Unsupervised Dimensionality Reduction and Noise Suppression in Marine Gravity Data in Oron-Calabar Waterway, Southern Nigeria

Abstract

Marine gravity data acquired along the Oron-Calabar waterway, southern Nigeria, a sedimentary-dominated and structurally complex sector of the Gulf of Guinea, are often affected by high-frequency noise and signal superposition, limiting the resolution of basement architecture required for basin analysis. This study applies an unsupervised machine-learning framework for dimensionality reduction and noise suppression to enhance the Free-Air Gravity Anomaly (FAGA) field derived from marine gravimeter measurements. The observed gravity measurement ranged from 978,000 to 978,380mGal. Both Linear and nonlinear dimensionality reduction techniques were evaluated, and the results show that a Convolutional Autoencoder, compressing the gravity data into a 12-dimensional latent space, provided the most effective enhancement by reducing high-frequency noise amplitude by 42% and improving the signal-to-noise ratio by 8.5dB. The enhanced gravity field significantly improved structural resolution, allowing clearer identification of residual anomalies ranging from −13 to +13mGal associated with a buried igneous ridge and a more sharply defined intra-basin fault, where gravity gradients increased from 8.3 to 13.7mGal after processing. Interpretation of the dominant latent components indicates that they capture regional long-wavelength gravity trends (−8 to +5mGal) linked to crustal thinning, superimposed localized high-amplitude anomalies (+15 to +30mGal) related to intrusive basement bodies, and intermediate-wavelength signals (−10 to +8mGal) corresponding to sedimentary depocenters. These results demonstrate that machine-learning-based dimensionality reduction provides a robust and objective approach for enhancing noisy marine gravity data and improving tectonic interpretation and resource assessment within the Oron–Calabar coastal corridor.