Convolutional Neural Network-Based Separation and Interpolation of Regional and Residual Gravity Anomalies for Application to Subsurface Structure Delineation in the Gongola Basin, Nigeria

Abstract

The Gongola Basin, a frontier sub-basin of the Benue Trough with recognized hydrocarbon potential, was selected because its gravity data are relatively sparse and noisy, limiting reliable structural interpretation. This study developed a convolutional neural network (CNN)-based framework integrating Fast Fourier Transform (FFT) filtering and U-Net interpolation to separate and reconstruct regional and residual gravity anomalies. FFT filtering effectively isolated long-wavelength regional fields from short-wavelength residual anomalies, while the CNN reconstructed residual anomaly maps with improved spatial continuity and preservation of structural boundaries. The resulting maps delineated sedimentary depocentres, basement highs, and fault-controlled structures associated with prospective hydrocarbon systems. Uncertainty analysis indicated stable predictions with minimal systematic bias. The proposed framework outperformed conventional interpolation approaches by preserving subtle geological features and reducing smoothing effects. These results demonstrate that integrating FFT filtering with deep learning provides an effective and scalable approach for gravity interpretation and petroleum exploration in data-limited frontier sedimentary basins.