Machine Learning Algorithms for High-Resolution Gravity Prediction

Namibia’s reliance on the coarse-resolution EGM2008 global geopotential model results in height conversion errors often above one meter, especially due to the lack of high-resolution terrestrial gravity data. This research analyzes Gaussian Process Regression (GPR), Support Vector Machines (SVM), Artificial Neural Networks (ANN), and Universal Kriging for the creation of high-resolution Bouguer gravity anomaly grids in the Khomas Region of Namibia. The study utilized 125 terrestrial gravity observations from the Bureau Gravimétrique International (BGI), along with SRTM elevation data and lithological density values, to train and test the models at spatial resolutions of 1.86 km and 9.26 km. Spatially Distributed Hold-Out Validation (SDHV) technique was used to check how accurate the model performed. GPR returned the best indicative accuracy (MAE = 4.67 mGal; RMSE = 5.33 mGal) and strongest agreement with Universal Kriging (r = 0.9073), while ANN exhibited overfitting tendencies at fine spatial scales and SVM produced the most conservative resolution-invariant outputs. All four techniques consistently replicated the prevailing northwest–southeast regional gravity gradient, with model divergence limited to short-wavelength spatial detail. These results show that machine learning may be used to simulate regional gravity fields and give a methodological framework to help with geoid refining and updating vertical datums in Namibia.