| Seafloor topography is an important research content of marine mapping,which has important economic,military,and scientific research values.It is difficult to obtain the global seafloor topography in a short time by the traditional ship-based measurement,and the inversion of seafloor topography by satellite altimetry gravity has become a feasible method to obtain the global seafloor topography quickly.However,there is still a great difference between the inversion accuracy and shipborne bathymetry,and the current inversion methods usually use only gravity anomaly data or vertical gravity gradient anomaly data for inversion,and do not integrate multisource gravity data for inversion.To address the problems of low inversion accuracy and single inversion data,this paper explores the technical ways to introduce machine learning and deep learning into the field of seafloor topography inversion,proposes a neural network inversion method to fuse multi-source data,compares and analyzes the accuracy performance of various inversion methods from multiple dimensions,and further analyzes the spectral characteristics of inversion models and their adaptability to different depths and topography.The main research contents and works of this paper are as follows:(1)The advantages and disadvantages of the three commonly used inversion methods are summarized.The results show that the calculation accuracy of gravity geological method(GGM)is high,but time-consuming and the inversion accuracy depends on the selection of the optimal density difference constant;the inversion accuracy of SAS method and vertical gravity gradient anomaly method based on frequency domain is lower than that of gravity geology method,and there are also disadvantages such as edge effect and poor presentation of short-wave topography,but their calculation time is short,the inversion speed is fast,and the dependence on parameters is low;in addition,the high frequency noise in vertical gravity gradient anomaly also makes the inversion results of vertical gravity gradient anomaly unstable.In addition,the high-frequency noise in the vertical gravity gradient anomaly also makes the inversion results of vertical gravity gradient anomaly unstable.(2)A neural network-based method for inversion of seafloor topography is proposed.Inspired by Goog Le Net,a back propagation(BP)neural network with two channels is constructed in this paper and named parallel connected BP neural network(PLBP).Operations such as detrending,cross-spectral analysis,and linear regression are used to split the gravity signal into a reference field and a residual field as feature data.The method integrates gravity anomalies and vertical gravity gradient anomalies for seafloor topography inversion,which improves the accuracy of seafloor topography inversion.(3)The accuracy performance of each inversion model was compared and analyzed by using single-beam and multibeam bathymetry data as the internal and external compliance testing standards,respectively.The results show that the relative accuracy of the neural network inversion results with 1’×1’ resolution and single-beam bathymetry difference is 1.76%,and the root mean square is 72.40 m,which is 19%better than the gravity geology method inversion accuracy.The accuracy of SAS method based on Fourier transform and vertical gravity gradient anomaly method is comparable and lower than that of GGM inversion;the results of external conformity accuracy check show that the standard deviation of difference of neural network model is 165.90 m,which is comparable to the accuracy of GGM model,and the SAS model and VGG model are worse in checking with multibeam data.(4)The inversion models were analyzed for spectral characteristics as well as for their adaptability.The results of the spectral analysis show that there are significant differences among the inversion models in the short-wave band in this study area,and the differences of these models are mainly distributed in the wavelength band from 6.2km to 38.4 km,in which the VGG model performs the best,and the neural network model performs the best in the short-wave band with wavelength less than 6.2 km.The results of the adaptation analysis show that the inversion accuracy improves overall with increasing water depth;the adaptation of each inversion model to different terrains is different,i.e.,the adaptation is higher in the area with flat terrain and lower in the area with large terrain relief,and the inversion model is more adaptable to the terrain relief with larger size and higher inversion accuracy compared with the dense small size terrain relief.This paper proposes a neural network-based water depth inversion method,which proves the high accuracy and feasibility of the neural network water depth inversion method,and provides a new idea and method for seabed terrain modeling.In the future,this method will continue to be promoted and applied in the research and practice of gravity inversion of water depth,providing more accurate and refined data support for better understanding and exploration of the ocean. |
