A convolutional neural network to optimize multi-mission satellite altimeter fusion for improving the marine gravity field

Abstract

Satellite altimetry is the main tool for constructing global or regional marine gravity fields. To improve the accuracy and spatial resolution, it is necessary to fuse multi-mission altimeters. How to determine the weights of multi-mission altimeters is a crucial issue, making the conventional calculation process very complex. In addition, traditional satellite inversion methods are often independent of shipborne gravity, which is used only as validation data, thus not take full advantages of high accuracy and resolution of shipborne gravity. In this study, we introduce a convolutional neural network (CNN) to merge the vertical deflections (DOVs) obtained from multi-altimeter missions to construct a marine gravity model in the South China Sea. High-accuracy shipborne gravity and a dataset comprising DOVs and geo-locations are employed as input data for neural network training. For the validation of CNN method, the gravity model is also computed by conventional Inverse Vening Meinesz (IVM) method. Independent shipborne gravity measurements and SIO V32.1, DTU17 models are used as validation data. The evaluation results show that the CNN-derived model achieves a higher level of accuracy, yielding a standard deviation (STD) of 3.21 mGal, with an improvement of 36.56% compared to IVM-derived model. More than 92% of the differences between the CNN-derived model and shipborne gravity are less than 5 mGal. In addition, spectral analysis results further show that the CNN-derived model has stronger energy at short wavelengths (less than 25 km) compared to other models. These findings reveal that CNN method is feasible for marine gravity recovery and the CNN-derived model can achieve higher accuracy. The CNN method can improve the accuracy and spectral characteristics of the constructed gravity model by taking advantage of the high accuracy and high resolution of shipborne gravity.