A complete feature set for classification of seismic sources with Distributed Acoustic Sensing (DAS) in the context of long-range monitoring

Distributed Acoustic Sensing (DAS) exploits Rayleigh light backscattering to extract images of seismic wave propagation along a fiber optic in time and distance. The spatial distribution of virtual point sensors represents an opportunity to develop innovative methods for seismic event sources detection and identification. We develop in this study a method based on Machine Learning solutions for events classification. This method relies on the development of features which translate the characteristics of the signals we observe into quantities that can be processed by machine learning algorithms to achieve the source classification. Three families of features investigating temporal and spatial characteristics and similarity of the signal are proposed, such as spatial and temporal analysis of the standard deviation, kurtosis or skewness of the signal or cross-correlation and dynamic time warping characterization and enables to quantify their individual contribution. Then we use a supervised machine learning model named XGBoost to perform classification based on these developed features. We tested this approach with a dataset recorded along a 91 km-long fiber optic deployed in the Pyrenees in France. The data acquisition has been achieved using a FEBUS A1-R DAS interrogator and with the support of TotalEnergies, from August 30 to September 20, 2022. During this period, 11 earthquakes and 6 quarry blasts have been recorded. The trained model is validated using cross-validation techniques. Our Machine Learning processing chain successfully detect and classify 13 regional events from continuous background noise made by natural and anthropogenic activities. In particular, spatial features help to reduce the contribution of moving vehicles, whose presence is unavoidable along existing long-distance telecommunication fiber sections installed alongside roads. In the continuity of this study, we investigate the potential of transfer learning from geophones deployed along the studied cable to DAS data or to another fiber optic cable installed in the same area.