Emmanuel Alofe, PhD

E. Alofe¹, W. Lu², R. Malehmir³

¹Department of Earth Sciences, Uppsala University, Uppsala, Sweden

²Tetra Tech Inc,Vancouver, Canada

³Michael Baker Int'l, USA

The application of deep learning (DL) for super-resolution of aeromagnetic maps has gained traction within the past decade. However, many of these applications are focused on improving the resolution of signatures on a sparsely gridded map. This research instead focuses on improving the resolution of signatures on aeromagnetic maps acquired from different heights, using a DL model as an alternative to downward continuation (DC). Using transformer-based DL architectures, we trained a network to recognize both signal and noise patterns in aeromagnetic maps generated from varying upward-continued heights with different noise intensities. The training datasets included both ground magnetic data acquired at five locations in Sweden and synthetic data generated by a model representative of high-susceptibility bodies within a lower-susceptibility host. Preliminary findings indicate that the DL model generally performed better at resolving signatures from lower-altitude data under high noise than from higher-altitude data with equivalent noise levels. Compared with downward-continued maps, the DL model produced superior resolution and a higher peak signal-to-noise ratio (PSNR) at certain height and noise thresholds. However, these effective thresholds depend on the PSNR of the input data. Beyond these height and noise thresholds, the DL model offered no clear advantage over DC. We conclude that the resolving power of the DL-based approach and its potential superiority over the DC filter is strongly dependent on acquisition height and noise levels in aeromagnetic datasets.