Spaceborne GNSS-R wave energy flux estimation model construction method

Abstract

The invention provides a satellite-borne GNSS-R wave energy flux estimation model construction method. Deep learning and physically guided machine learning are innovatively combined. According to the method, multi-scale feature extraction, a channel and a space attention mechanism are fused through a Dilatation ResBiLSTM-AttnNet model, a significant wave height (SWH) is estimated, and a self-adaptive CDF matching method is introduced to carry out deviation correction; wave periods are estimated using a physically guided based bagged tree (BT) model. And calculating the wave energy flux by combining the corrected SWH and the estimated wave period with a wave energy formula. The model carries out space-time matching and quality control preprocessing based on GNSS-R data, ERA5 data and WW3 data of a wind cloud series satellite (FY-3E / 3F / 3G). Through comparison and evaluation with various existing models, the method has higher precision in the aspects of SWH, wave period and wave energy estimation, the estimation of the satellite-borne GNSS-R technology on the wave energy is realized for the first time, and a brand new efficient solution is provided for ocean wave parameter monitoring and wave energy development.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and GNSS reflected signal wave energy flux estimation, and particularly to a method for constructing an on-orbit GNSS-R wave energy flux estimation model. Background Art

[0002] Ocean waves are an important part of the marine dynamic environment. The study of wave parameters is not only crucial for ocean engineering and shipping safety, but also has a profound impact on multiple fields such as climate change, environmental protection, and ocean resource development. Therefore, the estimation of significant wave height (SWH), wave period, and wave energy has important value.

[0003] Global Navigation Satellite System Reflectometry (GNSS-R) is a new remote sensing technology. By receiving GNSS signals reflected from the ground and analyzing their characteristics, various physical parameters of the ground or water surface can be inferred. The on-orbit GNSS-R technology has unique advantages such as low observation cost, wide coverage, short revisit period, and the ability to conduct all-weather observations. Currently, this technology has been widely applied in various fields such as land, ocean, and atmosphere. In addition, existing research has shown that there is great potential in using on-orbit GNSS-R technology to retrieve wave parameters.

[0004] For the estimation of SWH, traditional methods commonly include buoy systems and satellite radar altimeters. However, buoy systems and satellite radar altimeters are restricted by the monitoring range or time resolution, making it difficult to quickly and effectively measure the globe. Spaceborne GNSS-R technology provides another effective method for SWH estimation. The methods for estimating SWH based on spaceborne GNSS-R mainly include empirical modeling methods. Although this method is helpful for SWH estimation, empirical models are usually established based on specific datasets or specific environmental conditions, with limited generalization ability. Moreover, SWH is affected by various complex factors on the ocean, and such complex environments often involve nonlinear interactions, which are difficult to describe with simple mathematical expressions. In contrast, machine learning and / or deep learning methods have powerful capabilities to handle such nonlinear relationships. In recent years, researchers have turned their attention to using machine learning and deep learning methods (e.g., neural network (NN), bagging tree (BT), deep convolutional neural network (DCNN), Transformer, etc.) to invert SWH. However, the inversion accuracy of these methods still has certain limitations. Most of them do not consider the attention mechanism and perform bias correction on the estimation results to eliminate systematic errors. Moreover, these studies are mainly based on the US CYGNSS satellite, which only receives GPS satellite signals. Compared with the CYGNSS satellite, the Chinese Fengyun-3 series satellites (i.e., FY-3E / 3F / 3G) can receive reflected signals from the three global navigation systems of Beidou, GPS, and Galileo simultaneously, providing a valuable opportunity to improve the inversion accuracy by fusing multi-system navigation satellite signals. However, up to now, very few studies have been carried out on inverting SWH using the Beidou / GNSS reflected signals of the Chinese Fengyun-3 series satellites (only 1-2 papers have carried out preliminary research work).

[0005] Wave period is another important parameter in ocean waves. Generally, satellite altimeters can directly provide the significant wave height (SWH) and wind speed of the ocean, and the wave period can be derived from the SWH and wind speed. However, the revisit period of satellite altimeters is difficult to meet the needs of dynamic changes, and the altimeter signal will attenuate during heavy rain. At present, there is little attention paid to the research on retrieving wave period using spaceborne GNSS-R technology. There is only one piece of literature (as far as the inventor knows) that has demonstrated the great potential of estimating wave period using spaceborne GNSS-R. However, the estimation method is mainly based on physical models. However, there are many factors affecting the wave period in complex sea conditions, such as wind speed, wind direction, water depth changes, and terrain differences in different sea areas. This makes it difficult for simple empirical formulas to comprehensively and accurately describe the variation law of the wave period. When facing complex and changeable sea conditions, the adaptability and accuracy of the model face challenges. Machine learning technology undoubtedly provides a novel and efficient solution for wave period inversion. Therefore, deeply exploring the application of spaceborne GNSS-R technology in the field of wave period inversion is of great significance for promoting the development of this field. Further research can not only improve the accuracy and reliability of wave period estimation but also provide strong technical support for ocean dynamics research and ocean environmental monitoring.

[0006] Up to now, the research on wave energy estimation using GNSS technology, especially BDS / GNSS reflectometry, is still blank. In view of this, the present invention proposes a method for constructing a spaceborne GNSS-R wave energy flux estimation model. Summary of the Invention

[0007] Aiming at the current situation that the research on wave energy estimation using spaceborne GNSS-R technology is still blank, the present invention proposes a method for constructing a spaceborne GNSS-R wave energy flux estimation model. First, for the estimation of SWH, a new deep learning model (i.e., DilatedResBiLSTM-AttnNet) is proposed. In this model, channel and spatial attention mechanisms are added to enhance important features to improve the inversion performance of the model. In addition, to improve the inversion accuracy and eliminate systematic errors, an adaptive CDF matching method is introduced to correct the bias of SWH. Second, for the estimation of wave period, since the empirical model has challenges in multi-parameter input, the accuracy of the wave period is limited, which will lead to large deviations in calculating wave energy. To solve this problem, a physically guided bagged tree (BT) model is proposed to invert the wave period. Finally, the SWH estimated by the DilatedResBiLSTM-AttnNet model is corrected by the adaptive CDF matching method to obtain the bias-corrected SWH, which is combined with the wave period estimated by the physically guided BT model to estimate wave energy. The present invention successfully realizes the estimation of wave energy using spaceborne GNSS-R technology, and for the first time demonstrates the great potential of estimating wave energy using spaceborne GNSS-R.

[0008] The technical solution adopted by the present invention is as follows:

[0009] A method for constructing a spaceborne GNSS-R wave energy flux estimation model, comprising the following steps:

[0010] Step S1, data preparation and preprocessing: First, download FY-3E / 3F / 3G GNSS-R data of Fengyun series satellites from the official website; ERA5 data (wind speed, SWH, wave period); WW3 data (SWH, wave period). Then, perform data feature extraction and data quality control preprocessing;

[0011] Step S2, model construction: SWH is estimated by the constructed deep learning hybrid model (DilatedResBiLSTM-AttnNet) and the bias correction of SWH is performed using the adaptive CDF matching method. DilatedResBiLSTM-AttnNet includes: a multi-scale feature extraction module based on the residual network (ResNet) and dilated convolution, a fully connected network module that fuses multi-feature parameters, a feature enhancement module based on channel attention mechanism and spatial attention mechanism, and a feature relationship reasoning module based on BiLSTM; the wave period is estimated by the physics-guided bagging trees ensemble learning algorithm (Bagging Trees, BT);

[0012] Step S3, wave energy flux estimation: The estimated value of SWH after bias correction is used as the input of the wave period model to obtain the estimated value of the wave period, and together with the corrected SWH, it is used as the input of the wave energy formula to estimate the wave energy;

[0013] Step S4, model evaluation: Taking the WW3 data as a reference, the innovative model is evaluated using the root mean square error (RMSE), bias (Bias), correlation coefficient (CC), and mean absolute percentage error (MAPE) as accuracy indicators.

[0014] Preferably, for the data preparation and preprocessing in step S1, download the GNSS-R data of the Fengyun series satellite constellation (i.e., FY-3E / 3F / 3G) from the relevant website; the ERA5 wind speed, ERA5 significant wave height, and ERA5 wave period provided by the European Centre for Medium-Range Weather Forecasts (ECMWF); the WW3SWH and WW3 wave period provided by the third-generation wave height product (i.e., WaveWatch III, WW3).

[0015] Data preprocessing includes: extracting the longitude and latitude of the specular reflection point, the data acquisition time, the normalized bistatic radar cross section (NBRCS), and the leading edge slope (LES) observation values from the FY-3E / 3F / 3G GNSS-R observation data, and spatially and temporally matching the GNSS-R data, ERA5 data, and WW3 data respectively using the longitude and latitude of the specular reflection point and the data acquisition time; extracting the time, grid longitude and latitude, U wind component, V wind component, combined significant wave height of swell and wind waves, and peak wave period variables from the ERA5 data, and calculating the synthetic wind speed value at 10 m above the sea surface; extracting the time, grid longitude and latitude, and peak wave period variables from the WW3 data; using the GNSS-R data acquisition time to match and align the GNSS-R data, ERA5, and WW3 data in time using a linear interpolation algorithm, and then using the longitude and latitude information of the specular reflection point to match and align the GNSS-R data, ERA5, and WW3 data spatially using a bilinear interpolation algorithm.

[0016] In addition, to ensure the quality of the data, data quality control is carried out in the following manner:

[0017] (1) To ensure that all the data used is valid, it is necessary to check the "Bad_File_Flag" attribute of each data file. A "Bad_File_Flag" of 0 indicates that there is at least one valid DDM data point in the file, while a "Bad_File_Flag" of 1 indicates that there is no valid DDM data point in the file. Select the data files with a "Bad_File_Flag" of 0 for subsequent analysis;

[0018] (2) The observed values must be positive, and when they are Nan values, they need to be discarded;

[0019] (3) Select the points where the specular reflection point is on the ocean.

[0020] Preferably, the model construction described in step S2 includes the following contents:

[0021] SWH is estimated by the constructed deep learning hybrid model (DilatedResBiLSTM-AttnNet) and the bias correction of SWH is carried out using the adaptive CDF matching method. DilatedResBiLSTM-AttnNet includes: a multi-scale feature extraction module based on a residual network (ResNet) and dilated convolution, a fully connected network module that fuses multi-feature parameters, a feature enhancement module based on a channel attention mechanism and a spatial attention mechanism, and a feature relationship reasoning module based on BiLSTM. Among them, the principle and method of the adaptive CDF matching method for bias correction of SWH are as follows:

[0022] The CDF matching method is a method that basically eliminates systematic biases by applying certain corrections to the output of the model so that the cumulative distribution function of the corrected model output is the same as the reference value

[30] . The specific method is as follows: In the training set, first, the inversed SWH of the model and the reference SWH are arranged in ascending order, and then the difference sequence D between the reference value and the inversed SWH is calculated. After applying the difference correction to the inversed SWH, the distribution functions of the inversed SWH and the reference SWH sequences are the same. The difference sequence is fitted by a polynomial of order 0 to 10:

[0023]

[0024] where P n is a polynomial of order n, is the sequence of inversed SWH.

[0025] Define RMSE as the regression loss. In this polynomial, the order of the polynomial that minimizes the RMSE is adaptively found through the principle of nonlinear least squares. The corrected inversed result is expressed as:

[0026]

[0027] In addition, the wave period is estimated by a physics-guided bagging trees ensemble learning algorithm (BT). Among them, the physics-guided constraint equation is expressed as:

[0028]

[0029] where X is used to describe the relationship between spaceborne GNSS-R observations (i.e., NBRCS (or LES) observations), SWH, and wave period.

[0030] Preferably, for the wave energy flux estimation in step S3, the estimated value of SWH after bias correction is used as the input of the wave period model to obtain the estimated value of the wave period, and together with the corrected SWH, it is used as the input of the wave energy formula to estimate the wave energy. The mathematical expression of the wave energy formula is as follows:

[0031]

[0032] where H s is the SWH, T e is the wave period, ρ and g are the seawater density and gravitational acceleration respectively, ρ = 1025 kg / m3 , g = 9.81 m / s 2 .

[0033] Preferably, for the model evaluation described in step 4, with WW3 data as a reference, the innovative model evaluation is carried out using the root mean square error (RMSE), bias (Bias), correlation coefficient (CC), and mean absolute percentage error (MAPE) as accuracy indicators.

[0034] In terms of SWH estimation: The innovated DilatedResBiLSTM-AttnNet deep learning model is compared with seven other existing machine learning models or deep learning models (decision tree regression (DTR), Light Gradient Boosting Machine (LightGBM), Extreme Gradient Boosting (XGBoost), Bagged Trees (BT), Deep Convolutional Neural Network (DCNN), Convolutional-Bidirectional Long Short-Term Memory Network (CNN-BiLSTM), Transformer).

[0035] In terms of wave period estimation: Seven machine learning models, namely DTR, Extremely Randomized Trees, CatBoost, Lightbm, XGBoost, Random Forest (RF), and Support Vector Regression (SVR), are compared with the physically-guided bagged tree wave period model estimation method proposed in the present invention.

[0036] In terms of wave energy estimation: With WW3 as a reference, it is compared with spaceborne GNSS-R for wave energy estimation.

[0037] The present invention proposes a comprehensive estimation network for ocean significant wave height, wave period, and wave energy flux using spaceborne GNSS-R technology. This network innovatively combines GNSS reflection technology with deep learning / machine learning algorithms, breaking the limitations of traditional ocean wave monitoring methods and providing a new solution for the accurate estimation of ocean wave parameters and wave energy calculation. It has the following beneficial effects:

[0038] (1) Scientificity: The scientificity of the present invention is reflected in the rigor of the theoretical framework and the accuracy of the experimental design. The core of the research is based on the combination of ocean wave theory and GNSS reflection signals. By efficiently using multi-system GNSS signals and combining ERA5 wind speed data and WaveWatch III ocean wave model data, high-precision estimation of significant wave height and wave period is achieved. During the training process of the model, deep learning algorithms are used to extract complex features from large-scale data, ensuring the high accuracy of the estimation results. For the estimation of wave period, an ensemble learning algorithm is adopted to provide accurate wave period estimations under low wind speed and high wind speed conditions respectively. In addition, a bias correction method using the adaptive CDF matching method is used to correct the bias of the significant wave height to eliminate systematic errors, enabling this study to provide more accurate and reliable ocean wave parameter estimation results in variable environments such as high wind speed and complex sea conditions.

[0039] (2) Advancedness: GNSS-R technology, as an emerging ocean remote sensing technology, has the advantages of low cost, wide coverage, and high timeliness. It can provide more extensive and efficient data support than traditional satellite radar and buoy monitoring. In the process of wave parameter inversion, this invention innovatively combines GNSS reflection signals with deep learning / machine learning, using machine learning and deep learning to automatically learn the complex patterns of ocean waves, and uses spaceborne GNSS-R technology to estimate wave energy, thereby improving the ability to estimate ocean wave parameters.

[0040] (3) Uniqueness: The uniqueness of this invention lies in its combination of multi-system GNSS reflection signals and advanced algorithms to propose a new wave parameter inversion model. GNSS reflection technology has not been widely used in existing wave monitoring methods. This study uses GNSS reflection signals as a data source and innovatively combines deep learning / machine learning with physical models, providing new ideas for wave monitoring in complex sea conditions.

[0041] Compared with traditional buoy measurement and satellite remote sensing technology, spaceborne GNSS-R technology has the advantages of low cost, wide spatial coverage, and strong data timeliness. It can more efficiently monitor ocean waves over a large area and is particularly suitable for the real-time data requirements of wave monitoring systems. In addition, the adaptive CDF matching method has demonstrated unique advantages in correcting deviations in wave height and wave period estimation, further improving the stability and accuracy of the model. Finally, the present invention calculates high-precision wave energy through the corrected effective wave height and wave period data, providing accurate parameter support for the application of wave power generation technology. The accurate estimation of wave energy provides a solid data foundation for the development of marine energy, the design of wave power generation systems, and marine environmental monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings of the implementation cases.

[0043] Figure 1 This is a flow chart of a method for constructing a space-borne GNSS-R wave energy flux estimation model provided in an embodiment of the present invention.

[0044] Figure 2 This is the DilatedResBiLSTM-AttnNet model framework for SWH estimation provided in an embodiment of the present invention.

[0045] Figure 3 This is a scatter plot comparing the estimation performance of spaceborne GNSS-R SWH of different models provided in an embodiment of the present invention.

[0046] Figure 4The SWH estimation performance of different models under different sea conditions provided in the embodiments of the present invention.

[0047] Figure 5 The scatter plot of the wave period estimation performance comparison of different model spaceborne GNSS-R provided in the embodiments of the present invention.

[0048] Figure 6 The scatter plot (left), PDF (middle), and histogram (right) of the comparison between WW3 and GNSS-R wave energy provided in the embodiments of the present invention. Detailed implementation manners

[0049] The following will specifically elaborate on the present invention in combination with specific implementation manners and embodiments, and the advantages and various effects of the present invention will be presented more clearly therefrom. Those skilled in the art should understand that these specific implementation manners and embodiments are used to illustrate the present invention rather than limit the present invention.

[0050] Throughout the specification, unless otherwise specifically stated, the terms used herein should be understood as having the meanings commonly used in the art. Therefore, unless otherwise defined, all technical and scientific terms used herein have the same meanings as those generally understood by those skilled in the art to which the present invention belongs. In case of conflict, this specification shall prevail.

[0051] Unless otherwise specifically stated, various raw materials, reagents, instruments, and equipment used in the present invention can be obtained through market purchases or can be prepared by existing methods.

[0052] The present invention provides a method for constructing a spaceborne GNSS-R wave energy flux estimation model, which will be described in detail below with reference to the accompanying drawings:

[0053] Example 1

[0054] In order to verify the feasibility and reliability of the method for constructing the spaceborne GNSS-R wave energy flux estimation model of the present invention, FY-3E / 3F / 3G GNSS-R data, ERA5 data (wind speed, SWH, wave period), and WW3 data (SWH, wave period) were downloaded from relevant public websites. The data time was from January 2020 to December 2020. Among them, ERA5 significant wave height, ERA5 wave period, and ERA5 wave energy flux data (calculated using the wave energy formula by combining ERA5 significant wave height and wave period data), and WW3 SWH, WW3 wave period, and WW3 wave energy data (calculated using the wave energy formula by combining WW3 significant wave height and wave period data) were used as reference data, and ERA5 wind speed was used as auxiliary data for wave period estimation modeling. The time and space resolution information of the three types of data used in the present invention is shown in Table 1.

[0055] Table 1 Time and spatial resolution information of three types of data used in the present invention

[0056]

[0057] A method for constructing a spaceborne GNSS-R wave energy flux estimation model, and the implementation process of the technical solution is as shown in the appendix Figure 1 as follows, including the following steps:

[0058] Step S1, data preparation and preprocessing: First, download Fengyun GNSS-R data (FY-3E / 3F / 3G); ERA5 data (wind speed, SWH, wave period); WW3 data (SWH, wave period) from the official website. Then, perform data feature extraction and data quality control preprocessing;

[0059] Step S2, model construction: SWH is estimated by the constructed deep learning hybrid model (DilatedResBiLSTM-AttnNet) and the bias correction of SWH is performed using the adaptive CDF matching method. DilatedResBiLSTM-AttnNet includes: a multi-scale feature extraction module based on a residual network (ResNet) and dilated convolution, a fully connected network module that fuses multi-feature parameters, a feature enhancement module based on a channel attention mechanism and a spatial attention mechanism, and a feature relationship reasoning module based on BiLSTM; the wave period is estimated by a physics-guided bagging tree ensemble learning algorithm (Bagging Trees, BT);

[0060] Step S3, wave energy flux estimation: The estimated value of SWH after bias correction is used as the input of the wave period model to obtain the estimated value of the wave period, and together with the corrected SWH, it is used as the input of the wave energy formula to estimate the wave energy;

[0061] Step S4, model evaluation: Using WW3 data as a reference, the innovative model is evaluated with the root mean square error (RMSE), bias (Bias), correlation coefficient (CC), and mean absolute percentage error (MAPE) as accuracy indicators.

[0062] As an implementation manner of this embodiment, the data preparation and preprocessing in Step S1, that is, download Fengyun series satellite constellation GNSS-R data (i.e., FY-3E / 3F / 3G) from relevant websites; ERA5 wind speed, ERA5 significant wave height, ERA5 wave period provided by the European Centre for Medium-Range Weather Forecasts (ECMWF); WW3SWH and WW3 wave period provided by the third-generation wave height product (i.e., WaveWatch III, WW3).

[0063] Data preprocessing includes: extracting the longitude and latitude of the specular reflection point, data acquisition time, normalized bistatic radar cross-section (NBRCS), and leading edge slope (LES) observation values from FY-3E / 3F / 3G GNSS-R observation data, and performing spatio-temporal matching on GNSS-R data, ERA5 data, and WW3 data respectively using the longitude and latitude of the specular reflection point and the data acquisition time; extracting time, grid longitude and latitude, U wind component, V wind component, combined significant wave height of swell and wind waves, and peak wave period variables from ERA5 data, and calculating the synthetic wind speed value at 10 m above the sea surface; extracting time, grid longitude and latitude, and peak wave period variables from WW3 data; using the GNSS-R data acquisition time to perform time matching and alignment on GNSS-R data, ERA5, and WW3 data using the linear interpolation algorithm, and then using the longitude and latitude information of the specular reflection point to perform spatial matching and alignment on GNSS-R data, ERA5, and WW3 data using the bilinear interpolation algorithm.

[0064] In addition, to ensure the quality of the data, data quality control is carried out in the following way:

[0065] (1) To ensure that all the data used is valid, it is necessary to check the "Bad_File_Flag" attribute of each data file. A "Bad_File_Flag" of 0 indicates that there is at least one valid DDM data point in the file, while a "Bad_File_Flag" of 1 indicates that there is no valid DDM data point in the file. Select the data files with a "Bad_File_Flag" of 0 for subsequent analysis;

[0066] (2) The observed values must be positive, and when they are Nan values, they need to be discarded;

[0067] (3) Select the points where the specular reflection point is on the ocean.

[0068] As an implementation manner of this embodiment, the model construction described in step S2 includes the following contents:

[0069] SWH is estimated by the constructed deep learning hybrid model (DilatedResBiLSTM-AttnNet) and the bias correction of SWH is performed using the adaptive CDF matching method. DilatedResBiLSTM-AttnNet includes: a multi-scale feature extraction module based on the residual network (ResNet) and dilated convolution, a fully connected network module that fuses multi-feature parameters, a feature enhancement module based on the channel attention mechanism and the spatial attention mechanism, and a feature relationship reasoning module based on BiLSTM. Among them, the principle and method of the adaptive CDF matching method for bias correction of SWH are as follows:

[0070] The CDF matching method is a method that basically eliminates systematic bias by applying a certain correction to the output of the model so that the cumulative distribution function of the corrected model output is the same as the reference value

[30] . The specific method is as follows: in the training set, first, the inversed SWH of the model and the reference SWH are arranged in ascending order, and then the difference sequence D between the reference value and the inversed SWH is calculated. After the difference correction is performed on the inversed SWH, the distribution functions of the inversed SWH and the reference SWH sequences are the same. The difference sequence is fitted by a polynomial of order 0 to 10:

[0071]

[0072] where P n is a polynomial of order n, is the sequence of inversed SWH.

[0073] Define RMSE as the regression loss. In this polynomial, the order of the polynomial that minimizes the RMSE is adaptively found through the principle of nonlinear least squares. The corrected inversion result is expressed as:

[0074]

[0075] In addition, the wave period is estimated by a physics-guided bagging trees ensemble learning algorithm (Bagging Trees, BT).

[0076] where the physics-guided constraint equation is expressed as:

[0077]

[0078] where X is used to describe the relationship between the spaceborne GNSS-R observation values (i.e., NBRCS (or LES) observation values), SWH, and wave period.

[0079] Specifically, the above formula (3) is used as the input quantity of the physics-guided BT model to enhance the accuracy of wave period estimation. It should be noted that the input variable information of the DilatedResBiLSTM-AttnNet model and the physics-guided BT model is shown in Table 2:

[0080]

[0081] As an implementation manner of this embodiment, for the wave energy flux estimation in step S3, the wave period estimation value is obtained by using the SWH estimated value corrected by bias correction as the input of the wave period model, and together with the corrected SWH, it is used as the input of the wave energy formula to estimate the wave energy.

[0082] In oceanography, ocean waves can generally be regarded as a combination of different random wave conditions, involving many individual wave components with different directions, amplitudes, and frequencies. The wave energy flux is defined by the wave energy flux equation:

[0083]

[0084] where ρ and g are the seawater density and gravitational acceleration respectively, σ is the frequency, θ is the wave propagation direction, C g (σ, h) and E(σ, θ) are the wave group velocity and spectral energy density respectively.

[0085] The calculation formula for the wave group velocity is:

[0086]

[0087] where h is the wave period, L is the wavelength, T is the wave period, and the wave number k = 2π / L.

[0088] The wave energy estimation equation for SWH can be derived from formula (4), and the calculation expression is as follows:

[0089]

[0090] where H s is the SWH, T e is the wave period. Since is a dispersion equation.

[0091] Therefore, the wave energy cannot be directly calculated using this formula. The waves need to be divided into deep-water waves and shallow-water waves for calculation. The research object of this invention focuses on the wave energy under deep-water waves. Therefore, it can be obtained that under deep-water conditions (i.e., the water depth is greater than half of the wavelength h > L / 2, ), the calculation method for the wave energy J is:

[0092]

[0093] where H s is the SWH, T e is the wave period, ρ and g are the seawater density and gravitational acceleration respectively, ρ = 1025 kg / m 3 , g = 9.81 m / s 2 .

[0094] Appendix Figure 1 shows the flowchart of the method for constructing the spaceborne GNSS-R wave energy flux estimation model provided by this invention. Appendix Figure 2The DilatedResBiLSTM-AttnNet model framework for SWH estimation of the present invention is shown. The DilatedResBiLSTM-AttnNet hybrid model mainly consists of a multi-scale feature extraction module based on ResNet and dilated convolution, a fully connected network module for fusing multi-feature parameters, a feature enhancement module based on channel attention mechanism and spatial attention mechanism, and a feature relationship reasoning module based on BiLSTM. ResNet can better solve the degradation problem caused by the increase in the depth of the deep convolutional network. The multi-scale feature extraction module uses ResNet34 and dilated convolution to extract two-dimensional information in BRCS and effective scattering area. The fully connected network module for fusing multi-feature parameters is used to process the spaceborne GNSS-R feature observations and variable parameters. After merging the features extracted from the three inputs, two attention mechanisms, channel and spatial, are introduced to enhance the importance of specific features. Then, the feature relationships are inferred through two layers of BiLSTM. Finally, the inverted SWH value is output through a fully connected network, and the Dropout technique is applied in this process to prevent overfitting. The mean squared error (mean_squared_error) is used as the loss function in this model, and the Adam optimizer is used for optimization.

[0095] As an implementation manner of this embodiment, for the model evaluation in step 4, taking the WW3 data as a reference, the root mean square error (RMSE), bias (Bias), correlation coefficient (CC), and mean absolute percentage error (MAPE) are used as accuracy indicators to evaluate the innovative model.

[0096] In terms of SWH estimation: The innovative DilatedResBiLSTM-AttnNet deep learning model is compared with seven other existing machine learning models or deep learning models (Decision Tree Regression (DTR), Light Gradient Boosting Machine (LightGBM), Extreme Gradient Boosting (XGBoost), Bagged Trees (BT), Deep Convolutional Neural Network (DCNN), Convolutional-Bidirectional Long Short-Term Memory Network (CNN-BiLSTM), Transformer).

[0097] It is very important to comprehensively evaluate the performance of the model. We compared and analyzed the DilatedResBiLSTM-AttnNet deep learning model proposed in the present invention with seven other machine learning models or deep learning models (DTR, LightGBM, XGBoost, BT, DCNN, CNN-BiLSTM, and Transformer). Attached Figure 3Shows the scatter plot of the comparison of the performance of different model spaceborne GNSS-R SWH estimations ((a)-(h) are DTR, lightgbm, XGBoost, BT, DCNN, CNN-BiLSTM, Transformer, and DilatedResBiLSTM-AttnNet), attached Figure 4 Shows the performance of different model spaceborne GNSS-R SWH estimations under different sea conditions. Table 3 shows the performance comparison of different models for inverting SWH and WW3SWH data, from Figure 3 、 Figure 4 And Table 3, the following experimental results can be obtained:

[0098] (1) It can be directly seen from the Figure 3 scatter plot that among the eight models, the proposed DilatedResBiLSTM-AttnNet model in this paper is better than the other seven machine learning or deep learning methods. From Table 3, it can be seen that among the eight models of decision tree regression (DTR), lightweight gradient boosting machine (LightGBM), extreme gradient boosting (XGBoost), bagged tree (BT), deep convolutional neural network (DCNN), convolutional-bi-directional long short-term memory network (CNN-BiLSTM), Transformer, and DilatedResBiLSTM-AttnNet, the DilatedResBiLSTM-AttnNet model has the best performance in terms of RMSE, CC, and MAPE, with values of 0.41m, 0.83, and 17.13% respectively, while DTR shows the worst performance, with RMSE of 0.58m, CC of 0.69, and MAPE of 21.35%;

[0099] (2) Figure 4 Shows the PDF curve distribution of the SWH estimated by eight models in the SWH range from 0 to 8m. From Figure 5 it can be seen that among the eight models, the PDF curve distribution of DTR has the best fit with ERA5, followed by the proposed DilatedResBiLSTM-AttnNet model, and the fit of lightgbm is the worst. Generally speaking, the proposed DilatedResBiLSTM-AttnNet model in the present invention has the best performance in estimating SWH.

[0100] Table 3 Performance comparison of different models for inverting SWH and WW3 data

[0101]

[0102] In terms of wave period estimation: Seven machine learning models, namely DTR, Extremely Randomized Trees, CatBoost, Lightbm, Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Support Vector Regression (SVR), are used for comparison with the physically-guided bagged tree wave period model estimation method proposed in the present invention.

[0103] To evaluate the generalization performance of the wave period estimation model proposed in the present invention, the wave period data of WW3 is used to test and evaluate the performance of the proposed physically-guided BT model and eight models, namely DTR, Extremely Randomized Trees, CatBoost, Lightbm, XGBoost, RF, and SVR, for estimating the wave period. Figure 5 Fig. shows the scatter plots of the comparison of the wave period estimation performance of different models for spaceborne GNSS-R ((a)-(h) are DTR, Extremely Randomized Trees, CatBoost, Lightbm, XGBoost, RF, and SVR respectively). Table 4 shows the comparison of the wave period estimation performance of eight machine learning methods. Figure 5 It shows that the wave periods predicted by the four models, namely BT, DTR, CatBoost, and RF, are relatively consistent with the wave periods of WW3, while there are large deviations between the predicted values of the four models, namely Extremely Randomized Trees, Lightgbm, XGBoost, and SVR, and the wave periods of WW3. In addition, it can be seen from the results in Table 4 that among the eight machine learning methods for estimating the wave period, the physically-guided ensemble learning method BT model proposed in this paper shows the best results, with RMSE, Bias, CC, and MAPE being 1.35 s, -0.01 s, 0.88, and 8.43% respectively, indicating that the physically-guided ensemble learning method BT model proposed in the present invention has achieved good results in estimating the wave period.

[0104] Table 4 Comparison of the wave period performance of eight machine learning methods

[0105]

[0106] In terms of wave energy estimation: Taking WW3 as a reference, the wave energy estimated by spaceborne GNSS-R is compared. Figure 6Scatter plots (left), PDFs (middle), and histograms (right) showing the comparison of wave energy between WW3 and GNSS-R are presented. The results in the figure indicate that there is a high correlation (CC = 0.94) between the wave energy estimated by the method proposed in the present invention and the WW3 wave energy data. However, the GNSS-R wave energy is slightly higher than the WW3 wave energy as a whole (Bias = 2.163 kW / m). At the same time, a positive bias of 2.16 kW / m can be seen from the histogram, and the GNSS-R wave energy is slightly higher than the WW3 wave energy as a whole, indicating that the method will overestimate the wave energy. In addition, the standard deviation of 9.1 kW / m indicates that the fluctuation of the bias is large. The PDF figure shows the distribution differences between GNSS-R and WW3 in different wave energy ranges. Generally speaking, there are no obvious differences between the two, demonstrating the great potential of the method proposed in the present invention in wave energy estimation.

[0107] Finally, it should also be noted that the terms "comprising", "including" or any other variant are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or elements inherent to such process, method, article or device. Although the preferred embodiments of the present invention have been described, those skilled in the art can make additional changes and modifications once they learn the basic creative concept. Therefore, the appended claims are intended to be construed as including the preferred embodiments and all changes and modifications falling within the scope of the present invention.

[0108] The above-described embodiments merely represent the specific implementation manners of the present application, and the description thereof is relatively specific and detailed, but it should not be construed as a limitation on the protection scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the technical solution of the present application, several deformations and improvements can still be made, and these all belong to the protection scope of the present application.

Claims

1. A method for constructing a satellite-borne GNSS-R wave energy flux estimation model, characterized in that It includes the following steps: Step S1, data preparation and preprocessing: First, download FY-3E / 3F / 3G GNSS-R data, ERA5 data (wind speed, SWH, wave period), and WW3 data (SWH, wave period) from the official website. Then, perform data feature extraction and data quality control preprocessing; Step S2, model construction: SWH is estimated through the constructed deep learning hybrid model (DilatedResBiLSTM-AttnNet) and the bias correction of SWH is performed using the adaptive CDF matching method. DilatedResBiLSTM-AttnNet includes: a multi-scale feature extraction module based on the residual network (ResNet) and dilated convolution, a fully connected network module that fuses multi-feature parameters, a feature enhancement module based on the channel attention mechanism and the spatial attention mechanism, and a feature relationship reasoning module based on BiLSTM; the wave period is estimated through the physics-guided bagging trees ensemble learning algorithm (Bagging Trees, BT); Step S3, wave energy flux estimation: The estimated value of the wave period is obtained by taking the estimated value of SWH after bias correction as the input of the wave period model, and together with the corrected SWH, it is used as the input of the wave energy formula to estimate the wave energy; Step S4, model evaluation: Taking the WW3 data as a reference, the innovative model is evaluated using the root mean square error (RMSE), bias (Bias), correlation coefficient (CC), and mean absolute percentage error (MAPE) as accuracy indicators.

2. The method according to claim 1, wherein The data preprocessing in step S1 includes: extracting the longitude and latitude of the specular reflection point, the data acquisition time, the normalized bistatic radar cross section (NBRCS), and the leading edge slope (LES) observation values from the FY-3E / 3F / 3G GNSS-R observation data; extracting the synthetic wind speed, significant wave height, and wave period from the ERA5 data, and extracting the significant wave height and wave period from the WW3 data; performing spatio-temporal matching on the GNSS-R data, ERA5, and WW3 data through linear interpolation and bilinear interpolation algorithms. In addition, to ensure the quality of the data, data quality control is carried out in the following way: (1) To ensure that all the data used is valid, it is necessary to check the "Bad_File_Flag" attribute of each data file. A "Bad_File_Flag" of 0 indicates that there is at least one valid DDM data point in the file, while a "Bad_File_Flag" of 1 indicates that there is no valid DDM data point in the file. Select the data files with a "Bad_File_Flag" of 0 for subsequent analysis; (2) The observed values must be positive, and when they are Nan values, they need to be discarded; (3) Select the points where the specular reflection point is on the ocean.

3. The method according to claim 1, characterized in that, In the step S2: The adaptive CDF matching method is a method that basically eliminates systematic errors by applying certain corrections to the output of the model so that the cumulative distribution function of the corrected model output is the same as that of the reference value. The specific method is as follows: In the training set, first arrange the inverted SWH of the model and the reference SWH in ascending order, then calculate the difference sequence D between the reference value and the inverted SWH. After correcting the inverted SWH by the difference, the distribution functions of the inverted SWH and the reference SWH sequences are the same. The difference sequence is fitted by a polynomial of order 0 to 10: where P n is an nth-order polynomial, is the sequence for inverting SWH. Define RMSE as the regression loss. In this polynomial, the polynomial order that minimizes the RMSE is adaptively found by the principle of nonlinear least squares. The corrected inversion result is expressed as: In addition, the wave period is estimated by a physics-guided bagging trees ensemble learning algorithm (Bagging Trees, BT), where the physics-guided constraint equation is expressed as: In the formula, X is used to describe the relationship between spaceborne GNSS-R observations (i.e., NBRCS (or LES) observations), SWH, and the wave period.

4. The method according to claim 1, wherein In the step S3: For the estimation of wave energy flux, the estimated value of SWH after bias correction is used as the input of the wave period model to obtain the estimated value of the wave period, and together with the corrected SWH, it is used as the input of the wave energy formula to estimate the wave energy. The mathematical expression of the wave energy formula is as follows: Where H s is the SWH, T e is the wave period, ρ and g are the seawater density and the acceleration of gravity respectively, ρ = 1025 kg / m3 , g = 9.81 m / s 2 .

5. The method according to claim 1, wherein The model evaluation in the step S4 includes: Using WW3 data as a reference, the innovative model is evaluated with the root mean square error (RMSE), bias (Bias), correlation coefficient (CC), and mean absolute percentage error (MAPE) as accuracy indicators. Regarding SWH estimation: The newly innovated DilatedResBiLSTM-AttnNet deep learning model is compared with seven other existing machine learning models or deep learning models (decision tree regression (DTR), light gradient boosting machine (LightGBM), extreme gradient boosting (XGBoost), bagging trees (BT), deep convolutional neural network (DCNN), convolutional bidirectional long short-term memory network (CNN-BiLSTM), Transformer). Regarding wave period estimation: Seven machine learning models, namely DTR, extremely randomized trees, gradient boosting decision tree (CatBoost), Lightbm, extreme gradient boosting (XGBoost), random forest (RF), and support vector regression (SVR), are compared with the physics-guided bagging tree wave period model estimation method proposed in the present invention. Regarding wave energy estimation: Using WW3 as a reference, it is compared with the wave energy estimated by spaceborne GNSS-R.

6. The method according to claim 2, wherein The spatio-temporal matching is specifically as follows: Linear interpolation algorithm is used for time matching, and bilinear interpolation algorithm is used for space matching to ensure the alignment of GNSS-R data, ERA5, and WW3 data at the same spatio-temporal resolution.

7. The method according to claim 1, characterized in that The DilatedResBiLSTM-AttnNet includes: a multi-scale feature extraction module based on the Residual Network (ResNet) and dilated convolution, a fully connected network module that fuses multi-feature parameters, a feature enhancement module based on the channel attention mechanism and the spatial attention mechanism, and a feature relationship inference module based on BiLSTM.

Images