Machine Learning Inference of Random Medium Properties
Kai Gao,Ryan T. Modrak
DOI: https://doi.org/10.1109/tgrs.2024.3367541
IF: 8.2
2024-03-02
IEEE Transactions on Geoscience and Remote Sensing
Abstract:Earth materials are heterogeneous across a range of spatial scales, but the resolvability of small structures is limited by sparse data coverage, noise, bandlimitedness, and other difficulties. In practice, heterogeneities below a certain size cannot be recovered from seismic data except through statistical medium descriptions, which even then can be difficult to uniquely determine. To improve the characterization of such heterogeneities, we develop a novel supervised machine learning (ML) model that provides insight about the recoverability of statistical medium properties from elastic waveform data and succeeds despite cycle-skipping and other challenges well known from elastic waveform inversion. We demonstrate the approach using random media generated by superimposing self-affine random variations on homogeneous and layered background structures. After training on sparsely-recorded, high-frequency waveforms from hundreds of different random medium realizations, we show the ability of our ML model to recover correlation lengths and other statistical properties of interest to near-surface and crustal seismology, among other fields. For frequency passbands and spatial offsets encountered in seismology, Gaussian correlation lengths and the amplitude of the random variations relative to the background model are recovered even in challenging scenarios involving unknown medium parameters, complex crustal structures, and low signal-to-noise ratio. In comparison, von Kármán correlation lengths, which are related to larger-wavelength variations of the medium than Gaussian correlation lengths, are not as well recovered. These results provide one of the first and most systematic investigations of the recoverability of statistical properties of heterogeneities below the resolution limit of deterministic seismic tomography, and suggest practical ML strategies for high-frequency waveform seismology.
imaging science & photographic technology,remote sensing,engineering, electrical & electronic,geochemistry & geophysics