A method and system for constructing a satellite-borne GNSS-R ocean significant wave height inversion model under high sea conditions

Abstract

The present application belongs to the technical field of multi-constellation spaceborne GNSS-R data ocean significant wave height inversion, and discloses a spaceborne GNSS-R ocean significant wave height inversion model construction method and system under high sea conditions, obtains modeling data, auxiliary data and verification data; the characteristic variable parameters and auxiliary variable parameters of the modeling data are extracted, and the extracted variable parameters are subjected to space-time matching; the variable parameters subjected to space-time matching are subjected to data quality control and data set division; the WaveMambaFormer model is constructed and trained by using the divided data set, and the SHAP interpreter is used to calculate the feature importance and global explanation, thereby enhancing the explainability of the WaveMambaFormer model; and the ERA5, WaveWatch III and Jason-3 significant wave height data are used as reference data to evaluate the inversion performance of the WaveMambaFormer model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ocean significant wave height inversion from multi-constellation spaceborne GNSS-R data, and particularly relates to a method and system for constructing a spaceborne GNSS-R ocean significant wave height inversion model under high sea conditions. BACKGROUND

[0002] The significant wave height (SWH) is one of the important indicators used by oceanographers to describe the sea state, and can be used to assess the wave energy of the marine environment and the safety of ship navigation. Thanks to the continuous development and progress of science and technology, a new satellite remote sensing technology is emerging as an effective method for estimating SWH, namely Global Navigation Satellite System Reflectometry (GNSS-R). Compared with traditional in-situ measurement methods such as buoys or ships, which are limited by coverage, spaceborne GNSS-R has the significant advantages of wide observation range, low observation cost, and the ability to measure all-weather and all-day, making it a promising method for measuring SWH.

[0003] Currently, the methods for estimating SWH using spaceborne GNSS-R technology mainly include SWH inversion empirical models based on spaceborne GNSS-R observation values and machine learning or deep learning methods. However, the empirical model method only considers limited input features, and due to the influence of various complex factors on the sea surface, it is difficult to represent the relationship between SWH and GNSS-R characteristic values with a simple empirical model. Machine learning or deep learning can handle non-linear problems and has significant advantages and potential in SWH inversion. However, traditional machine learning has limitations in advanced feature extraction. Deep learning methods based on deep neural networks can handle complex data, and some deep learning methods have been applied to SWH inversion, and the inversion accuracy of SWH has been greatly improved. However, in fact, the SWH inversion accuracy is still not very high, and the range of inverted wave height is small. The reason is that SWH is affected by many complex factors on the sea surface, and to obtain higher SWH inversion accuracy, it is necessary to continuously improve the inversion model and consider more auxiliary data that affect SWH. In addition, for deep learning models, their internal structure is usually complex and difficult to interpret. Quantifying the importance of each feature to model prediction can help identify the most important and influential features, thereby optimizing feature selection. Therefore, the interpretability of the model is also very important.

[0004] In summary, since SWH is affected by various complex factors on the sea surface in actual situations, in order to obtain higher inversion accuracy and greater range of wave height inversion accuracy, a new SWH deep learning inversion model needs to be constructed, and more SWH influencing factors need to be considered. In addition, the explanation of the importance of model features is also crucial, SHAP (Shapley Additive exPlanations) can analyze the influence of each feature on the prediction result, quantify the importance of each feature to the model prediction, and at the same time, by averaging the SHAP values of multiple samples, the overall influence of the feature on the model prediction can be understood, and the global feature importance analysis is generated, which enhances the explainability of the deep learning model. SUMMARY

[0005] In order to overcome the shortcomings of the existing satellite-borne GNSS-R ocean significant wave height inversion model, such as low inversion accuracy and low inversion efficiency, and the current deep learning inversion model based on satellite-borne GNSS-R technology has weak explainability and insufficient feature explanation, the present application provides a satellite-borne GNSS-R ocean significant wave height inversion model construction method and system under high sea conditions, based on satellite-borne GNSS-R technology, combined with SHAP interpreter and fusion of CNN, Transformer and SSM, a hybrid network (called WaveMambaFormer model) is constructed for estimating ocean significant wave height, which can well enhance the explainability of the model and improve the efficiency and inversion accuracy of the model.

[0006] The present application provides a satellite-borne GNSS-R ocean significant wave height inversion model construction method under high sea conditions, the method comprises:

[0007] Step S1, obtaining modeling data, auxiliary data and verification data, wherein the modeling data comprises satellite-borne GNSS-R data and ERA5 significant wave height data; the auxiliary data comprises ERA5 data, SMAP data and OSCAR data; the verification data comprises ERA5, WaveWatch III and Jason-3 significant wave height data;

[0008] Step S2, extracting satellite-borne GNSS-R feature variable parameters and auxiliary variable parameters, and performing spatio-temporal matching on the extracted variable parameters;

[0009] Step S3, performing data quality control and data set division on the spatio-temporally matched variable parameters;

[0010] Step S4, constructing and training the WaveMambaFormer model using the divided data set, and calculating the feature importance and global explanation using the SHAP interpreter to enhance the explainability of the WaveMambaFormer model;

[0011] Step S5, the inversion performance of the WaveMambaFormer model is evaluated respectively with ERA5, WaveWatch III and Jason-3 significant wave height data as reference data.

[0012] Preferably, in the step S2, the satellite-borne GNSS-R characteristic variable parameters and auxiliary variable parameters are extracted, and the extracted variable parameters are spatiotemporally matched, including:

[0013] The extracted space-borne GNSS-R variable parameters in step S2.1 include bistatic radar cross section BRCS, effective scattering area eff_scatter, DDM peak signal-to-noise ratio Ddm_peak_snr, DDM waveform leading edge slope Ddm_sp_les, DDM specular point normalized bistatic radar cross section Ddm_sp_nbrcs, normalized specular point signal-to-noise ratio Ddm_sp_normalized_snr, DDM specular point signal-to-noise ratio Ddm_sp_snr, DDM track ID Ddm_track_id, GNSS satellite block code used to generate the DDM waveform Gnss_block_flag, GNSS satellite PRN code used to generate the DDM waveform Gnss_prn_code, altitude of the low earth orbit satellite corresponding to the DDM sampling time Rx_alt, latitude of the low earth orbit satellite corresponding to the DDM sampling time Rx_lat, longitude of the low earth orbit satellite corresponding to the DDM sampling time Rx_lon, pitch angle of the low earth orbit satellite corresponding to the DDM sampling time Rx_pitch, roll angle of the low earth orbit satellite corresponding to the DDM sampling time Rx_roll, X component of the low earth orbit satellite velocity corresponding to the DDM sampling time Rx_vel_x, Y component of the low earth orbit satellite velocity corresponding to the DDM sampling time Rx_vel_y, Z component of the low earth orbit satellite velocity corresponding to the DDM sampling time Rx_vel_z, specular point receiver antenna gain Sp_antenna_gain, azimuth angle of the specular point in the antenna coordinate system Sp_az_antenna, azimuth angle of the specular point in the satellite body coordinate system Sp_az_body, azimuth angle of the specular point in the orbit coordinate system Sp_az_orbit, azimuth angle of the specular point in the pattern coordinate system Sp_az_pattern, GNSS signal incidence angle at the specular point Sp_inc_angle, altitude of the specular point Sp_alt, latitude of the specular point Sp_lat, longitude of the specular point Sp_lon, elevation angle of the specular point in the antenna coordinate system Sp_theta_antenna, elevation angle of the specular point in the satellite body coordinate system Sp_theta_body, elevation angle of the specular point in the orbit coordinate system Sp_theta_orbit, elevation angle of the specular point in the pattern coordinate system Sp_theta_pattern, X component of the specular point velocity Sp_vel_x, Y component of the specular point velocity Sp_vel_y, Z component of the specular point velocity Sp_vel_z, X component of the GNSS satellite velocity at the signal transmission time corresponding to the DDM sampling time Tx_vel_x, Y component of the GNSS satellite velocity at the signal transmission time corresponding to the DDM sampling time Tx_vel_y,The DDM extracts the Z component of the velocity of the GNSS satellite Tx_vel_z corresponding to the signal transmission time at the intermediate time point.

[0014] Step S2.2, after spatiotemporal matching of the characteristic variable parameters and metadata variables in the Tianmu No. 1 GPS-R / BDS-R / GLONASS-R / Galileo-R L1 level data and ERA5 significant wave height data, ERA5 wind speed, ERA5 rainfall, ERA5 sea surface temperature SST, SMAP sea surface salinity SSS, and OSCAR sea surface ocean current data, and ERA5 significant wave height data, WaveWatch III significant wave height data, and Jason-3 significant wave height data, a dataset after spatiotemporal matching is obtained.

[0015] Preferably, in step S3, the data quality control of the spatiotemporally matched variable parameters comprises:

[0016] The mirror reflection point on the sea is selected, i.e., Sp_surface_type = 0 is set; the observation value must be positive, and when the value is a Nan value, it needs to be discarded; when the satellite roll angle is greater than 30 degrees, the yaw angle is greater than 5 degrees, or the pitch angle is greater than 10 degrees in absolute value, the data will be discarded; the detected radio frequency interference RFI data is deleted; if the uncertainty of the estimated mirror point delay and Doppler shift exceeds the preset requirement, the data is deleted; if the height of the satellite exceeds the nominal height range, the data is deleted; if the temperature and automatic gain control value of the delay Doppler map DDM power calibration condition exceed the preset range, the data is deleted.

[0017] Preferably, in step S4, the WaveMambaFormer model is composed of a CNN-based DDM feature extraction module, a Transformer-based module, and a state space model State Space Model, SSM module.

[0018] First, the local spatial information of the signal in the bistatic radar scattering cross section BRCS is extracted through the multi-layer convolutional layer of CNN, and after the convolution operation, the global average pooling is used to reserve preset significant features and reduce the parameter amount, then the reserved significant features are input into the Transformer to further capture long-distance dependence and time sequence relationship; and the long-range dependence relationship between different time points is captured through the self-attention mechanism to extract the information in eff_scatter from the reflection signal; finally, the dynamic multi-path SSM module extracts features from the variable parameters.

[0019] Preferably, in step S4, training the WaveMambaFormer model comprises:

[0020] A log-cosh loss function is introduced as the loss function, and the calculation formula of the log-cosh loss is as follows:

[0021]

[0022] In the formula, y represents the predicted significant wave height of the sea surface, y i represents the reference significant wave height.

[0023] Preferably, in the step S4, the feature importance and global explanation are calculated by using the SHAP interpreter to enhance the explainability of the WaveMambaFormer model, which includes:

[0024] The trained WaveMambaFormer model and data are prepared, a deep learning SHAP interpreter DeepExplainer is created, and then SHAP values are calculated based on test data, and feature importance analysis and global explanation are performed through a visualization method; wherein the visualization method includes: summary plot, waterfall plot, dependence plot and heat map plot.

[0025] Preferably, in the step S5, the inversion performance of the WaveMambaFormer model is evaluated by taking the ERA5, WaveWatch III and Jason-3 significant wave height data as reference data, respectively, which includes:

[0026] The root mean square error RMSE, the bias Bias, the mean absolute percentage error MAPE and the Pearson correlation coefficient R are used as indexes for evaluating the inversion performance of the model, and the related calculation formulas are as follows:

[0027]

[0028] In the formula, m is the number of data samples, y i,M and y i,T are the estimated significant wave height of the model and the significant wave height obtained from the reference data, respectively, and are the average values of y i,M and y i,T , respectively.

[0029] The application also provides a satellite-borne GNSS-R ocean significant wave height inversion model construction system in high sea conditions, which is used to realize any one of the methods, and includes an acquisition module, an extraction module, a division module, a construction module and an evaluation module.

[0030] The acquisition module is configured to acquire modeling data, auxiliary data and verification data, wherein the modeling data comprises satellite-borne GNSS-R data and ERA5 significant wave height data; the auxiliary data comprises ERA5 data, SMAP data and OSCAR data; and the verification data comprises ERA5, WaveWatch III and Jason-3 significant wave height data.

[0031] The extraction module is configured to extract satellite-borne GNSS-R feature variable parameters and auxiliary variable parameters, and perform spatio-temporal matching on the extracted variable parameters.

[0032] The division module is configured to perform data quality control and data set division on the spatio-temporally matched variable parameters.

[0033] The construction module is configured to construct and train a WaveMambaFormer model by using the divided data set, and calculate feature importance and global explanation by using a SHAP interpreter to enhance the explainability of the WaveMambaFormer model.

[0034] The evaluation module is configured to evaluate the inversion performance of the WaveMambaFormer model by taking ERA5, WaveWatch III and Jason-3 significant wave height data as reference data, respectively.

[0035] Compared with the prior art, the present application has the following beneficial effects:

[0036] The application discloses a high-sea-state satellite-borne GNSS-R ocean significant wave height inversion model construction method and system, the inversion model is called WaveMambaFormer, and the interpretability of the model is enhanced through a SHAP (Shapley Additive exPlanations) interpreter, mainly including three modules, namely, a CNN-based DDM feature extraction module, a Transformer module and a state space model (SSM) module. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the present application, the following briefly introduces the drawings needed in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0038] Figure 1 A flow chart of the high-sea-state satellite-borne GNSS-R ocean significant wave height inversion model construction method provided in the embodiments of the present application;

[0039] Figure 2 A WaveMambaFormer model structure diagram in the embodiments of the present application;

[0040] Figure 3 Image interpretation of a bistatic radar cross section (BRCS) and an effective scattering area (eff_scatter) SHAP interpreter in the embodiments of the present application;

[0041] Figure 4SHAP waterfall plot in the embodiment of the present application;

[0042] Figure 5 SHAP summary feature plot in the embodiment of the present application;

[0043] Figure 6 SHAP dependence plot in the embodiment of the present application;

[0044] Figure 7 Comparison results of effective wave height inversed by different GNSS reflectometry signals and ERA5 data in the embodiment of the present application;

[0045] Figure 8 Global distribution of deviation between effective wave height inversed by different GNSS reflectometry signals and ERA5 data in the embodiment of the present application;

[0046] Figure 9 Comparison results of effective wave height inversed by different GNSS reflectometry signals and WaveWatch III data in the embodiment of the present application;

[0047] Figure 10 Comparison results of effective wave height inversed by different GNSS reflectometry signals and Jason-3 c-band (left) and Jason-3 ku-band (right) data in the embodiment of the present application Figure 1 ;

[0048] Figure 11 Comparison results of effective wave height inversed by different GNSS reflectometry signals and Jason-3 c-band (left) and Jason-3 ku-band (right) data in the embodiment of the present application Figure 2 ;

[0049] Figure 12 Accuracy of comparison of effective wave height inversed by different models and ERA5 data in different effective wave height intervals in the embodiment of the present application;

[0050] Figure 13 Accuracy of comparison of effective wave height inversed by different models and WaveWatch III data in different effective wave height intervals in the embodiment of the present application;

[0051] Figure 14 Accuracy of comparison of effective wave height inversed by different models and Jason-3 c-band data in different effective wave height intervals in the embodiment of the present application;

[0052] Figure 15 Accuracy of comparison of effective wave height inversed by different models and Jason-3 ku-band data in different effective wave height intervals in the embodiment of the present application. DETAILED DESCRIPTION

[0053] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of the present application.

[0054] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the embodiments of the present application should be understood as the common meanings of the technical terms or scientific terms by those skilled in the art to which the embodiments of the present application belong. The terms “first”, “second” and similar terms used in the embodiments of the present application do not represent any order, number or importance, but are only used to distinguish different components. The terms “include” or “contain” and similar terms mean that the components or objects before the terms cover the components or objects listed after the terms and their equivalents, and do not exclude other components or objects. The terms “connect” or “connected” and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. The terms “up”, “down”, “left”, “right” and the like are only used to represent relative positional relationships, and when the absolute positions of the described objects are changed, the relative positional relationships may also be changed accordingly.

[0055] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0056] Embodiment one

[0057] The fusion of CNN, Transformer and state space model (SSM) can effectively improve the performance of the significant wave height inversion model. The advantages of local feature extraction of CNN, global dependence modeling of Transformer and dynamic modeling of SSM are used to improve the prediction accuracy, robustness and adaptability to complex wave patterns of the model. This comprehensive method can provide strong support for the significant wave height inversion task in the space-borne GNSS-R data, especially when dealing with wave data with spatio-temporal complexity, it shows great advantages.

[0058] In addition, the combination of CNN (convolutional neural network), Transformer and state space model (SSM) for the construction of the significant wave height inversion model has the following main advantages:

[0059] (1) The advantage of CNN: extracting local features

[0060] 1) Strong local feature extraction capability: CNNs are good at extracting local features from Delay Doppler Map (DDM) image data. In significant wave height inversion, CNNs can effectively extract these local features through convolution operations. 2) Spatial information preservation: When dealing with spatial data, CNNs can preserve important spatial information through convolution layers and help identify local patterns and structures of waves.

[0061] (2) Advantages of Transformer: Modeling global dependencies

[0062] 1) Long-range dependency modeling capability: Transformers can capture global dependencies over long time series or spatial distances through self-attention mechanisms. In significant wave height inversion, the evolution of waves is influenced by multiple temporal and spatial factors, and Transformers can effectively model these long-range dependencies, making them suitable for tasks that require global context information.

[0063] 2) Parallel processing: Unlike traditional RNNs and LSTMs, Transformers can process data in parallel, greatly improving training efficiency. This is very useful for large-scale time series data processing and wave height inversion.

[0064] (3) Advantages of State Space Model (SSM): Modeling dynamic systems

[0065] Dynamic system modeling capability: SSMs are particularly good at modeling changes in hidden states during dynamic processes. In the significant wave height inversion problem, the wave evolution process is a dynamic system, and SSMs can effectively capture the hidden states and dynamic characteristics of wave evolution in time series.

[0066] Smoothness and robustness: Through state space models, smoothing and state estimation can be performed even when observation data is incomplete or noisy, improving the robustness and accuracy of the model.

[0067] (4) Advantages of model fusion: complementary advantages

[0068] 1) Comprehensive feature extraction and modeling capability: By combining the local feature extraction capability of CNNs, the global dependency modeling capability of Transformers, and the dynamic modeling capability of SSMs, we can fully utilize the advantages of different models in significant wave height inversion. CNNs can extract fine-grained local wave features, Transformers can capture long-term evolution patterns of waves, and SSMs can smooth the temporal dynamics of waves.

[0069] 2) Improve prediction accuracy: This multi-model fusion strategy can effectively capture the spatial and temporal characteristics of waves, improving the overall accuracy and generalization ability of the significant wave height inversion model.

[0070] (5) Adaptability to complex wave environments (such as high sea states)

[0071] Handling complex wave patterns: The wave height inversion model needs to consider various complex wave environments and interference factors, such as wind speed, rainfall, and other external conditions. By integrating CNN, Transformer, and SSM, these complex influencing factors can be more comprehensively modeled, enhancing the model's adaptability to different wave patterns.

[0072] (6) Efficiency and scalability

[0073] Model optimization and precision improvement: By jointly using these advanced network structures, the prediction accuracy can be improved while ensuring computational efficiency. The parallel processing capability of Transformer and the feature extraction capability of CNN help to accelerate the training process, while SSM can effectively model dynamic systems, reducing the demand for computing resources.

[0074] Therefore, the present application proposes a method for constructing a satellite-borne GNSS-R ocean significant wave height inversion model under high sea conditions. Specifically, a WaveMambaFormer model that integrates CNN, Transformer, and SSM is proposed for ocean significant wave height inversion under high sea conditions, and a SHAP interpreter is used to enhance the model's interpretability.

[0075] The specific implementation process includes: a method for constructing an ocean significant wave height inversion model under high sea conditions by integrating China Tianmu-1 satellite-borne GPS-R / BDS-R / GLONASS-R / Galileo-R data, including the following steps:

[0076] Step S1, obtain modeling data (including satellite-borne GNSS-R data, ERA5 significant wave height data), auxiliary data (including ERA5 data, SMAP data, OSCAR data), and verification data (including ERA5, WaveWatch III, and Jason-3 significant wave height data);

[0077] Step S2, extract satellite-borne GNSS-R feature variable parameters and auxiliary variable parameters, and perform spatio-temporal matching on the data set;

[0078] Step S3, data quality control and data set division;

[0079] Step S4, construction and training of the WaveMambaFormer model, and calculation of feature importance and global explanation using the SHAP interpreter to enhance the model's interpretability;

[0080] Step S5, respectively taking ERA5, WaveWatch III and Jason-3 significant wave height data as reference data, using root mean square error (RMSE), bias (Bias), mean absolute percentage error (MAPE) and Pearson correlation coefficient (R 2 ) as evaluation indexes of model inversion performance to evaluate the inversion performance of the model.

[0081] Preferably, the spaceborne GNSS-R data in the modeling data used in step S1 is Tianmoyi No. 1 GPS-R / BDS-R / GLONASS-R / Galileo-R L1 level data (provided free of charge by Spaceflight Tianmoyi (Chongqing) Satellite Technology Co., Ltd. and Spaceflight Science and Technology (Beijing) Space Information Application Co., Ltd.), and the ERA5 data in the auxiliary data includes ERA5 wind speed, ERA5 rainfall and ERA5 sea surface temperature (SST) data, the SMAP data includes SMAP sea surface salinity (SSS), and the OSCAR data is OSCAR sea surface current (OSCAR_currents) data including east-west direction current speed (OSCAR_currents_u) and north-south direction current speed (OSCAR_currents_v).

[0082] Preferably, step S2 includes the following sub-steps:

[0083] The extracted space-borne GNSS-R variable parameters in step S2.1 include bistatic radar cross section (BRCS), effective scatter area (eff_scatter), DDM peak signal-to-noise ratio (Ddm_peak_snr), DDM waveform front slope (Ddm_sp_les), DDM specular point normalized bistatic radar cross section (Ddm_sp_nbrcs), DDM specular point normalized signal-to-noise ratio (Ddm_sp_normalized_snr), DDM specular point signal-to-noise ratio (Ddm_sp_snr), DDM track ID (Ddm_track_id), GNSS satellite block code used to generate the DDM waveform (Gnss_block_flag), GNSS satellite PRN code used to generate the DDM waveform (Gnss_prn_code), altitude of the low earth orbit satellite corresponding to the DDM sampling mid-time (Rx_alt), latitude of the low earth orbit satellite corresponding to the DDM sampling mid-time (Rx_lat), longitude of the low earth orbit satellite corresponding to the DDM sampling mid-time (Rx_lon), pitch angle of the low earth orbit satellite corresponding to the DDM sampling mid-time (Rx_pitch), roll angle of the low earth orbit satellite corresponding to the DDM sampling mid-time (Rx_roll), X component of the low earth orbit satellite velocity corresponding to the DDM sampling mid-time (Rx_vel_x), Y component of the low earth orbit satellite velocity corresponding to the DDM sampling mid-time (Rx_vel_y), Z component of the low earth orbit satellite velocity corresponding to the DDM sampling mid-time (Rx_vel_z), specular point receiver antenna gain (Sp_antenna_gain), azimuth angle of the specular point in the antenna coordinate system (Sp_az_antenna), azimuth angle of the specular point in the satellite body coordinate system (Sp_az_body), azimuth angle of the specular point in the orbit coordinate system (Sp_az_orbit), azimuth angle of the specular point in the pattern coordinate system (Sp_az_pattern), GNSS signal incidence angle at the specular point (Sp_inc_angle), altitude of the specular point (Sp_alt), latitude of the specular point (Sp_lat), longitude of the specular point (Sp_lon), elevation angle of the specular point in the antenna coordinate system (Sp_theta_antenna), elevation angle of the specular point in the satellite body coordinate system (Sp_theta_body), elevation angle of the specular point in the orbit coordinate system (Sp_theta_orbit), elevation angle of the specular point in the pattern coordinate system (Sp_theta_pattern), X component of the specular point velocity (Sp_vel_x), Y component of the specular point velocity (Sp_vel_y), Z component of the specular point velocity (Sp_vel_z),a GNSS satellite velocity X component (Tx_vel_x) at a signal transmission time corresponding to the intermediate time of the DDM sampling, a GNSS satellite velocity Y component (Tx_vel_y) at the signal transmission time corresponding to the intermediate time of the DDM sampling, and a GNSS satellite velocity Z component (Tx_vel_z) at the signal transmission time corresponding to the intermediate time of the DDM sampling,

[0084] Step S2.2, after spatiotemporal matching of the characteristic variable parameters and metadata variables in the Tianmu No. 1 GPS-R / BDS-R / GLONASS-R / Galileo-R L1 level data with ERA5 significant wave height data, ERA5 wind speed, ERA5 rainfall, ERA5 sea surface temperature (SST), SMAP sea surface salinity (SSS), and OSCAR sea surface ocean current data, and ERA5 significant wave height data, WaveWatch III significant wave height data, and Jason-3 significant wave height data, a spatiotemporally matched data set is obtained.

[0085] As preferred, step S3 includes the following sub-steps:

[0086] Step S3.1, data filtering is performed using a basic quality control (QC) flag in the Tianmu No. 1 GNSS-R L1 data, and a single QC flag bit is used for data filtering. Data quality control includes selecting mirror reflection points on the ocean (i.e., setting Sp_surface_type = 0); the observation value must be positive, and when the value is a Nan value, it needs to be discarded; when the satellite roll angle is greater than 30 degrees, the yaw angle is greater than 5 degrees, or the pitch angle is greater than 10 degrees in absolute value, the data will be discarded; delete the radio frequency interference (RFI) data detected; if the estimated mirror point delay and Doppler shift have high uncertainty, delete the data; if the height of the satellite exceeds the nominal height range, delete the data; if the temperature and automatic gain control value of the delay Doppler map (DDM) power calibration condition exceed the normal range, delete the data;

[0087] Step S3.2, the data set is divided into a training set accounting for 60% of the total data set and a test set accounting for 40% of the total data set.

[0088] As preferred, step 4 includes the following sub-steps:

[0089] Step 4.1, the effective wave height inversion model is called WaveMambaFormer, which is composed of three modules: a CNN-based DDM feature extraction module, a Transformer-based module, and a state space model (SSM) module. CNN is very suitable for processing local features in images or signals, Transformer has a unique self-attention mechanism that can capture long-time scale or long-distance dependencies in signals, and the dynamic modeling capability, noise suppression capability, and adaptability to time-varying processes of the SSM model are particularly suitable for describing the complexity and variability of sea surface conditions. Specifically, first, the local spatial information of the signal in the BRCS is extracted through multiple convolution layers of CNN, and global average pooling is used after convolution to retain significant features and reduce parameter quantity, greatly improving the training speed of the model. Then these features are input into the Transformer to further capture long-range dependencies and temporal relationships. Second, due to the highly nonlinear relationship between the scattering of ocean surface waves and reflected signals, Transformer can handle such complex relationships and capture long-range dependencies between different time points through self-attention mechanisms, thus more accurately extracting information from eff_scatter in the reflected signal. Finally, the dynamic multi-path SSM module extracts features from variable parameters.

[0090] The main part of the Transformer module is the encoder module, which is used to extract global information and consists of multi-head attention mechanisms, residual connections, layer normalization, feedforward neural networks, and fully connected networks. The multi-head attention mechanism consists of multiple self-attention mechanisms. For a single self-attention mechanism, if the input vector matrix is represented as , then Q (query), K (key), and V (value) are obtained by multiplying the input matrix X and three learnable weight matrices W Q , W K , and W V , respectively. The relevant calculation formulas are as follows:

[0091] Q = XW Q (1)

[0092] K = XW K (2)

[0093] V = XW V (3)

[0094] After obtaining the matrices Q, K, and V, the output of the self-attention mechanism Self-Attention can be calculated, and the calculation formula is as follows:

[0095]

[0096] where d k is the number of columns of Q, K matrix.

[0097] Residual connection can pass the gradient between different layers, effectively alleviate the problem of gradient disappearance, and layer normalization helps to achieve faster convergence speed of the model during the training phase and improve the generalization ability of the model.

[0098] SSM can describe the dynamic changes of the system, especially suitable for data modeling with time dependence, can handle multiple variables and their mutual influence at the same time, and is suitable for modeling complex systems. SSM comes from control system theory, which originally deals with continuous functions, including state equation and observation equation, and the calculation expression is as follows:

[0099] h'(t) = Ah(t) + Bx(t) (5)

[0100] y(t) = Ch(t) (6)

[0101] where x(t) is the input sequence, h(t) is the hidden state representation, and y(t) is the predicted output sequence. Discretize SSM in discrete form and apply it to deep learning, A and B are discretized by zero-order hold (ZOH) method, and the time scale parameter is Δ.

[0102]

[0103] After discretization, equations (1) and (2) can be rewritten as:

[0104]

[0105] The final output y is obtained by the following calculation formula:

[0106]

[0107] where L represents the length of the input sequence, represents a structured convolution kernel.

[0108] In the model training process, MSE or MAE is usually used as the loss function, but they have their own defects. MSE is very sensitive to outliers because the square term will amplify large errors. MAE has a certain robustness to outliers, but it will show a gradient discontinuity problem when the error is close to 0. However, Log-Cosh punishes outliers less than MSE, and the gradient changes smoothly when the error is close to 0. Log-Cosh Loss function combines the smoothness of MSE and the robustness of MAE, which can effectively optimize small errors and avoid the excessive influence of outliers on the model. It is particularly suitable for regression tasks in deep learning models, and can balance high precision and robustness. Therefore, we introduce the Log-Cosh Loss function as the loss function. The calculation formula of Log-Cosh Loss is as follows:

[0109]

[0110] where, represents the predicted significant wave height, y i represents the reference significant wave height.

[0111] Step 4.2, use SHAP (SHapley Additive exPlanations) interpreter to calculate feature importance and global explanation, and enhance the interpretability of the model. It usually includes the following steps: prepare the trained WaveMambaFormer model and data, create a deep learning SHAP interpreter (DeepExplainer), then calculate SHAP values based on test data, and perform feature importance analysis and global explanation through visualization methods (Summary plot, Waterfall Plot, Dependence plot and Heatmap plot).

[0112] Step 4.3, use the feature variable parameters and metadata variables in the Tianmu-1 GPS-R / BDS-R / GLONASS-R / Galileo-R L1 level data, ERA5 wind speed, ERA5 rainfall, ERA5 sea surface temperature (SST), SMAP sea surface salinity (SSS) and OSCAR sea surface current data as input data for WaveMambaFormer model training, and ERA5 significant wave height data as target data. Among them, the GNSS satellite batch code (Gnss_block_flag) used to generate DDM waveforms helps to fuse GPS-R / BDS-R / GLONASS-R / Galileo-R data to realize multi-constellation GNSS-R construction of significant wave height inversion model.

[0113] As preferred, the inversion performance of the model is evaluated by taking ERA5, WaveWatch III and Jason-3 significant wave height products as reference data respectively in step S5. The root mean square error (RMSE), bias, mean absolute percentage error (MAPE) and Pearson correlation coefficient (R) are used as indicators to evaluate the inversion performance of the model, and the calculation formulas are as follows:

[0114]

[0115] In the formula, m is the number of data samples, y i,M and y i,T are the significant wave height estimated by the model and obtained from the reference data respectively, and are the average values of y i,M and y i,T respectively.

[0116] It should be noted that the method of the embodiments of the present disclosure can be executed by a single device, such as a computer or a server, etc. The method of the embodiments can also be applied to a distributed scenario, and be completed by multiple devices cooperating with each other. In the case of such a distributed scenario, one of the multiple devices can only execute one or more steps in the method of the embodiments of the present disclosure, and the multiple devices can interact with each other to complete the method.

[0117] It should be noted that some embodiments of the present disclosure have been described above. Other embodiments are within the scope of the appended claims. In some cases, the size of the serial number of the steps in the above embodiments does not mean the order of execution, and the execution order of the processes should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present disclosure. The actions or steps recited in the claims can be executed in an order different from that in the above embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are possible or can be advantageous.

[0118] Embodiment Two

[0119] Based on the same inventive concept, the present disclosure also provides a high sea state satellite-borne GNSS-R ocean significant wave height inversion model construction system corresponding to any of the above-mentioned embodiment methods, which is used to implement any of the above-mentioned methods, and comprises an acquisition module, an extraction module, a division module, a construction module and an evaluation module.

[0120] The acquisition module is configured to acquire modeling data, auxiliary data, and verification data, wherein the modeling data comprises satellite-borne GNSS-R data and ERA5 significant wave height data; the auxiliary data comprises ERA5 data, SMAP data, and OSCAR data; and the verification data comprises ERA5 data, WaveWatch III data, and Jason-3 significant wave height data.

[0121] The extraction module is configured to extract satellite-borne GNSS-R feature variable parameters and auxiliary variable parameters, and perform spatio-temporal matching on the extracted variable parameters.

[0122] The division module is configured to perform data quality control and data set division on the spatio-temporal matched variable parameters.

[0123] The construction module is configured to construct and train the WaveMambaFormer model by using the divided data set, and calculate feature importance and global explanation by using a SHAP interpreter to enhance the explainability of the WaveMambaFormer model.

[0124] The evaluation module is configured to evaluate the inversion performance of the WaveMambaFormer model by taking ERA5 significant wave height data, WaveWatch III significant wave height data, and Jason-3 significant wave height data as reference data, respectively.

[0125] The system of the above embodiment is configured to implement the high sea state satellite-borne GNSS-R ocean significant wave height inversion model construction method of any one of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be described herein.

[0126] It should be noted that the high sea state satellite-borne GNSS-R ocean significant wave height inversion model construction system is embodied in the form of a functional unit. The term "module" herein can be implemented in the form of software and / or hardware, and is not specifically limited.

[0127] For example, the "module" can be a software program, a hardware circuit, or a combination of the two, which implements the above functions. The hardware circuit can include an application specific integrated circuit (ASIC), an electronic circuit, a processor (such as a shared processor, a dedicated processor, or a group processor) and a memory for executing one or more software or firmware programs, and other suitable components for supporting the described functions.

[0128] Embodiment three

[0129] The application provides a high sea condition ocean significant wave height inversion model construction method fusing China Tianmu No.1 satellite-borne GPS-R / BDS-R / GLONASS-R / Galileo-R data, which will be described in detail below with reference to the drawings:

[0130] In order to verify the effectiveness of the method, Tianmu No.1 GPS-R / BDS-R / GLONASS-R / Galileo-R L1 data (provided by Space Tianmu (Chongqing) Satellite Technology Co., Ltd., Space Science and Technology Corporation of China (Beijing) Space Information Application Co., Ltd. for free) from March 2024 to April 2024, ERA5 significant wave height data, ERA5 wind speed data, ERA5 rain rate data, ERA5 sea surface temperature (SST) data, SMAP sea surface salinity (SSS) data, OSCAR sea surface currents (OSCAR_currents) data and ERA5, WW3 and Jason3 significant wave height data are selected, wherein the Tianmu No.1 GPS-R / BDS-R / GLONASS-R / Galileo-R L1 data and the ERA5 significant wave height data are used for model establishment; the ERA5 wind speed data, the ERA5 rain rate data, the ERA5 sea surface temperature (SST) data, the SMAP sea surface salinity (SSS) data and the OSCAR sea surface currents (OSCAR_currents) data are used as auxiliary data; the ERA5 significant wave height data, the WaveWatch III significant wave height data and the Jason-3 significant wave height data are used as reference data, and the root mean square error (RMSE), the bias (Bias), the mean absolute percentage error (MAPE), the determination coefficient (R 2 ) and the Pearson correlation coefficient (R) are used as five indexes for evaluating the inversion performance of the model.

[0131] The basic configuration of the experimental platform of the embodiment is shown in Table 1:

[0132] Table 1 Configuration of experimental platform

[0133]

[0134] A high sea condition ocean significant wave height inversion model construction method fusing China Tianmu No.1 satellite-borne GPS-R / BDS-R / GLONASS-R / Galileo-R data, technical solution implementation process as shown in the attached Figure 1 , comprising the following steps:

[0135] Step S1, obtaining modeling data (including spaceborne GNSS-R data, ERA5 significant wave height data), auxiliary data (including ERA5 data, SMAP data, OSCAR data) and verification data (including ERA5, WW3 and Jason3 significant wave height data);

[0136] Step S2, extracting spaceborne GNSS-R feature variable parameters and auxiliary variable parameters, and performing spatio-temporal matching on the data set;

[0137] Step S3, data filtering, data quality control and data set division;

[0138] Step S4, construction and training of WaveMambaFormer model, and calculation of feature importance and global explanation by SHAP interpreter to enhance the explainability of the model;

[0139] Step S5, evaluation of the inversion performance of the model by taking ERA5, WaveWatch III and Jason-3 significant wave height data as reference data respectively.

[0140] As an embodiment of the present embodiment, in step S1, the spaceborne GNSS-R data in the modeling data used is Tianmoyihao GPS-R / BDS-R / GLONASS-R / Galileo-R L1 level data (provided free of charge by Hangtian Tianmoyihao (Chongqing) Satellite Technology Co., Ltd. and Hangtian Keji (Beijing) Space Information Application Co., Ltd.), and the ERA5 data in the auxiliary data includes ERA5 wind speed, ERA5 rain rate and ERA5 sea surface temperature (SST) data, the SMAP data includes SMAP sea surface salinity (SSS), and the OSCAR data is OSCAR sea surface current (OSCAR_currents) data including east-west direction flow rate (OSCAR_currents_u) and north-south direction flow rate (OSCAR_currents_v).

[0141] As an embodiment of the present embodiment, step S2 includes the following sub-steps:

[0142] The extracted space-borne GNSS-R variable parameters in step S2.1 include bistatic radar cross section (BRCS), effective scatter area (eff_scatter), DDM peak signal-to-noise ratio (Ddm_peak_snr), DDM waveform front slope (Ddm_sp_les), DDM specular point normalized bistatic radar cross section (Ddm_sp_nbrcs), DDM specular point normalized signal-to-noise ratio (Ddm_sp_normalized_snr), DDM specular point signal-to-noise ratio (Ddm_sp_snr), DDM track ID (Ddm_track_id), GNSS satellite block code used to generate the DDM waveform (Gnss_block_flag), GNSS satellite PRN code used to generate the DDM waveform (Gnss_prn_code), altitude of the low earth orbit satellite corresponding to the DDM sampling mid-time (Rx_alt), latitude of the low earth orbit satellite corresponding to the DDM sampling mid-time (Rx_lat), longitude of the low earth orbit satellite corresponding to the DDM sampling mid-time (Rx_lon), pitch angle of the low earth orbit satellite corresponding to the DDM sampling mid-time (Rx_pitch), roll angle of the low earth orbit satellite corresponding to the DDM sampling mid-time (Rx_roll), X component of the low earth orbit satellite velocity corresponding to the DDM sampling mid-time (Rx_vel_x), Y component of the low earth orbit satellite velocity corresponding to the DDM sampling mid-time (Rx_vel_y), Z component of the low earth orbit satellite velocity corresponding to the DDM sampling mid-time (Rx_vel_z), specular point receiver antenna gain (Sp_antenna_gain), azimuth angle of the specular point in the antenna coordinate system (Sp_az_antenna), azimuth angle of the specular point in the satellite body coordinate system (Sp_az_body), azimuth angle of the specular point in the orbit coordinate system (Sp_az_orbit), azimuth angle of the specular point in the pattern coordinate system (Sp_az_pattern), GNSS signal incidence angle at the specular point (Sp_inc_angle), altitude of the specular point (Sp_alt), latitude of the specular point (Sp_lat), longitude of the specular point (Sp_lon), elevation angle of the specular point in the antenna coordinate system (Sp_theta_antenna), elevation angle of the specular point in the satellite body coordinate system (Sp_theta_body), elevation angle of the specular point in the orbit coordinate system (Sp_theta_orbit), elevation angle of the specular point in the pattern coordinate system (Sp_theta_pattern), X component of the specular point velocity (Sp_vel_x), Y component of the specular point velocity (Sp_vel_y), Z component of the specular point velocity (Sp_vel_z),The DDM acquires the GNSS satellite velocity X component (Tx_vel_x), Y component (Tx_vel_y), and Z component (Tx_vel_z) at the corresponding signal transmission time at the intermediate time.

[0143] Step S2.2: The feature variable parameters and metadata variables in the Tianmu-1 GPS-R / BDS-R / GLONASS-R / Galileo-R L1 level data are spatiotemporally matched with ERA5 significant wave height data, ERA5 wind speed, ERA5 rainfall, ERA5 sea surface temperature (SST), SMAP sea surface salinity (SSS), and OSCAR sea surface current data, as well as ERA5 significant wave height data, WaveWatch III significant wave height data, and Jason-3 significant wave height data to obtain the spatiotemporally matched dataset.

[0144] As one implementation of this embodiment, step S3 includes the following sub-steps:

[0145] Step S3.1, data filtering involves using basic quality control (QC) flags in the Tianmu-1 GNSS-R L1 data, applying data filtering with a single QC flag bit. Data quality control includes selecting specular reflection points on the ocean (i.e., setting Sp_surface_type = 0); observations must be positive, discarding values ​​with a Nan value; data will be discarded if the absolute value of the satellite roll angle is greater than 30 degrees, yaw angle is greater than 5 degrees, or pitch angle is greater than 10 degrees; data with detected radio frequency interference (RFI) will be deleted; data will be deleted if the estimated specular point delay and Doppler shift have high uncertainty; data will be deleted if the satellite altitude exceeds the nominal altitude range; and data will be deleted if the temperature and automatic gain control values ​​of the DDM power calibration conditions are outside the normal range.

[0146] Step S3.2: The dataset is divided into two parts: the training set accounts for 60% of the total dataset and the test set accounts for 40% of the total dataset.

[0147] As one implementation of this embodiment, step 4 includes the following sub-steps:

[0148] Step 4.1, the constructed effective wave height inversion model is called WaveMambaFormer, and the model structure is shown in the attached figure. Figure 2As shown, the model is composed of a CNN-based DDM feature extraction module, a Transformer-based module, and a state space model (SSM) module. CNN has high training efficiency, and using global average pooling after convolution operation can retain significant features and reduce parameter quantity, greatly improving the training speed of the model; Transformer can use self-attention mechanism and feedforward neural network to capture the relationship between elements in the input sequence to process long-distance dependence. The main part of the Transformer module is the encoder module, which is used to extract global information and consists of multi-head attention mechanism, residual connection, layer normalization, feedforward neural network, and fully connected network.

[0149] The multi-head attention mechanism is composed of multiple self-attention mechanisms. For a single self-attention mechanism, if the input vector matrix is represented as , then Q (query), K (key), and V (value) are obtained by multiplying the input matrix X and three learnable weight matrices W Q , W K , and W V , and the relevant calculation formula is as follows:

[0150] Q = XW Q (1)

[0151] K = XW K (2)

[0152] V = XW V (3)

[0153] After obtaining the matrices Q, K, and V, the output of the self-attention mechanism Self-Attention can be calculated, and the calculation formula is as follows:

[0154]

[0155] In the formula, d k is the number of columns of the Q and K matrices.

[0156] Residual connection can pass gradients between different layers, effectively alleviating the problem of gradient vanishing, and layer normalization helps to achieve faster convergence speed of the model during training and improve the generalization ability of the model.

[0157] SSM can describe the dynamic changes of the system, and is particularly suitable for data modeling with time dependence. It can handle multiple variables and their mutual influence at the same time, and is suitable for modeling complex systems. SSM originates from control system theory, which deals with continuous functions and contains state equations and observation equations. The calculation expression is as follows:

[0158] h'(t) = Ah(t) + Bx(t) (5)

[0159] y(t) = Ch(t) (6)

[0160] In the formula, x(t) is the input sequence, h(t) is the hidden state, and y(t) is the predicted output sequence. The SSM is applied to deep learning in discrete form, A and B are discretized by the zero-order hold (ZOH) method, and the time scale parameter is Δ.

[0161]

[0162] After discretization, equations (1) and (2) can be rewritten as:

[0163]

[0164] The final output y is obtained by the following calculation formula:

[0165]

[0166] In the formula, L represents the length of the input sequence, represents the structured convolution kernel.

[0167] During model training, the Log-Cosh Loss function combines the smoothness of mean square error (MSE) and the robustness of mean absolute error (MAE), which can effectively optimize small errors and avoid the excessive influence of outliers on the model. It is particularly suitable for regression tasks in deep learning models, and can balance high precision and robustness. Therefore, we introduce the Log-Cosh Loss function as the loss function, and the calculation formula of Log-Cosh Loss is as follows:

[0168]

[0169] In the formula, represents the predicted significant wave height, y i represents the reference significant wave height.

[0170] Step 4.2, in order to calculate the feature importance and global explanation using the SHAP interpreter, enhance the explainability of the model, and Figure 3 The SHAP explanation of the image is shown, from which the SHAP values of each pixel on the image can be visually observed. The Figure 4The diagram shows a SHAP waterfall plot, where each bar represents the SHAP value of a feature. Positive values ​​indicate a positive contribution of the feature to the prediction result, while negative values ​​indicate a negative contribution. For the model proposed in this invention, the expected value of the significant wave height is 2.767. The SHAP values ​​for significant wave height are predominantly positive, leading to an improved inversion rate. The final sample's SHAP value is 3.105, with sp_lat having the largest negative impact in this sample. (Appendix) Figure 5 The diagram shows the SHAP comprehensive feature analysis (feature importance distribution plot (left), summary plot (middle), and heatmap (right)). The feature importance distribution plot (left) averages the absolute values ​​of the SHAP values ​​for each feature, revealing that windspeed, Sp_lat, and SST are the three most important features. The summary plot (middle) provides a global interpretation of the model's SHAP values; each point represents a sample, with redder colors indicating larger feature values ​​and bluer colors indicating smaller feature values. Windspeed is a significant feature and is generally positively correlated with the model's output SHAP value. The heatmap (right) clusters data points based on feature SHAP values, using supervised clustering and the heatmap to display the overall substructure of the dataset. f(x) on the X-axis represents the sum of the SHAP values ​​for each sample, with the gray dashed line being the baseline. The bar chart to the right of the Y-axis shows the global importance of each model input. The heatmap effectively visualizes and interprets the model.

[0171] Appendix Figure 6 The SHAP dependency graph illustrates how one feature interacts with another feature to affect the model's prediction. The graph shows how the maximum absolute value, average absolute value, and 95th percentile absolute value of a feature interact with another feature to affect the model's prediction. The interaction between features can be seen intuitively from the graph.

[0172] Furthermore, in step 4.3, the characteristic variable parameters and metadata variables from the Tianmu-1 GPS-R / BDS-R / GLONASS-R / Galileo-R L1 level data, along with ERA5 wind speed, ERA5 rainfall rate, ERA5 sea surface temperature (SST), SMAP sea surface salinity (SSS), and OSCAR ocean current data, are used as input data for training the WaveMambaFormer model. ERA5 significant wave height data is used as target data. The GNSS satellite batch code (Gnss_block_flag) used to generate the DDM waveform is used as input to help fuse GPS-R / BDS-R / GLONASS-R / Galileo-R data to realize the construction of a multi-constellation GNSS-R significant wave height inversion model.

[0173] In one implementation of this embodiment, step S5 involves evaluating the model's inversion performance using ERA5, WaveWatch III, and Jason-3 significant wave height products as reference data. Root mean square error (RMSE), bias, mean absolute percentage error (MAPE), and Pearson correlation coefficient (R) are used as indicators to evaluate the model's inversion performance, and the relevant calculation formulas are as follows:

[0174]

[0175] In the formula, m is the number of data samples, and y i,M and y i,T These are the effective wave heights estimated by the model and the effective wave heights obtained from the reference data, respectively. and They are y i,M and y i,T The average value.

[0176] To evaluate the differences in sea surface significant wave (SSW) performance among different GNSS reflection signals from Tianmu-1, the trained WaveMambaFormer model was used to retrieve SSW height. ERA5 SSW data, WaveWatchIII SSW data, and Jason-3 SSW data were used as reference data to evaluate the retrieval performance. Figure 7 This paper presents a comparison of the significant wave height retrieved from different GNSS reflection signals with ERA5 data, as shown in the appendix. Figure 7 Among the five GNSS reflection signals—BDS-R, GPS-R, Galileo-R, GLONASS-R, and Multi-GNSS-R—the BDS-R signal inversion yielded the best results, with RMSE, Bias, MAPE, and R... 2 The values ​​were 0.287m, -0.003m, 8.69%, and 0.936, respectively. Multi-GNSS-R also showed good inversion results, with RMSE, Bias, MAPE, and R... 2 The effective wave heights obtained from Multi-GNSS-R signal inversion are 0.655m, -0.034m, 19.43%, and 0.708, respectively. Compared to BDS-R signals, the effective wave height range obtained from Multi-GNSS-R signal inversion is mainly between 0-12m, representing a wider range. In contrast, the effective wave height range obtained from BDS-R signal inversion is mainly between 0-8m, representing a smaller range. Overall, the results obtained from Multi-GNSS-R signal inversion are the best. (Appendix) Figure 8The global distribution of the bias between the significant wave height retrieved from different GNSS reflection signals and ERA5 data can be clearly seen. The bias between the significant wave height retrieved from Multi-GNSS-R and ERA5 data is the smallest in the global range, and has higher accuracy. The results retrieved from Galileo-R and GLONASS-R have larger bias.

[0177] Figure 4 Figure 9 Figure 4 Figure 10 Figure 4 Figure 11 Figure 4 shows the comparison results of the significant wave height retrieved from different GNSS reflection signals and WaveWatch III data, and the comparison results of the significant wave height retrieved from different GNSS reflection signals and Jason-3 c-band (left) and Jason-3 ku-band (right) data, respectively, wherein Figure 9 Figure 4 Figure 10 Figure 4 Figure 11 It can be obtained that:

[0178] (1) When the WaveWatch III significant wave height and Jason-3 significant wave height data are used as reference data, the significant wave height retrieved from BDS-R and Multi-GNSS-R has the best inversion accuracy, while the results retrieved from Galileo-R and GLONASS-R have poor inversion accuracy. This is mainly because the range of the significant wave height retrieved from Galileo-R and GLONASS-R is larger. When the WaveWatch III significant wave height data are used as reference data, the range of the wave height is about 0-12 m; when the Jason-3 significant wave height data are used as reference data, the range of the wave height is about 0-8 m. The range of the significant wave height retrieved from BDS-R is smaller. When the WaveWatch III significant wave height data are used as reference data, the range of the wave height is about 0-8 m; when the Jason-3 significant wave height data are used as reference data, the range of the wave height is about 0-5 m, so that the inversion accuracy of BDS-R is higher than that of Galileo-R and GLONASS-R. Compared with BDS-R with smaller inversion range, the significant wave height retrieved from Multi-GNSS-R not only has a larger wave height range, but also has better inversion accuracy than GPS-R, Galileo-R and GLONASS-R. Overall, Multi-GNSS-R has the best inversion performance.

[0179] (2) In Figure 4 Figure 10 Figure 4 Figure 11It can be seen that in addition to Galileo-R, the inversion results obtained by taking Jason-3c-band data as reference data are generally better than those obtained by taking Jason-3 ku-band data. When taking Jason-3 ku-band data as reference data, Multi-GNSS-R still has good inversion performance, and the RMSE, Bias, MAPE and R 2 0.542 m, 0.075 m, 19.69%, and 0.762, respectively.

[0180] In addition, in order to further evaluate the inversion performance of the model, the present application compares the effective wave height results obtained by the WaveMambaFormer model with those obtained by bagging tree (BT), artificial neural network (ANN), deep convolutional neural network (DCNN), CNN-BiLSTM, and Transformer. Tables 1, 2, 3, and 4 are the precisions of the inversion effective wave heights of different models compared with ERA5 data, WaveWatch III data, Jason-3 c-band data, and Jason-3 ku-band data, respectively, wherein IR is the improvement rate of WaveMambaFormer model in RMSE compared with the other five machine learning or deep learning models. From Tables 2, 3, 4, and 5, it can be seen that when ERA5 data, WaveWatch III data, Jason-3 c-band data, and Jason-3 ku-band data are respectively taken as reference data, the precision of the WaveMambaFormer model is better than that of the other five machine learning or deep learning models. However, due to the difference in the range of the inversion effective wave height, the inversion precision obtained by taking Jason-3 c-band data and Jason-3 ku-band data as reference data is better than that obtained by taking ERA5 data and WaveWatch III data as reference data. When the range of the effective wave height is 0-12 m, the optimal RMSE of 0.655 m is obtained by taking ERA5 data as reference data, and the improvement rates (IR) in RMSE are 12.90%, 8.07%, 20.19%, 26.13%, and 12.25%, respectively. When the range of the effective wave height is 0-7 m, the optimal RMSE and Bias of 0.543 m and -0.007 m, respectively, are obtained by taking Jason-3 ku-band data as reference data, and the IRs are 19.90%, 4.26%, 19.86%, 25.26%, and 15.70%, respectively.

[0181] Table 2 Precision of inversion effective wave height of different models compared with ERA5 data (range of effective wave height: 0-12 m)

[0182]

[0183] Table 3 Accuracy of different models inverting significant wave height compared with WaveWatch3 data (significant wave height range:

[0184] 0-12m)

[0185]

[0186] Table 4 Accuracy of different models inverting significant wave height compared with Jason3 c-band data (significant wave height range:

[0187] 0-7m)

[0188]

[0189] Table 5 Accuracy of different models inverting significant wave height compared with Jason3 ku-band data (significant wave height range: 0-7m)

[0190]

[0191] The performance of each model at different wave heights is also a concern and an important indicator of evaluating the goodness of the model. Figures 1-3 show the accuracy of different models inverting significant wave height in different significant wave height intervals when ERA5 significant wave height, WaveWatch III significant wave height and Jason-3 significant wave height data are used as reference data, respectively. Figure 12 Figure 13 Figure 14 Figure 15 Figures 1-3 show that: Figure 12 Figure 13 Figure 14 Figure 15

[0192] (1) WaveMambaFormer model, Transformer model and artificial neural network model have better inversion accuracy than other models in different wave height intervals, and have good performance in RMSE, Bias, MAPE and R. With the continuous increase of wave height, the RMSE of the model becomes larger and larger, which shows that the increase of wave height makes the performance of the model worse. MAPE presents a trend of first decreasing and then increasing, that is, the accuracy of the prediction model is higher in the range of moderate wave height. R presents a trend of first increasing, then decreasing and then increasing. When the significant wave height is between 1m-2m and the wave height is higher, the correlation coefficient is the largest and the linear relationship is stronger.

[0193] ​​​​​​​(2) In terms of Bias, when the wave height is less than 3 m, the Bias value of the model is positive, and the Bias value is small. However, when the wave height is greater than 3 m, the Bias of the effective wave height obtained by the five inversion models is less than 0, that is, the five inversion models all have the phenomenon of underestimation when the wave height is high. Especially when the ERA5 effective wave height and the WaveWatch III effective wave height data are used as the reference data, with the wave height becoming higher and higher, the Bias of the model also becomes larger and larger, and the predicted value of the model deviates from the true value more and more. The main reason may be that the training sample of the high effective wave height is less, which limits the inversion performance of the model when the wave height is high. When the Jason-3c wave band data and the Jason-3ku wave band data are used as the reference data, the model has the maximum Bias when the wave height is in the interval of 5 m-6 m. Compared with the other five machine learning or deep learning models, the WaveMambaFormer model has the best performance in terms of Bias. Although the Transformer model also has good inversion results when the wave height is low, with the continuous increase of the wave height, the performance of the WaveMambaFormer model in terms of Bias is obviously better than that of the other five machine learning (or deep learning) models.

[0194] The embodiments of the present disclosure are intended to cover all such alternatives, modifications, and variations of the present disclosure falling within the broadest scope of the appended claims. Accordingly, any and all such modifications, variations, or equivalents that fall within the spirit and scope of the embodiments of the present disclosure are included herein.

Claims

1. A method for constructing an ocean significant wave height inversion model under high sea conditions, characterized in that, The method comprises: Step S1, obtaining modeling data, auxiliary data and verification data, wherein the modeling data comprises: space-borne GNSS-R data, ERA5 significant wave height data; the auxiliary data comprises: ERA5 data, SMAP data, OSCAR data; the verification data comprises: ERA5, WaveWatch III and Jason-3 significant wave height data; Step S2, extracting space-borne GNSS-R feature variable parameters and auxiliary variable parameters, and performing space-time matching on the extracted variable parameters; Step S3, performing data quality control and data set division on the space-time matched variable parameters; Step S4, constructing and training a WaveMambaFormer model by using the divided data set, and calculating feature importance and global explanation by using a SHAP interpreter to enhance the explainability of the WaveMambaFormer model; Step S5, respectively taking ERA5, WaveWatch III and Jason-3 significant wave height data as reference data to evaluate the inversion performance of the WaveMambaFormer model; In the step S4, the WaveMambaFormer model is composed of a CNN-based DDM feature extraction module, a Transformer-based module and a state space model State Space Model, SSM module; Firstly, the local spatial information of the bistatic radar cross section BRCS signal is extracted through the multi-layer convolution layer of the CNN, and after the convolution operation, the global average pooling is used to retain the preset significant features and reduce the parameter quantity, then the retained significant features are input into the Transformer to further capture the long-distance dependence and time sequence relationship; and the long-range dependence relationship between different time points is captured through the self-attention mechanism to extract the information in eff_scatter from the reflection signal; finally, the dynamic multi-path SSM module is used to extract features from the variable parameters.

2. The method of claim 1, wherein, In the step S2, the space-borne GNSS-R feature variable parameters and auxiliary variable parameters are extracted, and the extracted variable parameters are space-time matched, which comprises: The extracted space-borne GNSS-R variable parameters in step S2.1 include bistatic radar cross section BRCS, effective scattering area eff_scatter, DDM peak signal-to-noise ratio Ddm_peak_snr, DDM waveform leading edge slope Ddm_sp_les, DDM specular point normalized bistatic radar cross section Ddm_sp_nbrcs, normalized specular point signal-to-noise ratio Ddm_sp_normalized_snr, DDM specular point signal-to-noise ratio Ddm_sp_snr, DDM track ID Ddm_track_id, GNSS satellite block code used to generate the DDM waveform Gnss_block_flag, GNSS satellite PRN code used to generate the DDM waveform Gnss_prn_code, altitude of the low earth orbit satellite corresponding to the DDM sampling time Rx_alt, latitude of the low earth orbit satellite corresponding to the DDM sampling time Rx_lat, longitude of the low earth orbit satellite corresponding to the DDM sampling time Rx_lon, pitch angle of the low earth orbit satellite corresponding to the DDM sampling time Rx_pitch, roll angle of the low earth orbit satellite corresponding to the DDM sampling time Rx_roll, X component of the low earth orbit satellite velocity corresponding to the DDM sampling time Rx_vel_x, Y component of the low earth orbit satellite velocity corresponding to the DDM sampling time Rx_vel_y, Z component of the low earth orbit satellite velocity corresponding to the DDM sampling time Rx_vel_z, specular point receiver antenna gain Sp_antenna_gain, azimuth angle of the specular point in the antenna coordinate system Sp_az_antenna, azimuth angle of the specular point in the satellite body coordinate system Sp_az_body, azimuth angle of the specular point in the orbit coordinate system Sp_az_orbit, azimuth angle of the specular point in the pattern coordinate system Sp_az_pattern, GNSS signal incidence angle at the specular point Sp_inc_angle, altitude of the specular point Sp_alt, latitude of the specular point Sp_lat, longitude of the specular point Sp_lon, elevation angle of the specular point in the antenna coordinate system Sp_theta_antenna, elevation angle of the specular point in the satellite body coordinate system Sp_theta_body, elevation angle of the specular point in the orbit coordinate system Sp_theta_orbit, elevation angle of the specular point in the pattern coordinate system Sp_theta_pattern, X component of the specular point velocity Sp_vel_x, Y component of the specular point velocity Sp_vel_y, Z component of the specular point velocity Sp_vel_z, X component of the GNSS satellite velocity at the signal transmission time corresponding to the DDM sampling time Tx_vel_x, Y component of the GNSS satellite velocity at the signal transmission time corresponding to the DDM sampling time Tx_vel_y,The DDM takes the Z component of the velocity of the GNSS satellite at the time of signal transmission corresponding to the intermediate time Tx_vel_z; Step S2.2, performing space-time matching on the feature variable parameters and metadata variables in the Tianmu No. 1 GPS-R / BDS-R / GLONASS-R / Galileo-R L1 level data, ERA5 significant wave height data, ERA5 wind speed, ERA5 rainfall, ERA5 sea surface temperature SST, SMAP sea surface salinity SSS and OSCAR sea surface current data, and ERA5 significant wave height data, WaveWatch III significant wave height data and Jason-3 significant wave height data to obtain the space-time matched data set.

3. The method of claim 1, wherein, In the step S3, the data quality control of the space-time matched variable parameters comprises: The point of specular reflection on the sea is selected, i.e., Sp_surface_type = 0 is set; the observation value must be positive, and when the value is a Nan value, the data needs to be discarded; when the satellite roll angle is greater than 30 degrees, the yaw angle is greater than 5 degrees, or the absolute value of the pitch angle is greater than 10 degrees, the data will be discarded; the data of detected radio frequency interference (RFI) is deleted; if the uncertainty of the estimated specular point delay and Doppler shift exceeds the preset requirement, the data is deleted; if the height of the satellite exceeds the nominal height range, the data is deleted; if the temperature and automatic gain control value of the delay Doppler map (DDM) power calibration condition exceed the preset range, the data is deleted.

4. The method of claim 1, wherein, In the step S4, training the WaveMambaFormer model includes: A log-cosh loss function is introduced as a loss function, and the calculation formula of the log-cosh loss is: ； wherein represents the predicted significant wave height of the sea surface, represents the reference significant wave height.

5. The method of claim 1, wherein, In the step S4, calculating feature importance and global explanation by using a SHAP interpreter to enhance the explainability of the WaveMambaFormer model includes: Prepare the trained WaveMambaFormer model and data, create a deep learning SHAP interpreter DeepExplainer, and then calculate SHAP values based on test data, and perform feature importance analysis and global explanation by using a visualization method; wherein the visualization method includes: a summary plot, a waterfall plot, a dependence plot, and a heat map plot.

6. The method of claim 1, wherein, In the step S5, the inversion performance of the WaveMambaFormer model is evaluated by using ERA5, WaveWatchIII, and Jason-3 significant wave height data as reference data, respectively, including: Root mean square error (RMSE), bias, mean absolute percentage error (MAPE), and Pearson correlation coefficient (R) are used as indicators to evaluate the performance of the model, and the related calculation formula is: ； ； ； ； wherein is the number of data samples, and are the model-estimated significant wave height and the significant wave height obtained from the reference data, respectively, and are the mean values of and respectively.

7. A system for constructing an ocean significant wave height inversion model in high sea states, said system being configured to implement the method of any one of claims 1 to 6, characterized in that, including: The acquisition module, the extraction module, the division module, the construction module, and the evaluation module; The acquisition module is configured to acquire modeling data, auxiliary data, and verification data, wherein the modeling data includes satellite-borne GNSS-R data and ERA5 significant wave height data; the auxiliary data includes ERA5 data, SMAP data, and OSCAR data; and the verification data includes ERA5, WaveWatch III, and Jason-3 significant wave height data. The extraction module is configured to extract satellite-borne GNSS-R feature variable parameters and auxiliary variable parameters, and perform spatiotemporal matching on the extracted variable parameters. The division module is configured to perform data quality control and data set division on the spatiotemporally matched variable parameters. The construction module is configured to construct and train the WaveMambaFormer model by using the divided data set, and calculate feature importance and global explanation by using a SHAP interpreter to enhance the explainability of the WaveMambaFormer model. The evaluation module is configured to evaluate the inversion performance of the WaveMambaFormer model by taking ERA5, WaveWatch III, and Jason-3 significant wave height data as reference data, respectively.

Images