Leveraging DAS Ambient seismic noise interferometry and machine learning for 2D subsurface seismic imaging in an urban area

The application of ambient noise interferometry to distributed acoustic sensing (DAS) data recorded on existing telecommunication networks provides a promising opportunity for effectively imaging the urban subsurface with high resolution at local and regional scales. This approach holds significant potential for various applications, including assessing the suitability of the urban subsurface for safe utilization, such as in geothermal development, and evaluating risks associated with subsurface activities, particularly concerning geological hazards like subsidence and sinkholes. Such capabilities are essential for developing resilience strategies and mitigating potential impacts in urban environments. However, extracting coherent seismic signals from the ambient wavefield recorded by DAS in urban settings remains a challenge. One obstacle is the presence of diverse and complex noise sources, which are often unevenly distributed. These localized sources can introduce deviation into the result of ambient noise interferometry and generate nonphysical arrivals, complicating the analysis and interpretation of the results.

In this study, we present the analysis of 15 days of continuous passive DAS data recorded on a pre-existing fiber optic cable (dark fiber) spanning 11 km along a major urban road in Berlin, Germany. Our investigation reveals anthropogenic activities, predominantly traffic noise from vehicles and trains, as the primary seismic source. To retrieve Virtual Shot Gathers (VSGs), we apply interferometric analysis based on the cross-correlation approach. Before stacking, we design a selection scheme to identify high-quality VSGs, thereby optimizing the resulting stacked VSG. Then, Multichannel Analysis of Surface Waves (MASW) is applied to derive 1D shear-wave velocity models across successive array segments. We construct a 2D velocity model of the subsurface through the concatenation of individual 1D velocity models obtained from overlapping array subsections. This expansion into 2D necessitates automatically identifying high-quality VSGs, achieved through unsupervised learning methods such as clustering. This process is crucial for excluding transient incoherent and localized noise sources during selective stacking. To implement clustering, we initially reduce the dimensionality of the VSGs using principal component analysis. We then cluster the features within this reduced-dimensional space. Finally, we stack the VSGs in each cluster and select the best-stacked VSGs. We initially test the clustering algorithm on synthetic VSGs before applying it to DAS ambient noise field data to ensure its reliability and effectiveness in real-world scenarios.

The clustering results reveal distinct groups of VSGs that demonstrate consistent patterns across synthetic and field DAS datasets. These distinct groupings offer valuable insights into the temporal variations in human activities and allow a better understanding and interpretation of the recorded DAS ambient noise data, enabling the identification of viable ambient noise signals for further processing. Ultimately, this approach enhances the accuracy of dispersion measurements, enabling improved subsurface imaging in urban areas.