A Self-Supervised Deep Learning Approach for Blind Denoising and Waveform Coherence Enhancement in Distributed...

Area of Science:

• Geophysics and seismology
• Signal processing
• Machine learning

Background:

• Fiber-optic distributed acoustic sensing (DAS) offers dense, meter-scale seismic measurements.
• DAS technology enables monitoring in challenging environments like urban, glaciated, and submarine settings.
• Traditional noise-handling methods are insufficient for DAS data from new environments.

Purpose of the Study:

• To develop a novel denoising technique for DAS data.
• To address challenges posed by spatially incoherent noise in DAS recordings.
• To improve the applicability of DAS in seismological analysis.

Main Methods:

• A self-supervised Deep Learning approach is proposed.
• The method leverages the spatial density inherent in DAS measurements.
• No assumptions are made about noise characteristics beyond spatio-temporal incoherence.

Main Results:

• The Deep Learning approach successfully removes spatially incoherent noise.
• Demonstrated effectiveness on both synthetic and real-world DAS data.
• Performance is robust even when signals are significantly below the noise level.

Conclusions:

• The proposed method enhances DAS data quality for seismological applications.
• It facilitates the incorporation of DAS into conventional data processing workflows.
• This technique opens new possibilities for analyzing seismic signals in complex environments.