Refined Bathymetric Prediction Based on Feature Extraction of Gravity Field Signals: BATHY‐FE

• Published in: Journal of geophysical research. Machine learning and computation Vol. 1; no. 3
• Main Authors: Annan, Richard Fiifi, Wan, Xiaoyun
• Format: Journal Article
• Language: English
• Published: 01.09.2024
• Subjects: bathymetry; convolutional neural network; deflection of the vertical; feature extraction; gravity anomaly; vertical gravity gradient
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2024JH000205

The resolution of altimetry‐derived gravity signals renders traditional bathymetry inversion methods inadequate for detecting short‐wavelength bathymetric features. This study presents a new global seafloor topography model by merging regional models constructed via a deep learning approach recently developed by the authors. Feature extraction techniques were used to refine each regional bathymetric model, resulting in sharpened features that are fuzzily revealed (or not seen) in global bathymetric models such as GEBCO_2023 and SRTM15+V2, especially in areas close to the poles. This is proof that, with regard to current gravity field products, conventional bathymetry inversion methods are weaker at generalizing to uncharted locations of the global ocean floor. At test points, the global seafloor model has mean error and error standard deviation of 1.75 ± 81.15 m. The refined seafloor topography can be used to supplement existing seafloor maps, especially in the polar regions. It is useful for studying seafloor geography, and also for studies in which seafloor ruggedness is required. Plain Language Summary It is said that less than a quarter of the global ocean floor has been properly surveyed via shipboard echo‐sounding. Therefore, the need for predicted depths continues to exist. To satisfy this need, researchers have resorted to depth estimation techniques using the relationship between Earth's topography and its gravity field. Since the past few years, the mathematical formulas used for inverting bathymetry have yielded incremental accuracy improvements. This is evident in the yearly iterations of GEBCO bathymetric grids. The authors recently proposed a new technique based on convolutional neural network (CNN) that used a combination of four gravity field signals. Its applicability was demonstrated in the Gulf of Guinea. Here, with almost the same data sets used by the GEBCO team, we use feature extraction technique to refine the CNN approach on a global scale to derive a new bathymetric model. The resultant model reveals more seafloor features which are fuzzy or not seen in existing seafloor models. It is superior to existing models in most marine regions, especially at the polar regions. This study is proof that traditional methods of bathymetry inversion may have reached their accuracy limits. Key Points We extract deeply learned bathymetric features from gravity field products to predict global seafloor topography The new bathymetric product is superior to existing models in most marine regions, especially at areas close to the poles Results show that traditional methods of geodetic bathymetry recovery might have reached their prediction accuracy limits