Contribution of spatial features for classifying seismic events from Distributed Acoustic Sensing (DAS) data streams

Abstract

Distributed Acoustic Sensing (DAS) is an acoustic sensor instrument that turns a single optical fiber into a dense array of thousands of equally spaced seismometers. Geoscientists and companies have an interest in investing in DAS technologies for better understanding the Earth by observing natural and anthropogenic seismic events or assisting in large infrastructure monitoring with low installation and maintenance costs. However, this type of instrument generates a significantly larger amount of data than conventional seismometers, data that can be complex to store, exploit and interpret.Several strategies for classifying seismic events from fiber-optics DAS data exist in the scientific literature. Conventional approaches rely on the use of features that describe the waveforms and frequency content of signals recorded individually at virtual stations along the fiber; they do not integrate the spatial density of information permitted by DAS. Several studies on dense seismological arrays have introduced similarity measures between the different time series data such as cross-correlations, dynamic time warping (DTW) or compression-based dissimilarity.This study aims to quantify the contribution of spatial features for DAS data streams classification. We have chosen to explore spatial features related to both standard statistical measures (e.g., spatial mean, median, skewness, kurtosis), and advanced signal processing measures (e.g., auto-correlations, cross-correlations, DTW). This set of measures allows enriching a list of already used time series features which includes waveform, spectrum and spectrogram. A Random Forest (RF) classifier is then trained, and a Random Markov Field (RMF) algorithm is used after classification to account for redundant spatial and temporal information.The evaluation of the spatial feature contribution is based on the output of the RF-RMF processing chain. Anthropogenically-triggered seismic data were acquired at the FEBUS Optics test bench. We consider five seismic sources: footsteps, impacts, excavators, compactor and fluid leaks. A class of noise is added as the RF-RMF algorithm is developed for processing DAS streams inherently affected by  noise.  Accurate  classification results can be obtained using only time features, and ongoing tests show a 2% increase in the correct classification rate with the use of both time and spatial features. The improvement allowed by the addition of spatial features is tangible but limited on our test dataset, but we think it should have a much greater impact on natural sources and we will discuss this perspective.