Wave Parameter Retrieval From Synthetic Aperture Radar By Machine Learning

Waves are important research field in physical oceanography,accurate and effective acquisition of wave information is an important research content in oceanography.Due to the vast area of the ocean,traditional ocean observation methods,such as coastal observation stations and buoys,which can only obtain limited data and cannot obtain large-scale sea surface information.It is greatly affected by the weather and cannot be observed continuously for a long time.Due to the vast area of the ocean,traditional ocean observation methods,such as manual observation and buoy observation,can only perform single-point observation of the sea surface,and cannot obtain large-scale sea surface information.Synthetic Aperture Radar SAR(Synthetic Aperture Radar)has all-day,all-weather,multi-band and multi-polarization working modes,with high imaging resolution.It is an important means of observing ocean waves.Moreover,SAR is the only observation instrument that can obtain two-dimensional information of ocean waves,Therefore,obtaining wave information from SAR is an important research direction in the marine field in recent decades.This paper mainly discusses the development history of SAR satellite and the research status of significant wave height retrieval from SAR satellite in recent decades,briefly discusses the sea surface imaging mechanism and related theories of SAR,and then systematically analyzes the accuracy of ENVISAT ASAR(Advanced Synthetic Aperture Radar)wave mode data released by ESA(European Space Agency).Finally,based on the previous research of significant wave height retrieval methods,a new retrieval algorithm of significant wave height is established.SAR imaging theory of wave is complex,which controlled by three imaging theories: tilt modulation,hydrodynamic modulation and velocity bunching.The retrieval accuracy in different sea conditions is different.ASAR level-2 wave mode product provides wave information,including effective wave height,wave direction,wavelength and two-dimensional wave spectrum,so the precision of ASAR data are essential to model predictions.This paper systematically analyzes the accuracy of significant wave height of ASAR Lv2 production under different sea state,wind speed and two-dimensional spectrum types.Due to the special imaging mechanism of SAR satellite on the sea surface,there will be different measurement results under different sea conditions.Comparison with the buoy data of the national Buoy Center(NDBC),it shows that the ASAR significant wave height is underestimated under high sea state and overestimated under low sea conditions.The measurement result under medium sea conditions is better,and the RMSE is 0.65 m,SI is 20.73%.The wave spectra retrieved from ASAR wave mode images can be categorized into four types after carefully analyzing the energy distribution and spectral shapes.The accuracy of wave parameters of different spectra types was discussed by studying the corresponding two dimensional spectra,results showed that in cases of normal spectra with single direction,the SWH and wave direction were consistent with buoy data,while in cases of spectra with 180°wave direction ambiguity only SWH showed good consistence,and spectra with disordered shapes showed poor results.Two classical algorithms for retrieving wave parameters from SAR images are presented after years of research.One is the theoretical algorithm,according to the wave imaging principle of SAR,calculates and obtains the two-dimensional wave spectrum,and then obtains the wave parameters from the wave spectrum;The second is the empirical algorithm,which uses the mathematical model to calculate the wave parameters according to the relevant parameters obtained from SAR images.This paper uses back propagation neural network algorithm,combined with various parameters with empirical relationship with the significant wave height,such as azimuth cutoff wavelength,normalized radar cross section(NRCS),image signal-to-noise ratio(S2G,signal to noise),normalized image variance(NV)to propose a new algorithm.The more kinds of parameters in the model,the richer the wave information it contains,and the more accurate the retrieval result is.The correlation coefficient between the training model and the buoy data is 0.84,RMSE is 0.697,SI is 30.8%;The model accuracy is verified when the data are divided into offshore,deep ocean and different sea state.The results of model inversion are better than the original SAR data,which shows that the neural network algorithm can effectively retrieve the significant wave height,and the algorithm has strong applicability,which can be applied to different sea areas and different sea state.However,for extreme sea conditions such as typhoon,the result of significant wave height is poor.This study will supply valuable reference for subsequent related research.For example,the wave parameters retrieval algorithms of other series of satellites can consider different sea state,which may improve accuracy of the algorithm;The neural network algorithm model can also be used in the retrieval of wave parameters of other satellites,such as sentinel series satellites and Gao Fen-3 synthetic aperture radar satellite of China.