Data-driven identification of individual scattering structures

What is it about?

In the age of machine learning and data science, seismology, like many other scientific disciplines, experiences a rapid transformation, providing unique opportunities to overcome conventional barriers. Following this new mentality of data-driven research, we explore the possibility of isolating individual diffractions directly in the data domain by employing physically guides image segmentation. With physics-based classifiers that express the continuity of wavefields, we demonstrate that a successful automatic detection, characterization, and segmentation of even very faint diffracted signals can be achieved in highly scattering environments, which often reveal a high degree of wavefield complexity.

Featured Image

Why is it important?

Diffracted signals provide unique opportunities for imaging and inversion, because they are caused at small-scale highly localized subsurface features that are of interest to an interpreter. The presented scheme of physics-guided image segmentation enables the rapid extraction of collective information of diffracted wavefields and, therefore, provides "order" in regions of high wavefield complexity and strong interference. Grouping measurements with the same origin in depth has many important applications, like fully data-driven yet highly accurate time migration, or focusing constraints in wavefront-tomographic inversion.

Perspectives