Evaluating Data-driven Methods for Predicting Marine Geomagnetics from Disparate, Sparse Geophysical Data

The marine crustal geomagnetic field contains important information about rock-types, seafloor age, and tectonic history, yet for half of the global ocean, the magnetic field is not sampled at 10 km resolution. Because the global ocean is too vast and remote to directly measure the geomagnetic field everywhere in a cost-effective way, models are required to resolve marine geomagnetics. The state-of-the-art for producing global crustal geomagnetic models (e.g. Earth Magnetic Anomaly Model (EMAG)) relies on the spatial interpolation of values derived from shiptrack and aeromagnetic measurements of total magnetic intensity. These models, however, suffer from the low-dimensionality of spatial interpolation algorithms and the relatively poor estimation of uncertainty. Given the limitations of current methods for modeling the crustal geomagnetic field, data-driven approaches (machine learning (ML)) to producing global marine models capable of incorporating disparate, sparse geophysical measurements (e.g. gravity and crustal age) should be explored as part of a new paradigm for developing global geologic and geophysical models. Since ML is a data-driven approach, only the extent and quality of the available observations limit prediction accuracy. The model can also readily assimilate new data, account for poor sampling density and sampling bias, and update geomagnetic predictions with posterior quantitative uncertainties. We assess the feasibility of using the Global Predictive Seabed Model (GPSM) to predict with uncertainty geomagnetic values of interest as a step-change over geospatial interpolation in terms of grid accuracy and error estimation. We find that the total geomagnetic field, as well as reduced-to-pole magnetic potential can be predicted with a high correlation between observed and predicted values (R2 >0.7) using a knearest neighbor regression algorithm and observations from three EMAG tiles. Grids of predicted values have continuous coverages at two arc-minute resolution and error estimates at every grid cell. A 10-fold validation of the GPSM prediction yields error measurements at least twice as good as those from Kriging interpolation for the same set of observations over most of the grid coverage. Results suggest a greater role for ML in producing global geomagnetic models.