Unsupervised DAS Denoising
Ground-truth-free noise suppression for DAS-VSP
About
This project implements an attention-based deep-image-prior workflow for suppressing strong background noise in DAS-VSP data without pre-labelled clean targets (Cui et al. 2025). The pipeline combines patching, kurtosis-based feature selection, unsupervised denoising, and reconstruction.
Why it matters
Distributed acoustic sensing provides dense measurements in harsh borehole environments, but its sensitivity also captures strong background noise. A ground-truth-free method avoids the need to create matched clean/noisy training pairs for each survey.
Key result and audience
The public study demonstrates the workflow on synthetic data and field profiles from Utah FORGE and Groß Schönebeck, emphasizing transfer across surveys. The code is relevant to DAS processors, VSP researchers, and unsupervised-learning practitioners.
Public links
This public SEEM-KFUPM repository is a fork of the authors’ upstream implementation; GitHub preserves that provenance.