Up/down and P/S decomposition of DAS-VSP Data using multi-task deep learning method

Distributed acoustic sensing (DAS) has emerged as an important seismic observation method due to the advantages of low cost, easy deployment and high-density spatial sampling. The integration of DAS recording with vertical seismic profiling (VSP) has become increasingly prevalent across various applications of borehole geophysics. In seismic imaging and inversion of VSP data, it is essential to decompose the up/down-going and P/S-wave modes as necessary preconditioning steps. However, conventional wavefield decomposition methods based on physical principles face significant challenges when handling single-component DAS-VSP data. We resort to a deep learning (DL) method to address this mode decomposition problem. In the framework of supervised learning, the labeled data are constructed through elastic forward modeling aided by the Hilbert transform and the Helmholtz decomposition. Then, multi-task convolutional neural network (CNN) models are trained for one-pass or two-pass operation of up/down and P/S decomposition. In order to process field data, we rely on transfer learning to improve the generalization ability of the trained network models. Examples with synthetic and field walk-away DAS-VSP data examples demonstrate the effectiveness of the proposed method in wavefield decomposition for the borehole axial strain data.