How Deep Learning Can Help Solving Geophysical Inverse Problems

Abstract

This brief summarizes some of the main results I obtained during my Ph.D. studies at Politecnico di Milano, under the supervision of Professor Stefano Tubaro. The thesis provides contributions to understanding the advantages, and limitations, of data-driven deep learning approaches to geophysical inverse problems, with a special focus on Convolutional Neural Networks (CNNs). Exploration Geophysics aims at estimating accurate physical properties of the Earth subsurface from seismic data acquired close to the surface. Seismic data show a great variety of statistically relevant and independent patterns. I devise Deep Learning methods to solve several geophysical tasks by learning such patterns. First, I devise generative networks as a post-processing operator for refining reflectivity images. When trained on pure image datasets, these networks suffer from the lack of physical knowledge. Then, I show a different approach named Deep Priors, which are CNNs that precondition the inverse problem. In particular, I develop a scheme to interpolate seismic data. Finally, I leverage the features extraction ability of CNNs for buried landmine detection on Ground Penetrating Radar (GPR) acquisitions. While the presented methods are effective compared to the state of the art, improvements can be achieved by integrating pure data-driven algorithms within general inverse problems theory through a-priori information derived from domain knowledge.