Urban DAS data enhancement and coherent noise removal with deep learning

Distributed Acoustic Sensing (DAS) employs fiber-optic cables for high-resolution seismic measurements and is particularly suited for urban settings. In this study, we introduce a deep-learning-based method to remove coherent noise while simultaneously enhancing urban DAS measurements. The DAS data is recorded by an underground fiber that has been installed in the European XFEL facility in Hamburg, Germany, as part of the WAVE Initiative. By applying coherent wavefield subtraction, a well-established deterministic scheme for wavefield decomposition and enhancement, to a subset of the DAS data, we generate training data for a convolutional neural network (CNN). We train the CNN, which is based on our recently proposed DiffractioNet, to both remove coherent noise and enhance coherent waveforms. Applications of the trained neural network to unseen data with different noise levels demonstrate that the network successfully generalizes to diverse noise conditions, effectively denoising the data and enhancing complex seismic wavefields. This approach can be a valuable step towards understanding the complex seismic wavefields encountered in urban environments, which is of particular importance to high-precision particle physics research facilities.