation. This verifies the effectiveness and reliability of the proposed model in significant wave height estimation.

[0077] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0078] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A spaceborne GNSS-R wave height inversion model fused with multi-channel, multi-modal spatiotemporal 3D convolution and Transformer, characterized in that, Includes the following steps: Step S1: Preprocess the spaceborne GNSS-R multi-channel data to unify the feature dimensions of data from different channels; Step S2: Extract spatiotemporal features of multi-channel data using 3D convolution and ConvLSTM2D; Step S3: Use the Transformer encoder to perform deep encoding on the multimodal feature data; Step S4: Fuse the image branching processing result from step S2 with the feature input processing result from step S3; Step S5: Construct a fully connected network based on the fused feature representation to predict the effective wave height; Step S6: Use the trained model to estimate the effective wave height.

2. The spaceborne GNSS-R wave height inversion model based on the fusion of multi-channel, multi-modal spatiotemporal 3D convolution and Transformer as described in claim 1, is characterized in that... The preprocessing in step S1 includes: Data standardization, normalization, and dimensional adjustment; Extract onboard GNSS-R variable data, including bistatic scattering cross section (BRCS), effective scattering area, raw count, power delay Doppler map, peak-to-average delay Doppler map (DDMA), leading edge slope (LES), trailing edge slope (TES), sum of leading edge waveforms (LEWS), sum of trailing edge waveforms (TEWS), equivalent isotropic radiated power (gps_eirp), receiver antenna gain (sp_rx_gain), incident angle (sp_inc_angle), latitude and longitude of reflection point (sp_lat, sp_lon), and range correction gain (RCG); Extract environmental variables from the sea surface, including wind speed, wind direction, water depth, and rainfall.

3. The spaceborne GNSS-R wave height inversion model based on the fusion of multi-channel, multi-modal spatiotemporal 3D convolution and Transformer as described in claim 1, is characterized in that... The 3D convolutional structure in step S2 includes a combination of multiple convolutional layers, pooling layers, and ConvLSTM2D layers, used to extract spatiotemporal features from the following image data: Bistatic scattering cross section (BRCS), effective scattering area, raw counts, and power-delay Doppler plot. power_ana log).

4. The spaceborne GNSS-R wave height inversion model based on the fusion of multi-channel, multi-modal spatiotemporal 3D convolution and Transformer as described in claim 1, characterized in that, The Transformer encoder in step S3 includes: a multi-head attention mechanism module, a feedforward neural network module, and a layer normalization module.

5. The spaceborne GNSS-R wave height inversion model based on the fusion of multi-channel, multi-modal spatiotemporal 3D convolution and Transformer as described in claim 1, characterized in that, The feature fusion strategy in step S4 includes operations such as weighted summation, feature concatenation, and element-wise multiplication.

6. The spaceborne GNSS-R wave height inversion model based on the fusion of multi-channel, multi-modal spatiotemporal 3D convolution and Transformer as described in claim 1, characterized in that, The model training in step S5 uses: Adam optimization algorithm; Early stopping prevents overfitting; Mean square error is used as the damage function.

7. A spaceborne GNSS-R effective wave height estimation model system, used to execute the spaceborne GNSS-R wave height inversion model fused with multi-channel multimodal spatiotemporal 3D convolution and Transformer as described in claims 1-6, characterized in that, include: Data preprocessing module: used to perform step S1 in claim 1; Image branch processing module: used to perform step S2 in claim 1; Feature input processing module: used to execute step S3 in claim 1; Feature fusion module: used to perform step S4 in claim 1; Prediction module: used to perform step S5 in claim 1; Training module: Used to train model parameters; Estimation module: used to perform step S6 in claim 1.

8. The spaceborne GNSS-R effective wave height estimation model system according to claim 7, characterized in that, The data preprocessing module includes: Standardized unit; Normalized unit; Dimension adjustment unit.

9. The spaceborne GNSS-R effective wave height estimation model system according to claim 7, characterized in that, The image branch processing module includes a 3D convolutional neural network submodule and a ConvLSTM2D submodule.

10. The spaceborne GNSS-R effective wave height estimation model system according to claim 7, characterized in that, The feature input processing module includes: Transformer encoder submodule; Multi-head attention mechanism submodule; Feedforward neural network submodule.

Images