loading page

Segmentation of Common Conversion Point Stacks using a Neural Network
  • +2
  • Lucas Sawade,
  • Peter Makus,
  • Ryan Manzuk,
  • Jessica Irving,
  • Stephane Rondenay
Lucas Sawade
Princeton University

Corresponding Author:lsawade@princeton.edu

Author Profile
Peter Makus
GFZ Potsdam
Author Profile
Ryan Manzuk
Princeton University
Author Profile
Jessica Irving
University of Bristol
Author Profile
Stephane Rondenay
University of Bergen
Author Profile


Receiver functions, an important tool in understanding sub-surface interfaces, can be analysed through carefully implemented neural networks. We demonstrate this approach. Previously, we introduced our receiver function tool set, Pythonic Global Lithospheric Imaging using Earthquake Recordings (PyGLImER). PyGLImER enables us to: [1] create a database of teleseismic event displacement records at worldwide seismic stations, [2] compute receiver functions from these records, and [3] compute volumetric common conversion point (CCP) stacks from the receiver functions and their conversion points. CCP stacking is a standard tool to image the subsurface using receiver functions. The CCP stacks represent rich but large, three-dimensional volumes of data that contain information about discontinuities in Earth’s crust and upper mantle. One goal of the interpretation of CCPs is the identification of such discontinuities. Automated picking routines reduce discontinuities to singular peaks and troughs, thus discarding the wealth of information available over the whole depth range, such as integrated discontinuity impedance and regional geometry. However, the obvious alternative, manual picking, is not feasible for large data volumes. Here, we explore the possibility of fully-automated segmentation of 3D CCP volumes through the application of image processing routines and machine learning to successive volume cross-sections. With our picking tool, we manually label discontinuities in CCP slices to serve as training and validation sets.We use these labeled datasets as input to train a convolutional neural network (CNN) to perform pixel-wise identifications in subsurface images. When applied to all slices of the CCP stack, the CNN outputs a fully-segmented 3D model, which furnishes quantitative exploration of subsurface discontinuity morphology. Specifically, we can investigate the thickness/width, intensity, and topography of discontinuities across continents. This information has the potential to improve our understanding of, e.g., mantle transition zone phase transitions and, therefore, mantle dynamics.