Satellite-borne GNSS-R significant wave height estimation and deviation correction method

Abstract

The invention discloses a satellite-borne GNSS-R (Global Navigation Satellite System-Radio) significant wave height estimation and deviation correction method. According to the method, a satellite-borne GNSS-R significant wave height estimation model based on a BayesOpt-CNN-Trans-BiConvLSTM-BiLSTM and a channel space attention mechanism and a significant wave height deviation correction model based on the BayesOpt-BiLSTM are put forward, and a satellite-borne GNSS-R significant wave height deviation correction model based on the BayesOpt-BiLSTM is put forward. The construction method comprises the following steps of: 1, preparing input data of a significant wave height estimation model and inputting a significant wave height deviation correction model; second, a significant wave height estimation model is constructed, and a satellite-borne GNSS-R significant wave height estimation model based on a BayesOpt-CNN-Trans-BiConvLSTM-BiLSTM and a channel space attention mechanism is constructed; 3, constructing a significant wave height estimation deviation correction model, and constructing a Bayesian optimization-based bidirectional long-short-term memory network model for significant wave height estimation deviation correction; fourthly, effective wave height estimation and deviation correction are carried out, on a test data set, an effective wave height value is estimated through the model, an effective wave height difference value is predicted through the BiLSTM model based on Bayesian optimization, and an effective wave height value after deviation correction is obtained. According to the invention, efficient, high-precision and reliable satellite-borne GNSS-R significant wave height data can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of cross-application of artificial intelligence and global navigation satellite system remote sensing, and specifically relates to a method for spaceborne GNSS-R effective wave height estimation and deviation correction based on the fusion of intelligent optimization and deep learning. Background Art

[0002] In the field of ocean monitoring, accurate estimation of significant wave height (SWH) is a key link in marine engineering construction, navigation safety, and marine scientific research. Traditional methods for estimating significant wave height mostly rely on a single physical model or simple statistical methods. For example, the method based on spectral analysis estimates significant wave height by analyzing the wave frequency spectrum. However, under complex conditions such as extreme sea conditions and multi-directional irregular waves, it is limited by model assumptions and data noise, and the estimation accuracy is significantly reduced. Although the method based on buoy measurement can provide high-precision local data, it is difficult to meet the real-time and wide-area requirements of global ocean dynamic monitoring due to the high cost of buoy deployment, difficulty in maintenance, and limited spatial coverage.

[0003] With the rapid development of artificial intelligence (AI) technology, machine learning-based methods for estimating significant wave height have gradually become a research hotspot. These methods integrate multi-source data such as wind speed, wave reflection signals, and ocean environmental parameters to build prediction models, but they still face many technical bottlenecks: First, the setting of model hyperparameters relies heavily on manual experience and lacks an adaptive optimization mechanism, resulting in insufficient generalization of the model across different sea state datasets; second, the feature extraction methods for multimodal data (such as remote sensing images and time series signals) are limited, failing to fully exploit the spatiotemporal coupling information in the data; third, due to the complexity of the ocean environment and the existence of measurement errors, there are systematic deviations between the significant wave height estimates output by the model and the actual buoy measurements, and existing methods lack effective dynamic correction mechanisms.

[0004] In summary, the present invention proposes a method for spaceborne GNSS-R significant wave height estimation and bias correction based on the fusion of intelligent optimization and deep learning. This method automatically searches for the optimal hyperparameter combination of the deep learning model through the Bayesian optimization algorithm, overcoming the subjectivity and limitations of manual parameter adjustment. In the significant wave height estimation link, a dual-channel attention mechanism and a convolutional bidirectional sensor architecture are innovatively introduced. The former enhances the model's adaptive weight allocation ability for image features through channel attention and spatial attention modules, while the latter utilizes the synergy of convolutional neural networks and bidirectional recurrent structures to deeply mine the spatiotemporal features in the reflected signal and achieve efficient fusion of multimodal data. In the bias correction stage, a nonlinear bias correction model is constructed based on BiLSTM, breaking through the constraints of traditional linear correction methods. By learning the complex mapping relationship between the estimated value and the true value, the model output is dynamically adjusted, significantly improving the accuracy and reliability of significant wave height estimation, and providing a more practical technical solution for real-time monitoring of the marine environment. Summary of the Invention

[0005] In order to overcome the shortcomings of existing significant wave height estimation technologies, such as difficult hyperparameter optimization, insufficient feature mining, and poor bias correction, this paper proposes a space-borne GNSS-R significant wave height estimation and bias correction method based on the fusion of intelligent optimization and deep learning. By constructing a space-borne GNSS-R significant wave height estimation model with a BayesOpt-CNN-Trans-BiConvLSTM-BiLSTM and a channel spatial attention mechanism, combined with a Bayesian optimized BiLSTM bias correction model, high-precision estimation and dynamic correction of significant wave height can be achieved.

[0006] To achieve the above object, the present invention provides the following technical solution: a method for estimating and correcting the effective wave height of a space-borne GNSS-R, the method comprising the following steps:

[0007] Step S1: Preparation of input data for the significant wave height estimation model and input of the significant wave height deviation correction model

[0008] Data preparation: Download CYGNSS GNSS-R L1 observation data from the official website and preprocess and extract the bistatic scattering cross section (BRCS), effective scattering area (effective scattering area), power delay Doppler map (power_analog), delay Doppler map peak average (DDMA), normalized bistatic scattering cross section (NBRCS), leading edge slope (LES), trailing edge slope (TES), signal-to-noise ratio (SNR), equivalent isotropic radiated power (gps_eirp), receiver antenna gain (sp_rx_gain), incidence angle (sp_inc_angle), reflection point latitude and longitude (sp_lat, sp_lon), and range correction gain (RCG) observation variables; download ERA5 significant wave height and ERA5 wind speed from the ECMWF website; download buoy significant wave height data from the National Buoy Center (NDBC) website;

[0009] Step S2: Construction of significant wave height estimation model

[0010] A spaceborne GNSS-R significant wave height estimation model based on BayesOpt-CNN-Trans-BiConvLSTM-BiLSTM and a channel-space attention mechanism was constructed. This model uses Bayesian optimization (BayesOpt) to adjust model hyperparameters and integrates a key feature enhancement module based on a convolutional neural network (CNN) module, a Transformer module, a bidirectional convolutional long short-term memory (BiConvLSTM) network, a bidirectional long short-term memory (BiLSTM) network, and a channel-space attention mechanism.

[0011] Step S3: Construction of significant wave height estimation bias correction model

[0012] A bidirectional long short-term memory (BiLSTM) model based on Bayesian optimization is constructed to correct the bias of significant wave height estimation. The difference between the spaceborne GNSS-R significant wave height estimate and the buoy significant wave height after temporal and spatial matching is calculated. The difference sequence is used as the input target of the BiLSTM model. The longitude and latitude of the reflection point, signal-to-noise ratio (SNR), leading-edge slope (LES), normalized bistatic radar cross section (NBRCS), receiver antenna gain, range correction gain, incident angle, and azimuth are used as input variables of the BiLSTM model.

[0013] Step S4: significant wave height estimation and deviation correction

[0014] On the test dataset, the GNSS-R significant wave height value is estimated using the significant wave height estimation model. Meanwhile, the BiLSTM model based on Bayesian optimization is used to predict the significant wave height difference. The GNSS-R significant wave height value is corrected according to the significant wave height prediction difference. Finally, the bias-corrected spaceborne GNSS-R significant wave height estimation result is obtained.

[0015] Preferably, the input data preparation of the significant wave height estimation model and the input of the significant wave height deviation correction model described in step S1, data preparation, that is, downloading the CYGNSS GNSS-R L1 observation data from the official website, pre-processing and extracting the bistatic scattering cross section, that is, BRCS, effective scattering area, that is, effective scattering area, power delay Doppler map, that is, power_analog, delay Doppler map peak average, that is, DDMA, normalized bistatic scattering cross section, that is, NBRCS, leading edge slope, that is, LES, trailing edge slope, that is, TES, signal-to-noise ratio, that is, SNR, equivalent isotropic radiated power, that is, gps_eirp, receiver antenna gain, that is, sp_rx_gain, incident angle, that is, sp_inc_angle, reflection point latitude and longitude, that is, sp_lat, sp_lon and distance correction gain, that is, RCG observation variables; downloading ERA5 significant wave height and ERA5 wind speed from the ECMWF website; wherein, the buoy significant wave height data is downloaded from the National Buoy Center, that is, NDBC website.

[0016] Preferably, step S2 constructs an effective wave height estimation model, that is, constructs a satellite-borne GNSS-R effective wave height estimation model based on BayesOpt-CNN-Trans-BiConvLSTM-BiLSTM and channel space attention mechanism, the model including: using Bayesian optimization, i.e. Bayesian Optimization, BayesOpt to adjust the model hyperparameters, and integrating a key feature enhancement module based on a convolutional neural network, i.e. a CNN module, a Transformer module, a bidirectional convolutional long short-term memory network, i.e. BidirectionalConvLSTM, BiConvLSTM, a bidirectional long short-term memory network, i.e. Bidirectional LSTM, BiLSTM, and a channel space attention mechanism; wherein the CNN module is used to process three different image inputs, namely: a bistatic scattering cross section, i.e. BRCS, an effective scattering area, i.e. effective scattering The Transformer module uses the feature map processed by the CNN module as input. It processes sequence data through the multi-head self-attention mechanism (MultiHeadAttention) and the feedforward network (Dense), which can capture long-range dependencies. The ConvLSTM module uses the temporal and spatial dependent data processed by the Transformer module as input. The BiLSTM module mainly inputs the delay-Doppler peak average (DDMA), the normalized bistatic scattering cross section (NBRCS), the leading edge slope (LES), the trailing edge slope (TES), the signal-to-noise ratio (SNR), the equivalent isotropic radiated power (gps_eirp), the receiver antenna gain (sp_rx_gain), the incident angle (sp_inc_angle), the longitude and latitude of the reflection point (sp_lat and sp_lon), and the range correction gain (RCG) observation variable, which can capture long-term dependencies in time series.

[0017] Preferably, the effective wave height estimation bias correction model described in step S3 is constructed, and a bidirectional long short-term memory network based on Bayesian optimization, that is, a BiLSTM model is constructed for effective wave height estimation bias correction. The effective wave height estimation value of the space-borne GNSS-R matched with the effective wave height value of the buoy is subtracted to obtain the difference sequence between them and serve as the input target of the BiLSTM model. The longitude and latitude of the reflection point, signal-to-noise ratio (SNR), leading edge slope (LES), normalized bistatic radar cross section (NBRCS), receiver antenna gain, range correction gain, distance correction gain, incident angle and azimuth are used as input variables of the BiLSTM model.

[0018] Preferably, the significant wave height estimation and bias correction described in step 4, on the test data set, uses the significant wave height estimation model to estimate the GNSS-R significant wave height value, and simultaneously uses the BiLSTM model based on Bayesian optimization to predict the significant wave height difference, and corrects the GNSS-R significant wave height value according to the significant wave height prediction difference, and finally obtains the spaceborne GNSS-R significant wave height estimation result after bias correction.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] 1. The present invention discloses a method for estimating and correcting the effective wave height of satellite-borne GNSS-R (GNSS-R) based on the fusion of intelligent optimization and deep learning. The method mainly includes two models: a satellite-borne GNSS-R effective wave height estimation model based on BayesOpt-CNN-Trans-BiConvLSTM-BiLSTM and channel spatial attention mechanism, and a effective wave height estimation bias correction model based on a Bayesian optimized bidirectional long short-term memory network (BiLSTM) model.

[0021] 2. The method includes the following steps: preparing the input data of the significant wave height estimation model and the input data of the significant wave height bias correction model. The data includes downloading the CYGNSS GNSS-R L1 observation data from the official website, pre-processing and extracting the bistatic scattering cross section (BRCS), effective scattering area (effective scattering area), power delay Doppler map (power_analog), delay Doppler map peak average (DDMA), normalized bistatic scattering cross section (NBRCS), leading edge slope (LES), trailing edge slope (TES), signal-to-noise ratio (SNR), equivalent isotropic radiated power (gps_eirp), receiver antenna gain (sp_rx_gain), incidence angle (sp_inc_angle), reflection point latitude and longitude (sp_lat, sp_lon) and range correction gain (RCG) observation variables;

[0022] 3. Download ERA5 significant wave height and ERA5 wind speed from the ECMWF website; download buoy significant wave height data from the National Buoy Center (NDBC) website; construct a significant wave height estimation model and a significant wave height estimation bias correction model; finally, obtain the bias-corrected spaceborne GNSS-R significant wave height estimation result through significant wave height estimation and bias correction. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a flow chart of a method for spaceborne GNSS-R significant wave height estimation and deviation correction based on the fusion of intelligent optimization and deep learning provided in an embodiment of the present invention;

[0024] Figure 2 This is a structural diagram of a satellite-borne GNSS-R significant wave height estimation and bias correction model provided in an embodiment of the present invention;

[0025] Figure 3 A comparison result diagram between the effective wave height value of the satellite-borne GNSS-R and the effective wave height value of the buoy after bias correction estimated by the technical solution proposed in the embodiment of the present invention is provided;

[0026] Figure 4 A comparison diagram of the effective wave height values ​​of the satellite-borne GNSS-R and the ERA5 after bias correction estimated by the technical solution of the present invention provided in an embodiment of the present invention;

[0027] Figure 5 This is a comparison result diagram between the effective wave height value of the satellite-borne GNSS-R after bias correction estimated by the technical solution proposed in the present invention and the effective wave height value of WW3 provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0029] Example 1

[0030] This method combines BayesOpt-CNN-Trans-BiConvLSTM-BiLSTM with a channel-space attention mechanism to estimate significant wave height from spaceborne GNSS-R data. Furthermore, it addresses the challenge of processing spatially and temporally heterogeneous data in the inversion of significant wave height from spaceborne GNSS-R data by using a Bayesian optimization algorithm to achieve global optimal configuration of model parameters. This comprehensive approach effectively improves the adaptability and generalization performance of the inversion model when processing wave data with dynamic temporal and spatial variations, providing strong support for significant wave height inversion tasks in complex sea conditions.

[0031] In addition, the use of convolutional neural network (CNN) modules, Transformer modules, bidirectional convolutional long short-term memory (BiConvLSTM) networks, bidirectional long short-term memory (BiLSTM) networks, and channel-space attention in the construction of the significant wave height inversion model has the following main advantages:

[0032] (1) Advantages of feature extraction and information capture

[0033] 1) Comprehensive Feature Extraction: The CNN module possesses powerful local feature extraction capabilities, capturing subtle changes in local features such as signal strength and frequency in spaceborne GNSS-R data. These features are crucial for the initial representation of significant wave height. The Transformer module, through its self-attention mechanism, captures long-range dependencies in the data. In spaceborne GNSS-R data processing, it effectively correlates data from different time periods and spatial locations, providing a more comprehensive picture of the overall characteristics and changing trends of waves.

[0034] 2) Fusion of spatiotemporal information: BiConvLSTM combines the features of convolution operation and bidirectional LSTM. It can consider time series information while processing two-dimensional spatial data, and effectively model the dynamic changes of waves in space and time.

[0035] (2) Advantages of the attention mechanism

[0036] 1) Targeted Feature Enhancement: The channel-space attention mechanism adaptively learns the importance of each channel and spatial location. In spaceborne GNSS-R data, data from different channels and spatial locations contribute differently to significant wave height. This mechanism weights the feature maps to highlight key features closely related to significant wave height, suppress interference from irrelevant information, and improve the model's focus on useful information.

[0037] 2) Improved Model Adaptability: This attention mechanism enables the model to dynamically adjust its attention allocation based on varying input data, better adapting to complex and changing ocean environments and wave characteristics. Whether in calm or complex sea conditions, it can more accurately focus on features that are crucial for retrieving significant wave height, enhancing the model's robustness and adaptability.

[0038] (3) Model performance and optimization advantages

[0039] 1) Superior Model Performance: The integration of multiple modules enables the model to leverage the strengths of each module, improving the accuracy and stability of significant wave height inversion. CNN, Transformer, BiConvLSTM, and BiLSTM collaborate to process and analyze data from different perspectives, providing a more comprehensive and in-depth representation of features for significant wave height inversion, thereby improving the overall performance of the model.

[0040] 2) Effectiveness of Hyperparameter Optimization: Combining Bayesian optimization with model hyperparameter tuning allows for efficient search for optimal hyperparameter combinations within a vast hyperparameter space. This enables precise model optimization based on diverse datasets and task requirements, further improving the performance of significant wave height inversion and reducing errors caused by inappropriate hyperparameter selection.

[0041] In order to verify the feasibility and reliability of the spaceborne GNSS-R effective wave height estimation and bias correction method based on the fusion of intelligent optimization and deep learning in the present invention, the CYGNSS GNSS-R L1 observation data were downloaded from the relevant public website, and the bistatic scattering cross section (BRCS), effective scattering area, power delay Doppler map (power_analog), delay Doppler map peak average (DDMA), normalized bistatic scattering cross section (NBRCS), leading edge slope (LES), trailing edge slope (TES), signal-to-noise ratio (SNR), equivalent isotropic radiated power (gps_eirp), receiver antenna gain (sp_rx_gain), incidence angle (sp_inc_angle), reflection point latitude and longitude (sp_lat, sp_lon) and range correction gain (RCG) observation variables, ERA5 significant wave height, ERA5 wind speed, and buoy significant wave height data were extracted by preprocessing. A method for estimating and correcting the effective wave height of satellite-borne GNSS-R based on the fusion of intelligent optimization and deep learning. The technical solution implementation process is as follows: Figure 1 As shown, the following steps are included:

[0042] Step S1, data preparation and preprocessing: Download CYGNSS GNSS-R L1 observation data from the official website and extract the bistatic scattering cross section (BRCS), effective scattering area (ESSA), power delay-Doppler pattern (power_analog), delay-Doppler pattern peak-to-average (DDMA), normalized bistatic scattering cross section (NBRCS), leading edge slope (LES), trailing edge slope (TES), signal-to-noise ratio (SNR), equivalent isotropic radiated power (gps_eirp), receiver antenna gain (sp_rx_gain), incidence angle (sp_inc_angle), reflection point latitude and longitude (sp_lat, sp_lon), and range correction gain (RCG) observation variables. Download ERA5 significant wave height and ERA5 wind speed from the ECMWF website. Download buoy significant wave height data from the National Navigation Buoy Center (NDBC) website.

[0043] Step S2, effective wave height estimation model construction, a spaceborne GNSS-R effective wave height estimation model based on BayesOpt-CNN-Trans-BiConvLSTM-BiLSTM and channel spatial attention mechanism is constructed, which includes: using Bayesian optimization (Bayesian Optimization, BayesOpt) to adjust the model hyperparameters, and fusing the key feature enhancement module based on convolutional neural network (CNN) module, Transformer module, bidirectional convolution long short-term memory network (Bidirectional ConvLSTM, BiConvLSTM), bidirectional long short-term memory network (Bidirectional LSTM, BiLSTM) and channel spatial attention mechanism;

[0044] Step S3, effective wave height estimation bias correction model construction, a bidirectional long short-term memory network (BiLSTM) model based on Bayesian optimization is constructed for effective wave height estimation bias correction, the spaceborne GNSS-R effective wave height estimation value after spatio-temporal matching is subtracted from the buoy effective wave height value, to obtain the difference value sequence between them and as the input target of the BiLSTM model, the reflector longitude and latitude, signal-to-noise ratio (SNR), leading edge slope (LES), normalized bistatic radar cross section (NBRCS), receiver antenna gain, distance correction gain, incident angle and azimuth as the input variables of the BiLSTM model;

[0045] Step S4, effective wave height estimation and bias correction, on the test data set, the GNSS-R effective wave height value is estimated by using the effective wave height estimation model, and the effective wave height difference value is predicted by using the BiLSTM model based on Bayesian optimization, the GNSS-R effective wave height value is corrected according to the effective wave height prediction difference value, and finally the spaceborne GNSS-R effective wave height estimation result after bias correction is obtained.

[0046] As an embodiment of the present embodiment, step S1 includes the following substeps:

[0047] Bistatic scattering cross section (BRCS), effective scattering area, power delay-Doppler pattern, delay-Doppler pattern peak average, normalized bistatic scattering cross section (NBRCS), leading edge slope (LES), trailing edge slope (TES), signal-to-noise ratio (SNR), equivalent isotropically radiated power (gps_eirp), receiver antenna gain (sp_rx_gain), angle of incidence (sp_inc_angle), reflection point latitude and longitude (sp_lat, sp_lon), and range correction gain (RCG) observations were extracted from the CYGNSS GNSS-R L1 observation data. ERA5 significant wave height and ERA5 wind speed were downloaded from the ECMWF website. The buoy significant wave height data were downloaded from the National Navigation Buoy Center (NDBC) website. The GNSS-R data acquisition time was used to align the GNSS-R data, ERA5 data, and NDBC data using a linear interpolation algorithm. Then, the longitude and latitude information of the mirror reflection point were used to align the GNSS-R data, ERA5 data, and NDBC data using a bilinear interpolation algorithm.

[0048] To ensure data quality, data quality control is performed in the following ways:

[0049] (1) Observation validity criteria: All observations must be positive values ​​within a reasonable physical range. For ERA5 and NDBC buoy data, if invalid identifiers (such as special fill values ​​in ERA5 data and error codes in NDBC data) or NaN values ​​appear, they must be discarded. For variables extracted from CYGNSS, when the signal-to-noise ratio (SNR) is less than 0 dB and the incident angle (sp_inc_angle) is greater than 60°, the corresponding data record is considered invalid and discarded.

[0050] (2) Spatial validity determination: For CYGNSS data, the longitude and latitude of the reflection point (sp_lat, sp_lon) are used to perform spatial matching with the global ocean boundary vector data, and only the data with the mirror reflection point located in the ocean area are retained; for ERA5 data, it is necessary to ensure that the data coverage area is consistent with the research target sea area, and exclude land areas or invalid grid point data; for NDBC buoy data, it is necessary to verify the consistency of the buoy position coordinates with the actual ocean monitoring points, and exclude abnormal data caused by positioning errors.

[0051] (3) Spatiotemporal matching consistency: For CYGNSS, ERA5, and NDBC buoy data, the time interval is less than 15 minutes and the spatial space is less than 25 km.

[0052] As an implementation of this embodiment, step S2 includes the following sub-steps:

[0053] Step S2.1, the effective wave height estimation model is constructed, first, the features of the image are extracted through the multi-layer convolutional layer of the CNN, and data normalization is performed by using the batch normalization layer, the shape of the feature map is adjusted by using the maximum pooling layer, and the processed feature map is obtained. The image features processed by convolution are expressed in a new tensor through splicing operation, and then the channel attention mechanism and the spatial attention mechanism are used to enhance the features of the new tensor obtained by splicing.

[0054] Step S2.2, the input of the BiConvLSTM module is the data with space-time dependence processed by the Transformer module. The ConvLSTM2D is a two-dimensional version of the convolutional long short-term memory network (Convol utional Long Short-Term Memory), which combines the characteristics of convolutional neural network (CNN) and long short-term memory network (LSTM). When processing sequence data, it can capture spatial and temporal features. Padding is performed before convolution operation, and then sliding convolution is performed on the spatial dimension of the input data by using the convolution kernel in the ConvLSTM2D, so as to extract spatial features. The LSTM unit in the ConvLSTM2D will update the unit state and hidden state according to the current input and the hidden state at the last time through a series of gating mechanisms (input gate, forget gate and output gate), so as to capture time sequence information. The hyperbolic tangent activation function tanh is applied to the output to map the output value to the interval (-1, 1). The sigmoid activation function is applied to the output of the loop connection to map the output value to the interval (0, 1), which is used to control the gating mechanism.

[0055] Step S2.3, for the feature sequence, i.e. bistatic radar cross section (BRCS), effective scattering area, power delay Doppler plot (power_analog), delay Doppler plot mean average (DDMA), normalized bistatic radar cross section (NBRCS), leading edge slope (LES), trailing edge slope (TES), signal-to-noise ratio (SNR), equivalent isotropically 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) observation variables. The BiLSTM module is used for processing, which can capture long-term dependencies in time series.

[0056] Step S2.4, the outputs of different modules obtained in steps S2.2 and S2.3 are merged to construct a fully connected layer network and add Dropout regularization.

[0057] As an embodiment of the present embodiment, step S3 includes the following sub-steps:

[0058] Step S3.1, obtain the spaceborne GNSS-R significant wave height estimation data, buoy significant wave height value data, and perform space-time matching on the spaceborne GNSS-R significant wave height estimation and buoy significant wave height value. Specifically, according to the time stamp and geographic location information, the spaceborne GNSS-R significant wave height estimation and buoy significant wave height value data pairs at similar time (time difference 15 minutes) and similar geographic location (distance difference 25 km) are screened out.

[0059] Step S3.2, difference between the spaceborne GNSS-R significant wave height estimation and the buoy significant wave height value after space-time matching to obtain the difference sequence therebetween. The spaceborne GNSS-R significant wave height estimation is H sat , and the buoy significant wave height value is H buoy , and the difference sequence is taken as the input target of the BiLSTM model.

[0060] D = H sat -H buoy

[0061] Step S3.3, data preprocessing, remove missing values, outliers in the difference sequence, reflector longitude and latitude, signal-to-noise ratio (SNR), leading edge slope (LES), normalized bistatic radar cross section (NBRCS), receiver antenna gain, range correction gain, incidence angle and azimuth angle data. Normalize the difference sequence, reflector longitude and latitude, signal-to-noise ratio (SNR), leading edge slope (LES), normalized bistatic radar cross section (NBRCS), receiver antenna gain, range correction gain, incidence angle and azimuth angle data. Divide the processed data into training set, validation set and test set. According to the proportion of 70%, 15% and 15%, the training set is used for model training, the validation set is used for adjusting the hyperparameters of the model, and the test set is used for evaluating the final performance of the model.

[0062] Step S3.4, construct a bidirectional long short-term memory network based on Bayesian optimization (BayesOpt-BiLSTM) model. The input layer takes the reflector longitude and latitude, signal-to-noise ratio (SNR), leading edge slope (LES), normalized bistatic radar cross section (NBRCS), receiver antenna gain, range correction gain, incidence angle and azimuth angle as input data, and the model estimates the difference sequence of the significant wave height and the buoy significant wave height as input target. Define the Bayesian optimizer and perform Bayesian optimization search, take the mean square error on the validation set as the objective function, and use the best hyperparameters to construct the model.

[0063] Wherein, the mean square error refers to the average value of the square of the difference between the predicted value and the true value. Meanwhile, the number of validation set samples is m, and the true bias value y val of the model is predicted

[0064]

[0065] The BiLSTM model is trained using the optimized hyperparameters, the training set data is input into the model, the model parameters are updated through the back propagation algorithm to minimize the loss function, and the test set data is used to evaluate the trained model. The mean square error, mean absolute error and other indicators on the test set are calculated to evaluate the performance of the model.

[0066] As an embodiment of the present embodiment, step S4 includes the following sub-steps:

[0067] Step S4.1, on the test data set, the spaceborne GNSS-R significant wave height estimation model fused with BayesOpt-CNN-Trans-BiConvLSTM-BiLSTM and channel space attention mechanism is used to calculate the significant wave height value, and the preliminary estimated value is obtained. Then the reflector longitude and latitude, signal-to-noise ratio (SNR), leading edge slope (LES), normalized bistatic radar cross section (NBRCS), receiver antenna gain, distance correction gain, incident angle and azimuth angle in the test data are input into the BiLSTM model based on Bayesian optimization to predict the significant wave height difference value.

[0068] Step S4.2, according to the predicted significant wave height difference value, the spaceborne GNSS-R significant wave height preliminary estimation value is corrected to obtain the corrected significant wave height estimation value; the error index between the corrected spaceborne GNSS-R significant wave height estimation result and the true value (buoy, ERA5 and WW3 significant wave height value) is calculated to evaluate the effect of bias correction. The structure of the spaceborne GNSS-R significant wave height estimation and bias correction model proposed in the present application is shown in Figure 2 .

[0069] Figure 3-Figure 5 The comparison results between the bias-corrected spaceborne GNSS-R significant wave height value estimated by the technical solution proposed in the present application and the buoy, ERA5 and WW3 significant wave height values are given. From Figure 3-Figure 5 it can be seen that: when the buoy, ERA5 and WW3 data are taken as reference values respectively, the precision of the estimated significant wave height after bias correction is increased by 38.14%, 32.24% and 16.43% respectively compared with that before bias correction, and the correlation coefficient is also significantly improved. The comparison and analysis of the verification experiment fully prove that the spaceborne GNSS-R significant wave height estimation and bias correction method based on intelligent optimization and deep learning fusion proposed in the present application can significantly improve the prediction precision of ocean significant wave height.

[0070] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0071] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for estimating and correcting the significant wave height of a spaceborne GNSS-R, characterized by: The method comprises the following steps: Step S1, preparation of input data for the significant wave height estimation model and the significant wave height bias correction model: download CYGNSS GNSS-R L1 observation data from the official website, pre-process and extract the bistatic scattering cross section (BRCS), effective scattering area (effective scattering area), power delay Doppler map (power_analog), delay Doppler map peak average (DDMA), normalized bistatic scattering cross section (NBRCS), leading edge slope (LES), trailing edge slope (TES), signal-to-noise ratio (SNR), equivalent isotropic radiated power (gps_eirp), receiver antenna gain (sp_rx_gain), incident angle (sp_inc_angle), reflection point latitude and longitude (sp_lat, sp_lon), and range correction gain (RCG) observation variables; download ERA5 significant wave height and ERA5 wind speed from the ECMWF website; download buoy significant wave height data from the National Buoy Center (NDBC) website; Step S2: Construction of significant wave height estimation model A spaceborne GNSS-R significant wave height estimation model based on BayesOpt-CNN-Trans-BiConvLSTM-BiLSTM and a channel-space attention mechanism was constructed. This model uses Bayesian optimization (BayesOpt) to adjust model hyperparameters and integrates a key feature enhancement module based on a convolutional neural network (CNN) module, a Transformer module, a bidirectional convolutional long short-term memory (BiConvLSTM) network, a bidirectional long short-term memory (BiLSTM) network, and a channel-space attention mechanism. Step S3: Construction of significant wave height estimation bias correction model A Bayesian optimization-based bidirectional long short-term memory (BiLSTM) model was constructed to correct the bias in significant wave height estimation. The difference between the spaceborne GNSS-R significant wave height estimate and the buoy significant wave height, which was time-space matched, was calculated. The resulting difference sequence was used as the input target for the BiLSTM model. The longitude and latitude of the reflection point, the signal-to-noise ratio (SNR), the leading-edge slope (LES), the normalized bistatic radar cross section (NBRCS), the receiver antenna gain, the range correction gain, the angle of incidence, and the azimuth were used as the input variables of the BiLSTM model. Step S4: significant wave height estimation and deviation correction On the test dataset, the GNSS-R significant wave height value is estimated using the significant wave height estimation model. Meanwhile, the BiLSTM model based on Bayesian optimization is used to predict the significant wave height difference. The GNSS-R significant wave height value is corrected according to the significant wave height prediction difference. Finally, the bias-corrected spaceborne GNSS-R significant wave height estimation result is obtained.

2. The method for estimating and correcting the significant wave height of a spaceborne GNSS-R according to claim 1, wherein: In step S1, the input data of the significant wave height estimation model and the significant wave height deviation correction model are prepared. Data preparation includes downloading CYGNSS GNSS-R L1 observation data from the official website, pre-processing and extracting the bistatic scattering cross section, i.e. BRCS, effective scattering area, i.e. effective scattering area, power delay Doppler map, i.e. power_analog, delay Doppler map peak average, i.e. DDMA, normalized bistatic scattering cross section, i.e. NBRCS, leading edge slope, i.e. LES, trailing edge slope, i.e. TES, signal-to-noise ratio, i.e. SNR, equivalent isotropic radiated power, i.e. gps_eirp, receiver antenna gain, i.e. sp_rx_gain, incident angle, i.e. sp_inc_angle, reflection point latitude and longitude, i.e. sp_lat, sp_lon and range correction gain, i.e. RCG observation variables; downloading ERA5 significant wave height and ERA5 wind speed from the ECMWF website; wherein, the buoy significant wave height data is downloaded from the National Buoy Center, i.e. NDBC website.

3. The method for estimating and correcting the significant wave height of a spaceborne GNSS-R according to claim 1, wherein: Step S2: constructing a significant wave height estimation model, i.e., constructing a satellite-borne GNSS-R significant wave height estimation model based on BayesOpt-CNN-Trans-BiConvLSTM-BiLSTM and channel space attention mechanism. The model includes: using Bayesian optimization, i.e., BayesOpt, to adjust the model hyperparameters, and integrating a key feature enhancement module based on a convolutional neural network, i.e., a CNN module, a Transformer module, a bidirectional convolutional long short-term memory network, i.e., a Bidirectional ConvLSTM, a BiConvLSTM, a bidirectional long short-term memory network, i.e., a Bidirectional LSTM, a BiLSTM, and a channel space attention mechanism; wherein the CNN module is used to process three different image inputs, namely, a bistatic scattering cross section, i.e., a BRCS, an effective scattering area, i.e., an effective scattering area. The Transformer module uses the feature map processed by the CNN module as input. It processes sequence data through the multi-head self-attention mechanism (MultiHeadAttention) and the feedforward network (Dense), which can capture long-range dependencies. The ConvLSTM module uses the temporal and spatial dependent data processed by the Transformer module as input. The BiLSTM module mainly inputs the delay-Doppler peak average (DDMA), the normalized bistatic scattering cross section (NBRCS), the leading edge slope (LES), the trailing edge slope (TES), the signal-to-noise ratio (SNR), the equivalent isotropic radiated power (gps_eirp), the receiver antenna gain (sp_rx_gain), the incident angle (sp_inc_angle), the longitude and latitude of the reflection point (sp_lat and sp_lon), and the range correction gain (RCG) observation variable, which can capture long-term dependencies in time series.

4. The method for estimating and correcting the significant wave height of a spaceborne GNSS-R according to claim 1, wherein: In step S3, a significant wave height estimation bias correction model is constructed, and a bidirectional long short-term memory network based on Bayesian optimization, i.e., a BiLSTM model, is constructed for significant wave height estimation bias correction. The spaceborne GNSS-R significant wave height estimation value after time and space matching is subtracted from the buoy significant wave height value to obtain a difference sequence between them and serve as the input target of the BiLSTM model. The longitude and latitude of the reflection point, the signal-to-noise ratio, i.e., SNR, the leading edge slope, i.e., LES, the normalized bistatic radar cross section, i.e., NBRCS, the receiver antenna gain, the range correction gain, the incident angle, and the azimuth are used as the input variables of the BiLSTM model.

5. The method for estimating and correcting the significant wave height of a spaceborne GNSS-R according to claim 1, wherein: In step 4, the significant wave height estimation and bias correction are performed. On the test dataset, the GNSS-R significant wave height value is estimated using the significant wave height estimation model. Meanwhile, the significant wave height difference is predicted using the BiLSTM model based on Bayesian optimization. The GNSS-R significant wave height value is corrected based on the significant wave height prediction difference, and finally the bias-corrected spaceborne GNSS-R significant wave height estimation result is obtained.

Images