The Retrieval Method Of GNSS-R Sea Surface Height Based On Deep Learning

The global surface altimetry data can invert the global ocean gravity field and provide the data basis for underwater gravity matching-assisted navigation of submersible.The underwater gravity matching-assisted navigation can correct the system drift error in the inertial navigation of the submersible,and effectively improve its stealth and navigation accuracy.Therefore,high spatial resolution and high accuracy global ocean gravity field references are the key to improve underwater gravity matching assisted navigation.The GNSS-R altimetry is a new dual-base passive microwave radar measurement system,using the GNSS signal as the signal source and inverting the sea surface height(SSH)by measuring the time delay between the direct signal emitted by the GNSS satellite and the reflected signal reflected by the sea surface.Compared with active radar altimetry,GNSS-R altimetry has the advantages of abundant signal sources and low cost,and can realize global sea surface altimetry with high spatial and temporal resolution through multiple GNSS-R altimetry constellations.Deep learning methods are able to overcome the limitations imposed by imperfect theoretical models in the field of new remote sensing by approximating unknown predicted values in a datadriven way.The spaceborne GNSS-R receivers can receive millions of measurements per day,such an enormous amount of spaceborne observation data provides an opportunity to further optimize the data-driven deep learning algorithms.The retrieval model of spaceborne GNSS-R sea surface height based on deep learning is expected to improve the retrieval accuracy of spaceborne GNSSR SSH and provide a new theoretical and methodological reference for the accurate inversion of GNSS-R altimetry satellite SSH with centimeter-level accuracy in the future.With the scientific object of acquiring global high spatial resolution and high accuracy SSH and ocean gravity anomaly based on GNSS-R altimetry,and then improving underwater gravity matching navigation accuracy of submersible,this paper presents a high-precision spaceborne GNSS-R SSH method.The main research work and contributions of this paper are as follows:(1)To prestudy the accuracy of spaceborne GNSS-R ocean altimetry,taking the Baltic Sea airborne experimental data as an example,a new Machine Learning Weighted Average Fusion method(MLWAF)is constructed with the airborne time delay waveform as the input and the corresponding SSH as the output.In addition,to obtain a more appropriate feature set for the sea surface height retrieval model,three features sensitive to sea surface height variation,DER,HALF and LES,were constructed using the feature construction method.The effects of feature sets with different information details on the accuracy of SSH retrieval were analyzed.The experimental results show that the proposed MLWAF method achieves better retrieval results,with the mean absolute difference(MAD)about 0.25 m,root mean square error(RMSE)about 0.29 m,and pearson correlation coefficient(PCC)about 0.75 compared with the DTU15 validated model.The MAD and PCC are improved by 61% and 44%,respectively,compared with the traditional singlepoint retracking method.(2)To address the problem that spaceborne GNSS-R SSH retrieval error model is complicated and the accuracy is constrained,a new Multi-Hidden Layer neural network Feature Optimization(MHLFO)method is constructed with the time delay waveform data,metadata,atmospheric delay,and ocean wind speed of Tech Demo Sat-1(TDS-1)satellite as the input and the corresponding SSH as the output.By analyzing 14 feature sets with different information details,it is concluded that the elevation,signal-to-noise ratio(SNR),atmospheric delay,and ocean wind speed can provide essential contributions to the SSH retrieval based on GSMHLFO.The experimental results show that the proposed GSMHLFO with four hidden layers and 200 neurons per layer has better retrieval performance.Compared with DTU18,the mean absolute difference(MAD),the root mean square error(RMSE),and the Pearson correlation coefficient(PCC)equal 4.23 m,5.94 m,and 0.98,respectively.Compared with the HALF single-point retracking method,the performance metrics of MAD,RMSE and PCC are improved by 32.86%,25.00% and 8.99%,respectively.(3)To address the problem of low utilization of delay-Doppler Map(DDM)information in the traditional spaceborne GNSS-R SSH retrieval method and consider the heterogeneity of the input data,this paper proposes an end-to-end Modified Residual Multimodal Deep Learning(MRMDL)method that can utilize the entire range of DDM information.The MRes Net applicable to DDM structures is used to adaptively capture features of the full DDM and to convert the twodimensional DDM data into one-dimensional numerical form.Then,the extracted features and auxiliary parameters are fused as the input data for MHL-NN to retrieve the SSH.Second,the reliability of the model is verified by DTU18 validated model.and spaceborne radar altimeters(Jason3,HY-2C,HY-2B).Compared with the various Deep learning SSH retrieval methods,MRes Net can retain more information of DDM data and produce better retrieval results.Compared to the SSH provided by the DTU18 validation model and the spaceborne radar altimeter,the PCC are 0.98 and 0.97,respectively.However,the CYGNSS satellite is not primarily employed for ocean altimetry,and the MAD are 3.92 m and 4.32 m,respectively.There are 59 figures,16 tables and 209 references.