A deep learning-based method for inverting global sea surface significant wave height using spaceborne GNSS-R

Abstract

The present invention discloses a deep learning-based satellite GNSS-R global sea surface significant wave height inversion method, which uses a convolutional neural network and a fully connected network to deeply fuse the sea surface environment and satellite auxiliary information to construct a deep learning model for global sea surface significant wave height inversion, including the following steps: obtaining satellite GNSS-R observation data, ERA5 reanalysis data, IMERG rainfall data, and WaveWatchIII significant wave height data; extracting characteristic parameters from satellite GNSS-R data and performing spatiotemporal matching on all data sets; data quality control and data partitioning; deep learning model construction and training for significant wave height inversion; inputting a test data set into the trained significant wave height inversion model to obtain the inverted significant wave height value, and evaluating the result. The technical solution of the present invention can achieve high-precision and high-resolution global ocean significant wave height inversion, providing a new reference method for satellite GNSS-R sea surface significant wave height inversion.

Description

Technical Field

[0001] The present invention belongs to the interdisciplinary research field of GNSS reflection measurement ocean remote sensing and computer science, and specifically, relates to a deep learning-based satellite GNSS-R global sea surface effective wave height inversion method. Background Art

[0002] Observing the wave height distribution on a global scale is one of the key research areas in oceanography. Wave height is usually measured by significant wave height (SWH). SWH observation is also of great significance for the study of wave energy resource assessment and ocean storm disaster prediction. In addition, as one of the important ocean dynamic environment monitoring parameters, significant wave height has always attracted widespread attention from many scholars in the field of ocean remote sensing and is a hot topic in the field of ocean remote sensing.

[0003] At present, for the inversion of GNSS-R effective wave height for spaceborne platforms, domestic and foreign researchers mainly build effective wave height inversion empirical models based on SNR, DDMA and LES observations. Although some promising results have been achieved, the effective wave height estimation model based on SNR observations is only applicable to the inversion of effective wave height under low sea conditions, while the inversion accuracy of the model is low for high effective wave height sea conditions. In addition, the accuracy improvement of the empirical combination model is very limited compared with the single observation model (such as DDMA or LES), and the inversion accuracy is not ideal under high effective wave height sea conditions. The main reason is that it is challenging to use other factors related to the GNSS-R process (such as geometry and wind speed) as inputs of the empirical model, which has not been deeply considered in the modeling. The effective wave height is affected by a variety of complex factors, which make the relationship between the sea surface effective wave height and the GNSS-R eigenvalue more complicated and difficult to characterize with a simple empirical model. Deep learning can make full use of the self-learning and adaptive capabilities of neurons to deal with complex nonlinear problems, so the influence of different geophysical and geographical parameters can be considered, which opens a new window for using deep learning methods to perform spaceborne GNSS-R effective wave height inversion.

[0004] In summary, the present invention proposes a deep learning-based satellite GNSS-R global sea surface effective wave height inversion method. Summary of the invention

[0005] Abstract: In order to overcome the problems that the existing empirical models for inversion of sea surface significant wave height by satellite-borne GNSS-R have poor generalization ability and are difficult to promote for practical application, the present invention provides a method for inversion of global sea surface significant wave height by satellite-borne GNSS-R based on deep learning. Compared with the empirical model method which only considers a limited number of variable parameters related to the significant wave height, the method uses a convolutional neural network (CNN) and a fully connected network (FCN) to deeply fuse the sea surface environment and satellite auxiliary information to construct a deep learning model for inversion of global sea surface significant wave height, so as to improve the accuracy of inversion of global sea surface significant wave height based on satellite-borne GNSS-R.

[0006] In order to solve the above technical problems, the basic concept of the technical solution adopted by the present invention is:

[0007] A deep learning-based method for inverting global sea surface significant wave height using satellite-borne GNSS-R includes the following steps:

[0008] Step S1, obtaining satellite GNSS-R observation data, ERA5 reanalysis data (wind speed, wind direction, water depth and significant wave height), IMERG rainfall data and WaveWatch III significant wave height data;

[0009] Step S2, extracting characteristic parameters from the onboard GNSS-R data and performing spatiotemporal matching on all data sets;

[0010] Step S3, data quality control and data division;

[0011] Step S4, deep learning model construction and training for significant wave height inversion.

[0012] Step S5, input the test data set into the trained significant wave height inversion model to obtain the inverted significant wave height value, and evaluate the result.

[0013] Furthermore, the satellite-borne GNSS-R observation data includes L1b power DDM (power_analog), normalized bistatic radar cross section (NBRCS), leading edge slope (LES), signal-to-noise ratio (SNR), receiver antenna gain (sp_rx_gain), longitude and latitude (sp_lon, sp_lat) and incident angle (sp_inc_angle) of the mirror reflection point, and the data version is V3.0; the ERA5 wind speed is the composite wind speed of the U component and the V component 10m above the sea surface; the wind direction data is the wind direction data (10m wind direction) provided in the ERA5 reanalysis; the water depth data is the water depth (Modelbathymetry) provided in the ERA5 product; the significant wave height data is the significant wave height of combined wind waves and swell (Significant height of combined wind waves and swell) and WaveWatch provided in the ERA5 reanalysis. III significant wave height (sea_surface_wave_significant_height) data; the IMERG rainfall data is IMERG-F rainfall intensity, and the data version is IMERG-V6.

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

[0015] Step S2.1, extracting characteristic parameters from the onboard GNSS-R data;

[0016] Step S2.2, perform spatiotemporal matching on all data sets to obtain matched data sets.

[0017] Furthermore, the quality control of the matched data set described in step S3 includes that the difference between the ERA5 significant wave height and the WaveWatchIII significant wave height is less than 0.5m; the NBRCS, LES and SNR observation values ​​are greater than 0; the RCG value should be greater than 3; the gain of the receiving antenna in the direction of the reflection point (sp_rx_gain) is greater than 0dBi; all observations containing NaN values ​​are deleted; the uncertainty of BRCS (ddm_BRCS_uncert) is less than 1; data quality filtering uses basic quality control (QC) marks in CYGNSS L1B data. Data on the ocean where the mirror reflection point is more than 25km away from the land is selected to reduce modeling errors. The quality-controlled data set described in step S3 is divided into training set, validation set and test set according to random selection, accounting for 30%, 20% and 50% of the filtered data set respectively.

[0018] Furthermore, the deep learning model for significant wave height inversion described in step S4 is composed of two core neural networks, the first network is a convolutional neural network, the input is the power DDM, and the convolutional neural network can be used to effectively extract the characteristic values ​​related to the significant wave height of the sea surface from the power DDM, and the second network is a fully connected network, and the input is the sea surface environment and satellite auxiliary parameter information. Finally, the significant wave height inversion value is output by combining the two core networks and using the fully connected network.

[0019] Furthermore, the training process of the satellite-borne GNSS-R global sea surface effective wave height inversion model is specifically implemented as follows:

[0020] First, the CYGNSS data need to be preprocessed and selected before training the significant wave height inversion model;

[0021] Then, the DDM data collection time and the location of the mirror reflection point are used to match the time and space with other data products using temporal linear interpolation and spatial bilinear interpolation algorithms, and the quality-controlled data set is divided into training set, validation set and test set;

[0022] The model is trained using the training set data, where the number of hidden layer neurons in the fully connected layer is obtained through empirical training to obtain the best performance;

[0023] When training the significant wave height inversion model, a mini-batch gradient descent method was used due to the huge number of samples. The batch size N was set to 32. To ensure good training performance, the number of epochs was set to 100. The effective wave height deep learning network used was implemented under the Keras architecture and TensorFlow backend;

[0024] Rectified linear unit (ReLU) is used as the activation function, Adam is selected as the optimizer, and mean absolute percentage error (MAPE) is used as the loss function; the learning rate is set to 0.0005. In order to further prevent overfitting problems, random inactivation (dropout) is set to 0.1;

[0025] When the training is completed, the test data set is input into the trained significant wave height inversion model to obtain the inverted significant wave height value, and the results are evaluated using the ERA5 reanalysis significant wave height and WaveWatch III significant wave height data, respectively.

[0026] The present invention uses a deep learning model framework to realize the inversion of global sea surface effective wave height by satellite-borne GNSS-R. The model framework includes the process of extracting effective features from DDM, deeply integrating these features with the sea surface environment and satellite auxiliary information, regressing the selected model with actual reference data, and verifying the model performance. The deep learning model for inversion of satellite-borne GNSS-R effective wave height jointly constructed by convolutional neural network and fully connected network can effectively improve the inversion accuracy. The method provides automatic feature extraction from the entire DDM using convolutional layers, rather than from a small box containing specular reflection points, because the physical information reflected in the small box is limited. The obtained model has good generalization ability and can produce effective wave height inversion results with high precision and high spatiotemporal resolution for the global ocean, which means good performance in practical applications. In addition, the method provided by the present invention is expected to become a useful guide for creating more advanced SWH inversion algorithms in the future.

[0027] After adopting the above technical scheme, the present invention has the following beneficial effects compared with the prior art.

[0028] The present invention uses a convolutional neural network (CNN) and a fully connected network (FCN) to deeply fuse the sea surface environment and satellite-aided information to construct a deep learning model for global sea surface effective wave height inversion. The obtained model has good generalization ability and can produce effective wave height inversion results with high precision and high spatiotemporal resolution for the global ocean, which means it performs well in practical applications. The method provided by the present invention is expected to become a useful guide for creating more advanced SWH inversion algorithms in the future.

[0029] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The accompanying drawings are part of this application and are used to provide a further understanding of the present invention. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but do not constitute an improper limitation of the present invention. Obviously, the drawings described below are only some embodiments. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0031] Figure 1 This is a flow chart of a method for inverting global sea surface significant wave height using satellite-borne GNSS-R based on deep learning according to an embodiment of the present invention.

[0032] Figure 2 This is a structural diagram of a deep learning-based satellite GNSS-R global sea surface significant wave height inversion model in an embodiment of the present invention.

[0033] Figure 3 It is a scatter density diagram of the effective wave height inverted by the deep learning model in an embodiment of the present invention and the ERA5 effective wave height.

[0034] Figure 4 It is a scatter density diagram of the significant wave height inverted by the deep learning model in an embodiment of the present invention and the WaveWatch III significant wave height.

[0035] Figure 5 It is the global distribution map of ERA5 significant wave height and the global distribution map of significant wave height inverted by deep learning model in the embodiment of the present invention.

[0036] Figure 6 This is a histogram of the deviation distribution between the significant wave height inverted by the deep learning model in an embodiment of the present invention and the ERA5 significant wave height.

[0037] Figure 7 This is a histogram of the deviation distribution between the significant wave height inverted by the deep learning model in an embodiment of the present invention and the significant wave height of WaveWatch III.

[0038] It should be noted that these drawings and textual descriptions are not intended to limit the conceptual scope of the present invention in any way, but are intended to illustrate the concept of the present invention for those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The following embodiments are used to illustrate the present invention but are not used to limit the scope of the present invention.

[0040] Example 1

[0041] In order to verify the effectiveness of the method proposed in the present invention, CYGNSS GNSS-R observation data, ERA5 reanalysis data (wind speed, wind direction, water depth and significant wave height), IMERG rainfall data and WaveWatchIII significant wave height data for all days of 2019 were obtained for experiments, and the experimental results of the present invention were compared with the results of significant wave height inversion using the empirical model, the fully connected network (FCN) model and the convolutional neural network (CNN) model. The basic configuration of the experimental platform of this embodiment is shown in Table 1:

[0042] Table 1 Configuration of the experimental platform

[0043]

[0044] A deep learning-based method for inverting global sea surface effective wave height using satellite-borne GNSS-R. The technical solution implementation process is as follows: Figure 1As shown, the following steps are included:

[0045] Step S1, obtaining satellite GNSS-R observation data, ERA5 reanalysis data (wind speed, wind direction, water depth and significant wave height), IMERG rainfall data and WaveWatch III significant wave height data;

[0046] Step S2, extracting characteristic parameters from the onboard GNSS-R data and performing spatiotemporal matching on all data sets;

[0047] Step S3, data quality control and data division;

[0048] Step S4, constructing and training a deep learning model for effective wave height inversion. The constructed deep learning model for effective wave height inversion is shown in the attached figure. Figure 2 shown.

[0049] Step S5, input the test data set into the trained significant wave height inversion model to obtain the inverted significant wave height value, and evaluate the result.

[0050] In step S1, the satellite-borne GNSS-R observation data includes L1b power DDM (power_analog), normalized bistatic radar cross section (NBRCS), leading edge slope (LES), signal-to-noise ratio (SNR), receiver antenna gain (sp_rx_gain), longitude and latitude (sp_lon, sp_lat) of the mirror reflection point, and incident angle (sp_inc_angle); the ERA5 wind speed is the composite wind speed of the U component and the V component 10m above the sea surface; the wind direction data is the wind direction data (10m winddirection) provided in the ERA5 reanalysis; the water depth data is the water depth (Model bathymetry) provided in the ERA5 product; the significant wave height data is the significant wave height of combined wind waves and swell (Significant height ofcombined wind waves and swell) and WaveWatch III significant wave height (sea_surface_wave_significant_height) data provided in the ERA5 reanalysis; and the IMERG rainfall data is the IMERG-F rainfall intensity.

[0051] Wherein step S2 includes the following sub-steps:

[0052] Step S2.1, extracting NBRCS, LES, receiver antenna gain, reflection point longitude, reflection point latitude and incident angle characteristic parameters from the spaceborne GNSS-R data;

[0053] In step S2.2, the power DDM (power_analog) and related characteristic parameters in the CYGNSS GNSS-R L1 level data are temporally and spatially matched with the ECMWF ERA5 wind speed, ERA5 wind direction, ERA5 water depth, ERA5 significant wave height, WaveWatch III significant wave height, and IMERG rainfall data to obtain the matched data set.

[0054] The quality control of the matched dataset described in step S3 includes that the difference between the ERA5 significant wave height and the WaveWatch III significant wave height is less than 0.5m; the NBRCS, LES and SNR observation values ​​are greater than 0; the RCG value should be greater than 3; the gain of the receiving antenna in the direction of the reflection point (sp_rx_gain) is greater than 0dBi; all observations containing NaN values ​​are deleted; the uncertainty of BRCS (ddm_BRCS_uncert) is less than 1; data quality filtering uses basic quality control (QC) flags in CYGNSS L1B data. Data on the ocean where the specular reflection point is more than 25km away from the land is selected to reduce modeling errors. The quality-controlled dataset described in step S3 is divided into training set, validation set and test set according to random selection, accounting for 30%, 20% and 50% of the filtered dataset respectively.

[0055] The training process of the spaceborne GNSS-R global sea surface significant wave height inversion deep learning model described in step S4 specifically includes:

[0056] After quality control of all data sets, the quality-controlled data sets are divided into training sets, validation sets, and test sets;

[0057] After quality control, the dataset is normalized, the learning rate is set to 0.0005, and the dropout is set to 0.1 to prevent overfitting.

[0058] The rectified linear unit (ReLU) is used as the activation function and the mean absolute percentage error (MAPE) is used as the loss function (as shown in Formula (1)). The Adam optimizer is used to train on randomly shuffled mini-batches.

[0059]

[0060] In the formula, N is the number of samples, To predict wind speed, u i is the reference wind speed.

[0061] The batch size N is set to 32. To ensure good training performance, the number of epochs is set to 100. The significant wave height inversion model is implemented under the Keras architecture and TensorFlow backend.

[0062] When the training is completed, the test data set is input into the trained significant wave height inversion model to obtain the inverted significant wave height value, and the results are evaluated using ERA5 reanalysis significant wave height and WaveWatch III significant wave height data. The root mean square error (RMSE), mean absolute error (MAE), correlation coefficient (CC) and mean absolute percentage error (MAPE) are used as indicators to evaluate the inversion performance of the model.

[0063] The trained deep learning model for effective wave height inversion is used to invert the global sea surface effective wave height, and the experimental results of the present invention are compared with the results of effective wave height inversion using the empirical model, the fully connected network (FCN) model, and the convolutional neural network (CNN) model. The final results are shown in Tables 2 and 3, which respectively give the inversion accuracy statistics of the effective wave height inverted by different inversion models in the range of 0-7m and the comparison with the ERA5 reanalysis effective wave height and the WaveWatch III effective wave height. The following conclusions can be drawn from the table:

[0064] (1) The inversion accuracy of the empirical model method is poor. However, compared with the ERA5 reanalysis significant wave height, the proposed model improves RMSE, MAE, CC and MAPE by 40%, 45.65%, 13.25% and 54.26% respectively. Compared with the WaveWatch III significant wave height, the proposed model improves RMSE, MAE, CC and MAPE by 30.88%, 37.25%, 13.58% and 49.62% respectively.

[0065] (2) The proposed model has the highest inversion accuracy. When the ERA5 reanalysis significant wave height is used as a reference, the multi-protocol model improves by 29.41%, 32.43%, 5.62% and 33.29% in RMSE, MAE, CC and MAPE compared with the FCN model, respectively; when the WaveWatch III significant wave height is used as a reference, the multi-protocol model improves by 20.34%, 23.81%, 5.75% and 26.78% in RMSE, MAE, CC and MAPE compared with the FCN model, respectively;

[0066] (3) Among the deep learning methods, the inversion accuracy of the CNN model trained using only DDM images as input is poor, which indicates that integrating DDM and other auxiliary parameter modeling is very important for improving model performance.

[0067] Table 2 Accuracy of wind speed inversion of different models on the test data set compared with ERA5 data

[0068]

[0069] Table 3 Accuracy of wind speed inversion of different models on the test dataset compared with WaveWatch III data

[0070]

[0071] In addition, in order to compare the significant wave height inverted by the model proposed in this embodiment with other inversion methods, the significant wave height of ERA5 and WaveWatch III is significantly improved in terms of correlation, as shown in the attached figure. Figure 3 and attached Figure 4 As shown, it can be seen that the inversion results of the significant wave height model method proposed in this embodiment have good correlation with the ERA5 significant wave height and the WaveWatch III significant wave height, which is better than the existing empirical model (DDMA method), FCN and CNN model methods. Obviously, the proposed model has achieved good effective wave height inversion performance, and has a high correlation with both the ERA5 significant wave height and the WaveWatch III significant wave height data. However, both the empirical model and the CNN model have obvious underestimation when the significant wave height value is greater than 3m. The inversion performance of the FCN model and the proposed model is better than the other two methods under large significant wave height conditions, and the inverted significant wave height values ​​are relatively clustered along the 1:1 reference line. Among them, the inversion performance of the proposed model is significantly better than that of the FCN model, which shows that the significant wave height model proposed in this embodiment further enhances the robustness, stability and versatility of the model.

[0072] In order to further evaluate the global performance of the proposed model in retrieving significant wave height, Figure 5 The figure shows the global distribution of ERA5 significant wave height and significant wave height inverted by different models. It can be seen from the figure that the significant wave height inversion model proposed in this embodiment has achieved a high inversion accuracy in inverting the global sea surface significant wave height, which is consistent with the global significant wave height distribution of ERA5. The FCN model has the second best performance. However, the DDMA method and the CNN model have large deviations in most sea areas around the world. Figure 6 As shown in the figure, from the deviation distribution histogram between the significant wave height inverted by the proposed model and the significant wave height of ERA5 (the figure shows the average deviation (μ), standard deviation (σ), mean absolute error (MAE) and 80% quantile (Qua) ​​of the deviation, the blue bar graph represents the error distribution, and the red bar graph represents the probability density function fitting curve of the error), it can be seen that the deviation between the significant wave height inverted by the proposed model and the significant wave height of ERA5 is very concentrated (80% of the wind speed deviations are less than 0.38m) and is near the significant wave height deviation of 0m. As shown in the attached figure Figure 7As shown in the figure, the deviation distribution histogram between the significant wave height inverted by the proposed model and the significant wave height of WaveWatch III also shows that the inversion performance of the proposed model is better than that of the other three methods. The above analysis results show that the deep learning method proposed in the present invention has significant advantages in inverting the global sea surface significant wave height.

[0073] The above is only a preferred embodiment of the present invention, and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment, it is not used to limit the present invention. Any technician familiar with this patent can make some changes or modify the technical contents suggested above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the scope of the solution of the present invention.

Claims

1. A deep learning-based method for inverting global sea surface significant wave height using satellite-borne GNSS-R, characterized in that: The following steps are involved: Step S1, obtaining spaceborne GNSS-R observation data, ERA5 reanalysis data, IMERG rainfall data and WaveWatch III significant wave height data; Step S2, extracting characteristic parameters from the onboard GNSS-R data and performing spatiotemporal matching on all data sets; Step S3, data quality control and data division; Step S4, deep learning model construction and training for significant wave height inversion; The deep learning model consists of two core neural networks. The first neural network is a convolutional neural network, and the input is the power DDM. The convolutional neural network can be used to effectively extract the characteristic values ​​related to the sea surface significant wave height from the power DDM; The second neural network is a fully connected network, and its input is the sea surface environment and satellite auxiliary parameter information; We use ReLU as the activation function and MAPE as the loss function, and use the Adam optimizer to train on randomly shuffled mini-batches. Step S5, input the test data set into the trained significant wave height inversion model to obtain the inverted significant wave height value, and evaluate the result.

2. The deep learning-based satellite GNSS-R global sea surface significant wave height inversion method according to claim 1 is characterized in that: The satellite-borne GNSS-R observation data includes L1b power DDM (power_analog), normalized bistatic radar cross section (NBRCS), leading edge slope (LES), signal-to-noise ratio (SNR), receiver antenna gain (sp_rx_gain), longitude and latitude of the specular reflection point (sp_lon, sp_lat) and incident angle (sp_inc_angle); The ERA5 reanalysis data includes wind speed, wind direction, water depth and significant wave height; the ERA5 wind speed is the composite wind speed of the U component and the V component 10m above the sea surface; the wind direction data is the wind direction data (10m winddirection) provided in the ERA5 reanalysis; the water depth data is the water depth (Model bathymetry) provided in the ERA5 product; the significant wave height data is the significant wave height of combined wind waves and swell (Significant height of combined wind waves and swell) and WaveWatch III significant wave height (sea_surface_wave_significant_height) data provided in the ERA5 reanalysis; The IMERG rainfall data is IMERG-F rainfall intensity.

3. The deep learning-based satellite GNSS-R global sea surface significant wave height inversion method according to claim 1, characterized in that: Step S2 includes the following sub-steps: Step S2.1, extracting NBRCS, LES, receiver antenna gain, reflection point longitude, reflection point latitude and incident angle characteristic parameters from the spaceborne GNSS-R data; In step S2.2, the power DDM (power_analog) and related characteristic parameters in the CYGNSS GNSS-R L1 level data are temporally and spatially matched with the ECMWF ERA5 wind speed, ERA5 wind direction, ERA5 water depth, ERA5 significant wave height, WaveWatch III significant wave height, and IMERG rainfall data to obtain the matched data set.

4. The deep learning-based satellite GNSS-R global sea surface significant wave height inversion method according to claim 1, characterized in that: The quality control of the matched data set described in step S3 includes that the difference between the ERA5 significant wave height and the WaveWatch III significant wave height is less than 0.5m; the NBRCS, LES and SNR observation values ​​are greater than 0; the RCG value should be greater than 3; the gain of the receiving antenna in the direction of the reflection point (sp_rx_gain) is greater than 0dBi; all observations containing NaN values ​​are deleted; the uncertainty of BRCS (ddm_BRCS_uncert) is less than 1; data quality filtering uses basic quality control (QC) flags in CYGNSS L1B data; Select the ocean data where the specular reflection point is more than 25km away from the land. The quality-controlled data set described in step S3 is divided into a training set, a validation set and a test set according to random selection, which account for 30%, 20% and 50% of the filtered data set respectively.

5. The deep learning-based satellite GNSS-R global sea surface significant wave height inversion method according to claim 1, characterized in that: The training process of the deep learning model for effective wave height inversion is implemented as follows: After quality control of all data sets, the quality-controlled data sets are divided into training sets, validation sets, and test sets; After quality control, the dataset is normalized, the learning rate is set to 0.0005, and the dropout is set to 0.1 to prevent overfitting. We use ReLU as the activation function and MAPE as the loss function, and use the Adam optimizer to train on randomly shuffled mini-batches. The batch size N is set to 32. To ensure good training performance, the number of epochs is set to 100. The significant wave height inversion model is implemented under the Keras architecture and TensorFlow backend; When the training is completed, the test data set is input into the trained significant wave height inversion model to obtain the inverted significant wave height value, and the results are evaluated using the ERA5 reanalysis significant wave height and WaveWatch III significant wave height data, respectively.

Images