Inversion Of Seamount Chain Seafloor Topography Based On Grid Models With Different Resolutions And Machine Learnin

High precision seafloor topography plays an important role in ocean research and ocean engineering,including ship-based sonar detection and satellite altitude-gravity data inversion.In recent years,many scholars have used gravity anomaly and ship survey data to invert the seafloor topography by SAS(Sandwell&Smith)method,but ignored the influence of the nonlinear relationship between gravity anomaly and ship survey depth.Taking the Emperor Seishan chain as an example,the Convolutional Neural Networks(CNN)and SAS method are integrated in this thesis,and the nonlinear relationship between them is considered to construct the submarine terrain model,which is compared with SAS Method and GGM(Gravity-Geologic Method).Meanwhile,the accuracy of SAS inversion model under different resolutions and the influence of ship measuring point distribution on model accuracy are analyzed.Main results are summarized as follows:1.The inversion accuracy of terrain models under different resolutions of SAS method was evaluated.The residual error of inversion band fitting was optimized with the improvement of inversion resolution,and the standard deviation of difference(STD)at the check point decreased from 58.92 m to 47.01 m,which significantly reduced the difference with the high-precision terrain model and the grayed ship survey data.Relative errors are generally high near the seamount chain and decrease gradually with the increase of resolution.2.The influence of sparse density of ship survey data on the accuracy of submarine terrain inversion under different resolutions was analyzed.The accuracy of terrain model inversion in different density regions was optimized with the improvement of resolution.The standard deviation at the check point is reduced by 18.95%,both better than the other two regions,and the proportion of topographic difference distribution within-100 m to 100 m is increased.3.The preliminary model is obtained by using CNN to establish nonlinear mathematical relations.The residual error sum of squares(SSE),root mean square error(RMSE)and mean absolute error(MAE)of the preliminary terrain model are optimized,and the residual error is significantly reduced.The accuracy of the final model is comparable to that of SIO V19.1 and GEBCO-2020 models,and better than that of ETOPO1 model.The sea depth difference in the range of 200 m accounted for 77.31% and 78.43%,respectively,higher than other terrain models.The relative errors of CNN estimation model and NGDC grid model are smaller than those of other terrain models.4.Terrain was reproduced by SAS method and GGM method respectively,and compared with CNN inversion terrain.The results showed that the standard deviation of the check point,deviation from GEBCO-2020 model and relative error of CNN estimation model in the study area were lower than those of SAS method and GGM method,and were less affected by the distribution of ship measurement points,and the inversion effect was better.At the same time,CNN estimation model has higher correlation with GEBCO-2020 model and higher accuracy.