A CNN-based Method to Construct a Laterally Heterogeneous S-wave Velocity Model from Spatially Windowed Surface Wave Data

Jian Shen,Zhenming Shi,Liu,Ming Peng,Dengyi Wang,Shaojun Li,Georgios P. Tsoflias
DOI: https://doi.org/10.1109/tgrs.2024.3496855
IF: 8.2
2024-01-01
IEEE Transactions on Geoscience and Remote Sensing
Abstract:The multichannel analysis of surface waves (MASW) method has been broadly used to investigate subsurface structures. The inversion of a single dispersion curve (DC) from an individual surface wave shot record can only generate a 1D velocity model that fails to represent accurately the real geological condition in the presence of lateral velocity heterogeneity. To improve imaging of sharp lateral variations along 2D profiles we propose a convolutional neural network-based windowed multichannel analysis of surface wave method (CNN-WMASW). The CNN-WMASW method uses multichannel surface wave data that are spatially windowed by Gaussian functions to extract local dispersion curves. The dispersion curves are inverted simultaneously by the CNN-based data-driven method, which maps the phase velocity to the S-wave velocity from end to end without any initial model. The relationship between adjacent dispersion curves (DCs) is captured and considered by the CNN model, which is trained on a dataset created by waveform simulation rather than the theoretical solutions of the surface wave propagating mode in the layered model. We tested the effectiveness and reliability of the proposed CNN- WMASW method using numerical examples and two field data examples. Both results show improved lateral resolution in the S-wave velocity maps retrieved by the CNN- WMASW method compared to the traditional multichannel method in which strong lateral heterogeneity is observed.
What problem does this paper attempt to address?