Distributed acoustic sensing data enhancement using an iterative dictionary learning method

Abstract

Distributed acoustic sensing (DAS) has emerged rapidly in the past decade because of its superb features in sensing the elastic wavefield via a low-cost, high-density, and high-durability manner. The compromise for the unprecedentedly high resolution of DAS is the noise effect. There exists a mixture of many types of noise, including but not limited to random ambient and strong amplitude noise. To tackle the various types of challenging noise, we propose a novel denoising framework based on the dictionary learning scheme. Dictionary learning is comparable to sparse transforms like wavelet and curvelet but outperforms all the alternatives by adaptively learning the basis functions for sparsifying seismic data. Instead of applying dictionary learning in a traditional way as widely reported in the literature, we apply a robust and sophisticated way to real DAS data so that we can best utilize the feature-learning advantages of dictionary learning without sacrificing the signal-leakage problems in traditional denoising methods, especially when it comes to very complicated and noisy DAS datasets.