点击次数：

摘要：Seismic data processing often encounters the challenge of near-surface complexity. Characterization of near-surface velocity structures is important for the selection of strategies and parameters in structure imaging. Repeated processing of a dataset without characterizations could waste large computational costs. Furthermore, manual characterization is typically time-consuming especially for super-large shot gathers. In this study, we propose a dilated convolution neural network (D-CNN) method to characterize near-surface velocity structures. The seismic shot gathers are labeled as layered, gradient, and complex velocity structures, respectively. The D-CNN uses fewer convolution layers to get the same size of receptive field as conventional CNNs and mitigates the overfitting problem. The proper architecture of the D-CNN is determined by evaluating different numbers of dilated convolution layers. For mitigating the generalization problem of D-CNN, we develop a sample acquisition method to automatically generate synthetic training samples with various near-surface features, and the other part of the training samples is acquired from field data in Sichuan, China. We propose an automatic linear move out (LMO) method to process the seismic shot gathers for better performance of D-CNN. The effectiveness of the D-CNN method is demonstrated in both synthetic and real data tests by comparing it with the k-means clustering method. The comparison results show that the D-CNN method can achieve higher accuracy compared with the k-means method. By applying the D-CNN method to the field data example, we develop an efficient early waveform inversion (EEWI) method. The results show that EEWI can save more than 65% computational costs to achieve nearly the same reconstruction accuracy compared with the conventional early waveform inversion (EWI) method.

是否译文：否