Reconstruction Of Sparsely Sampled Seismic Data Via Residual U-Net And Earthquake Stress Drop For A Circular Crack In An Anisotropic Medium



Journal Title

Journal ISSN

Volume Title



Reconstruction of sparsely sampled seismic data is critical for maintaining the quality of seismic images when significant numbers of shots and receivers are missing. I present a reconstruction method in the shot-receiver-time (SRT) domain based on a residual U-Net machine learning architecture for seismic data acquired in a sparse 2D acquisition. The SRT domain retains a high level of seismic signal connectivity, which is likely the main data feature that most reconstruction algorithms rely on. I develop an “in-situ training and prediction” workflow by dividing the acquisition area into two separate subareas: a training subarea for establishing the network model using regularly sampled data, and a testing subarea for reconstructing the sparsely sampled data using the trained model. To establish a reference base for quantifying the reconstructed data, I devise a baseline reference using a small portion of field data. For a marine dataset, the average correlation between the true answers and the reconstructed missing traces is over 85% for regular missing traces, and over 77% for random missing traces, which is higher than that by the compressive sensing method (74%). The results show that my method can effectively learn the features of seismic data in the training process. My second line of research is on earthquake stress drop for a circular crack model in an anisotropic medium. The widely used circular crack model to infer earthquake stress drop assumes an isotropic background medium. Here, I study the effect of anisotropy on stress drop for a circular crack model. I obtain the static relationship between stress drop and slip for a circular crack model in an arbitrarily anisotropic medium. I then calculate the far-field waveforms to infer stress drop by assuming that the crack ruptures circularly and reaches the final displacement determined by the static solutions. Therefore, I can estimate the stress drops based on far-field waveforms. For a transversely isotropic medium with about 18% anisotropy, the misfit of inferred stress drop could be up to 43% if ignoring the possible anisotropy information. Hence, it is necessary to consider the medium anisotropy in stress drop estimation.



Reconstruction, Convolutional neural network, Stress drop


Portions of this document appear in: Tang, Shuhang, Yinshuai Ding, Hua-Wei Zhou, and Heng Zhou. "Reconstruction of sparsely sampled seismic data via residual U-Net." IEEE Geoscience and Remote Sensing Letters 19 (2020): 1-5.