Fast Dynamical-Statistical Reconstruction Of Three-Dimensional Ocean Data:A Research On The Framework And Key Techniques

The three-dimensional(3D)reconstruction of ocean data is an important technical means to estimate the 3D oceanic field based on multi-source data,including satellite observations and in situ observations,etc.It is a frontier field of Marine science research and an important direction of Marine hydrological support.Enlighted by the development and demand of Ocean Data Reconstruction(ODR),scientific issues and research objectives are drawn in terms of this article to make innovative exploration of ODR both theoretically and technically.From the perspective of reverse problem,a new definition of ODR is proposed,and a new concept of"information-dominated"ODR is put forward.New ODR methods are developed,validated and demonstrated,constituding a fast dynamical-statistical reconstruction model.This article is promising to provide theoretical exploration and technical support for the research and development of a portable,lightweighted and fast reconstruction system.The main results and innovations include:(1)Diagnosis of the inconsistency between satellite salinity and in situ salinity.By comparing the differences between the three satellite products and the near surface salinity of in situ observations,the inconsistent observation pairs that are significantly different from the three satellite products were extracted based on the improved3-Sigma criterion.The continuous inconsistent pairs of equatorial mooring buoys correspond to the obvious drift of mooring data.And the inconsistencies of Argo observations reveal that subfootprint variation and temporal undersampling are the important factors that induce the inherent difference between satellite salinity and in situ salinity.(2)High-resolution sea surface salinity(SSS)product reconstruction based on multi-fractal fusion.Through wave number spectrum analysis,four sea surface temperature(SST)products were compared,of which the most suitable product was selected as the template.Satellite SSS data was reconstructed by means of multi-fractal fusion algorithm to reduce the error and improve the resolution.A high-quality SSS product was generated with the nominal resolution of 0.1° and effective resolution of30-40km.(3)Dynamic-statistical reconstruction of ocean density,temperature and salinity based on Surface Quasi-Geostrophic(SQG)dynamics.Combining the dynamical mode of SQG with the empirical mode extracted from historical data,a new method of dynamical-statistical density reconstruction,SQG-mEOF-R,was proposed.Considering that the traditional SQG-based algorithms could not resolve the thermohaline variation directly,the LS-mEOFS algorithm of density-thermohaline transformation was proposed,and the technical framework of fast dynamical and statistical thermohaline reconstruction was constructed.The results of Observational System Simulation Experiment(OSSE)show that SQG-mEOF-R is superior to the pure dynamical reconstruction methods and pure statistical reconstruction methods.The thermohaline fields reconstructed by SQG-mEOF-R plus LS-mEOFs are superior to those reconstructed by the machine learning algorithms of generalized regression neural network and random forest.(4)Validation and application experiment of SQG-based dynamical-statistical reconstruction.Although the SQG-based methods have been verified in the model output data,hardly any research successfully demonstrated the superiority of SQG-based methods in terms of observed data.In this paper,the difficulties of the realistic application of SQG-based methods to observated data were analyzed from the SQG theory.The"First Guess"(FG)framework was proposed to circumvent the problems,so that the SQG-mEOF-R and LS-mEOFS methods can be put into practice.An optimization was further carried out by the empirical modes extracted from reanalysis data.The results manifest that the dynamic-statistical methods proposed in this article can be better than the statistical algorithms including random forest in the reconstruction of density,temperature and salinity in the upper ocean.(5)"Information facilitated"machine learning method and thermohaline reconstruction.By changing a variety of machine learning algorithms including deep learning network,as well as inputting the additional factors of temporal information and the SQG-mEOF-R density reconstruction,the thermohaline reconstructions are evaluated.The aim of the evaluated is to compare the role of model and information to the reconstruction improvement.The results allude to the limited contribution of changing nonlinear algorithms but the significant effect of inputting additional information.It is concluded that the machine-learning-based ocean data reconstruction is"information dominated"rather than"model dominated".(6)"Information embedded"deep learning network and thermohaline reconstruction.By means of empirical mode,the information of reanalysis data was embedded into the deep network,and a technical approach of deep network EOF embedding feed-forward neural network(EE-FFNN)was proposed.Compared with the reconstructions of multivariate linear regression and conventional deep network algorithm as well as reanalysis data,EE-FFNN could fulfill the potential of deep learning network in information mining and the capability of EOF in information compression.It turns out that the EE-FFNN can significantly improve the performance of subsurface thermohaline reconstruction.