Deep Learning Based Super Resolution of Gravity Data for Geophysical Exploration

Abstract

Gravity measurement from ground can be obtained by precision gyrostabilized gravimeters. However, land-based gravity measurements can be challenging in many hard-to-reach regions. By contrast, satellite-based gravity measurement is considered to be a very promising for large-scale exploration since it opens up the possibilities for geophysical study of remote regions. However, comparing with the gravity from ground, the resolution and precision of satellite-based gravity measurements is lower, which limits its application for geological explorations. The novel and effective methods in improving the low spatial resolution and precision of satellite-based gravity data are active quest. Super resolution is used for naming any technique that exploit the knowledge contained in several low-resolution image to form a high resolution. There are many applications of super-resolution, which successfully improved medical imaging systems, satellite imaging, astronomical imaging. In our works, deep learning based super resolution (SR) was adopted to improving the low resolution and precision of satellite-based gravity data. Satellite-based gravity data and land-based gravity were visualized to low resolution image and high resolution image, respectively. Based on the different cropping methods, HR and LR were cropped into small patches with different size and different overlap ratio. And then modified super-resolution residual network (SRResNet) were trained by paired HR and LR small patches; and then evaluated using average absolute difference gravity , improvement (), Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). The effectiveness of modified SRResNet of trained models on small patches of different size and different overlap ratio was also investigated and confirmed.