SUMMARY: During the inversion of seafloor topography (ST) using the backpropagation neural network (BPNN), the random selection of parameters may decrease the accuracy. To address this issue and achieve a more efficient global search, this paper introduces a genetic algorithm-backpropagation (GA-BP) neural network. Benefiting from the global search and parallel computing capabilities of the GA, this study refines the ST of the South China Sea using multisource gravity data. The results indicate that the GA-BP model, with a root mean square (RMS) value of 126.0 m concerning ship-measured water depths. It is noteworthy that when dealing with regions characterized by sparse survey line distributions, the GA-BP neural network stronger robustness compared to BPNN, showing less sensitivity to the distribution of survey data. Furthermore, the paper explores the influence of different data pre-processing methods on the neural network inversion of sea depths. This research introduces an optimization algorithm that reduces instability during BPNN initialization, resulting in a more accurate prediction of ST.

SUMMARY: The precise estimation of the mass changes in Madagascar is a challenge by using the Gravity Recovery and Climate Experiment (GRACE) mission Level-2 products since they are contaminated by noise. Although this issue can be alleviated by the empirical destriping method or spatial filtering, they result in potential signal distortion or signal leakage. To improve this, we propose a reconstructive filter, whose parameters are optimized by the signal-to-noise ratio. Subsequently, our optimal filter corresponding to the best signal-to-noise ratio (5.63) is used to estimate the mass changes (2002–2017) in Madagascar. Eventually, our results are compared with two reliable GRACE mascon products and other independent observations. Correspondingly, here are our major conclusions: (1) Compared with groundwater storage from the mascon products, our estimates have the highest Pearson correlation (0.5) with in situ observation and can detect the rapid increase of groundwater storage during the rainy season. (2) The Fourier spectrum analysis detects a ∼3.8-yr periodic signal in the terrestrial water storage changes in Madagascar, which is contributed from the interannual precipitation driven by climate factor (Indian Ocean Dipole) and the aliasing error for imperfect GRACE pre-process. Our work introduces an effective filter for processing GRACE Level-2 data and presents novel insights into mass changes in Madagascar.

It is significant to determine the refined Moho topography for understanding the tectonic structure of the crust and upper mantle. A novel method to invert the Moho topography from the on-orbit gravity gradients is proposed in the present study. The Moho topography of Tibet is estimated by our method, which is verified by previous studies. The research results show that: (1) the deepest Moho of Tibet, approximately 70 km, is located at the western Kunlun area, where it corresponds well to that of previous publications; (2) clear Moho folds can be observed from the inverted Moho topography, whose direction presents a clockwise pattern and is in good agreement with that of Global Positioning System; (3) compared with the CRUST 1.0, our inverted Moho model has a better spatial resolution and reveals more details for tectonic structure; (4) the poor density model of the crust in Tibet may be the main reason for the differences between the obtained gravity Moho model and seismic Moho model; (5) by comparing our inverted Moho with those from previous publications, our method is correct and effective. This work provides a new method for the study of Moho topography and the interior structure of the Earth.

SUMMARY: Before inverting Moho topography, the traditional Parker–Oldenburg method requires the determination of two important hyperparameters, the average Moho depth and Moho density contrast. The selection of these two hyperparameters will directly affect the inversion results. In this paper, a new method for estimating hyperparameters is proposed which is used to improve the Parker–Oldenburg method. The new method is improved by using simulated annealing to accurately estimate the average Moho depth and Moho density contrast based on the relationship between Moho depths and corresponding gravity anomalies at seismic control points. Synthetic tests show that compared to the improved Bott's method and the trial and error method, our method reduces the error in Moho density contrast and average Moho depth by 0.83 and 1.81 per cent, respectively. In addition, compared with the trial and error method, our method greatly improves the computational efficiency. In a practical example, we apply this method to invert the Moho topography in the northern South China Sea. The inversion results show that the Moho topography in the northern South China Sea ranges from 8.2 to 33 km. The root mean squared error between our Moho topography and the seismic validation points is 0.94 km. Compared with the CRUST 1.0 model, our Moho topography is more accurate.