Aug 2026· Journal of Geophysical Research - Solid Earth· Vol 131· 2 citations· 67 references
TL;DR
Five independent research groups are presented, each incorporating the fundamental physics of downward continuation within advanced data‐driven modeling frameworks, and each provides a substantial improvement in depth accuracy at shallow seamounts and deep trenches.
Abstract
Only one quarter of the global ocean floor has been directly surveyed; the remaining three quarters are inferred from satellite altimeter‐derived gravity data using techniques developed in the 1990s. These classical methods correlate gravity anomalies with known depths and extrapolate bathymetry in unsounded regions. However, spatial resolution has remained limited to ∼ ${\sim} $ 12 km full wavelength due to the smoothing effects of upward continuation. Two recent advances are now reshaping this field: the Surface Water and Ocean Topography mission has increased radar altimeter range precision by a factor of four, effectively doubling the two‐dimensional resolution, and modern machine learning (ML) approaches allow for scalable, high‐fidelity inversion. The authors of this paper organized a workshop, initially at the Technical University of Denmark, that brought together experts in marine gravity, seafloor mapping, and ML to tackle this challenge. Here, we present results from five independent research groups demonstrating consistent and substantial improvements (23%–46%) in bathymetric prediction accuracy, each incorporating the fundamental physics of downward continuation within advanced data‐driven modeling frameworks. In addition to improving the accuracy and resolution of the predicted bathymetry, each of the five models provides a substantial improvement in depth accuracy at shallow seamounts and deep trenches.
Despite millions of ship soundings, bathymetry from satellite derived gravity is still necessary to fill in approximately three quarters of the global oceans. The methods used to carry out this inference have not changed significantly since the 1990s; however new methodology involving Machine Learning (ML) improves the bathymetric predictions considerably. Here we utilize five independent ML models from a workshop at the Technical University of Denmark (DTU). We highlight the benefits achieved by either (a) inclusion of a new highly‐accurate gravity field from the Surface Water and Ocean Topography (SWOT) satellite (∼22% improvement in spatial resolution), or (b) highly flexible ML methods capable of inferring bathymetric regimes not previously possible. By taking advantage of these five independent models we can determine regions of high confidence in our bathymetric inversion as well as regions with challenging conditions. Building on model prediction confidence and features revealed from the dense gravity field obtained by SWOT, we present a global features‐of‐interest map that, if mapped by ship soundings, would yield the largest improvement of the global bathymetry. These features are primarily located in regions with sparse multibeam coverage and associated with large gravity anomalies in the marine gravity field, indicating a potential presence of complex seafloor topography.
B. Nilsson, B. Phrampus, F. Salajegheh et al.· Journal of Geophysical Resea...· 1 citation
Uniform density-contrast assumptions in gravity-derived bathymetry produce substantial systematic errors. This problem stands out in regions with strong lateral density variation, such as the central–northern South China Sea. Conventional constant or simple vertically varying density models fail to capture these complexities. To overcome this limitation, a spatially varying density-contrast field is constructed by integrating multi-source geophysical data (crustal, gravity, and bathymetric data) using a back-propagation (BP) neural network. This field is incorporated into an adaptive Parker–based inversion, yielding a high-resolution bathymetric grid with Root Mean Square (RMS) improvements of 2.6 m over the constant-density approach, together with the smallest systematic bias. The most significant gains occur in shallow reef-dominated waters (0 to −1500 m), where relative RMS reductions reach approximately 9%. By coupling neural network-derived density modelling with physically rigorous inversion, the approach overcomes limitations of uniform-density assumptions while retaining interpretability, providing an efficient and reliable approach to high-precision seafloor mapping in geologically complex regions.
Chunhong Wu, Chuang Xu, Laiyong Song et al.· Geophysical Journal Internat...· 0 citations
Earth system reanalysis datasets are foundational for weather and climate research and provide the gridded training data used by most machine learning weather prediction systems. Here we show results from a prototype system that suggest that machine learning models trained only on Earth system observations can potentially be used to generate multi-decade global reanalyses without using physics-based numerical models. The resulting gridded fields capture large-scale atmospheric structure and variability across multiple timescales, while exhibiting signs of physical coherence in several key dynamical diagnostics. Evaluations of the prototype against held-out independent atmospheric observations indicate that the root mean square vector error of upper-level winds is close to that of ERA5 when compared at a consistent resolution, and that the standard deviation of the error at the surface is between that of 4th- and 5th-generation ECMWF reanalyses (ERA-Interim and ERA5). Furthermore, while traditional reanalysis production is computationally expensive, typically taking several years to produce, the reanalysis presented here was generated during the course of a single working day. These results suggest that observation-trained machine learning models offer a promising new approach for reanalysis production from observations alone.
Peter Lean, E. Pinnington, P. Laloyaux et al.· 0 citations
Despite significant advances in observational systems such as the global Argo array of autonomous profiling floats, the spatiotemporal coverage of subsurface ocean observations remains limited compared to the dense data provided by satellite platforms. This study develops a data‐driven framework to reconstruct synthetic profiles of upper ocean temperature and salinity by training a self‐attention‐based neural network with satellite‐derived sea surface height (SSH) and sea surface temperature anomalies, using 17 years of collocated Argo float measurements. Daily synthetic profiles for the upper 650 m of the Philippine Sea were generated for the entirety of 2010 and assimilated into a regional ocean model via 4D‐Var data assimilation. Results show overall improved effectiveness of state estimation when synthetic profiles are utilized. Diagnostic variables like temperature, salinity, SSH, horizontal velocity all show improvement. Synthetic profiles of subsurface temperature had an overall positive impact on SSH analysis and forecast, especially in regions east of the Luzon Strait and the southern domain influenced by the North Equatorial Current. In the vertical range of 150–600 m, the impact of synthetic profiles on various observations was promising, leading to substantial reductions in the analysis and forecast error.
Guangpeng Liu, B. Powell· Journal of Advances in Model...· 0 citations
Mapping seafloor topography is of great significance for deep-sea navigation, marine resource exploration, and aquatic ecosystem conservation. Advances in bathymetric surveying technology have progressively enriched our understanding of the oceans. However, due to the high cost and low coverage of ship-based surveys, extensive regions of the global ocean remain unmeasured. Bathymetric prediction based on gravity anomalies remains the dominant approach for mapping the seafloor. Traditional physical methods are constrained by limitations in model assumptions, making it difficult to further improve the accuracy of depth estimation. Given their strong capability for nonlinear fitting, neural network methods have attracted growing interest in bathymetric estimation. Nevertheless, their inherent tendency to overfit often leads to poor generalization performance—particularly when the water depth distribution in the test set diverges from that of the training set, a common scenario in seafloor mapping due to ship tracks typically avoiding shallow or reef-prone areas. To address this limitation, this paper proposes a novel framework integrating traditional physical models with neural networks: the Geospatial Dual-Channel Attention Physics-constrained Network for bathymetry prediction. The bathymetric model derived from the Parker–Oldenburg method is incorporated into the network. Notably, an attention mechanism is integrated into the architecture to enhance the model’s fitting capacity. Extensive validation demonstrates the effectiveness of the proposed model, with 97.2561% of the residuals falling within \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 100$$\end{document} m.
Jia Hu, Yi-Feng Luo, Chao Wang et al.· Deutsche Hydrographische Zei...· 1 citation
Abstract The establishment of a reliable vertical reference framework remains a fundamental challenge in geodetically complex regions such as Ethiopia, where extreme topographic gradients and lithospheric heterogeneity severely compromise the predictive accuracy of Global gravity models (GGMs). This study presents a hybrid geoid modeling framework that integrates a large nationwide dataset comprising 92,433 airborne gravity observations and 22,370 terrestrial gravity points to refine GGM performance across the Ethiopian Plateau. Initial spectral analysis identified Earth Gravitational Model 2008 (GM2008) as the most suitable global baseline; however, it exhibited a significant Root Mean Square Error (RMSE) of 0.9239 m, primarily attributed to omission errors in resolving high-frequency gravitational signals. To mitigate these discrepancies, we evaluated six corrector surface architectures, ranging from classical parametric transformations to advance nonlinear machine learning algorithms. Our results demonstrate that a feedforward Artificial Neural Network (ANN) based on a multilayer perceptron (MLP) architecture provides an improved stochastic representation of the geopotential surface, achieving a 94.24 % reduction in the overall prediction error. By adaptively modeling nonlinear residuals, the MLP model refined the local hybrid geoid achieving an RMSE of 0.0532 m which represents a significant improvement compared to traditional parametric models. This research provides a scalable, methodological framework for the modernization of National Vertical Datums in similar topographically and geodynamically complex regions.
A. A. Gedamu, Sintayehu Abie Dires, Zerihun Geremew Gebresenbet· Journal of Applied Geodesy· 0 citations