Altimetry‐Derived Gravity Gradients Using Spectral Method and Their Performance in Bathymetry Inversion Using Back‐Propagation Neural Network

Abstract

AbstractVariations in bathymetry produce gravity gradients; therefore, gravity gradients should be suitable for inverting bathymetry. However, classical bathymetry inversion methods often use gravity anomalies. Studies wherein gravity gradients are employed always use only vertical gravity gradients, ; however, the remaining five gravity gradients also contain bathymetric information. This study presents a spectral method of inverting gravity gradients from altimetry‐derived deflection of the vertical. The accuracy of derived gravity gradients is first evaluated by Laplace's equation. Results show that, except the Arctic region, the mean and standard deviation (STD) of the Laplace operator are, respectively, less than 0.001 and 0.06 E in the Atlantic, Indian and Pacific oceans. Computed was compared with from Scripps Institution of Oceanography. The differences have means and STDs smaller than 0.06 and 6.7 E, respectively, in all study areas, except the Arctic which registered an STD of 13.81 E. All six tensors were then fed into a back‐propagation neural network to predict bathymetries of each study area. The resultant bathymetries compare well with reference models from ship‐borne depths, SRTM15+V2 and GEBCO_2021. Relative to ship‐borne depths, the Atlantic, Pacific, Indian and Arctic study areas yielded mean errors, error STDs and correlation coefficients of 5.40, 3.28, 0.92, 3.07 m; 110.03, 76.23, 104.58, 132.41 m; and 0.9861, 0.9615, 0.9760, 0.9803; respectively. The bathymetries were inverted again, however, with each gravity gradient component omitted. Analysis showed that was not the most influential component across the study regions; rather, each gravity gradient component can significantly contribute to bathymetry inversion.