Near-surface geophysics addresses challenges like groundwater monitoring and infrastructure development in urban areas. Traditional seismic surveys, typically employing weaker sources like sledgehammers, are limited in their depth of investigation due to the absence of low frequencies. Conversely, low-cost passive vehicle sources hold promise due to their capacity to generate valuable low frequencies, thus enhancing the depth of investigation. However, extracting subsurface responses (Green’s functions) from vehicle noise presents a non-trivial task, owing to the dynamic nature of these sources, characterized by high temporal variability. Conventional Green’s function estimation methods rely on averaging source effects from noise cross-correlations. They necessitate lengthy survey durations, rendering them impractical in dynamic urban environments where source variability is pronounced. In this study, we depart from conventional stacking approaches. Instead, we train symmetric autoencoders (SymAE) to learn the capability of disentangling subsurface information, which is coherent across multiple noise gathers, from the remaining source effects. We focus on a controlled experiment aimed at detecting a deep void, utilizing measured noise from a stationary vehicle with its engine turned on. A conventional MASW survey with a weight-drop source failed to yield reliable images of this void, primarily due to the absence of low frequencies. We demonstrate that SymAE holds the capability to extract backscattered surface waves attributed to the void, utilizing nearly half the recorded noise compared to that required for cross-correlation stacking approaches. Our unsupervised approach relies on the network’s capability to generate virtual seismic data through high-dimensional interpolation.

This content is only available via PDF.
You can access this article if you purchase or spend a download.