LAUSR

Seismic full-waveform inversion (FWI) is a powerful tool to image the high-resolution subsurface properties by fitting synthetic seismogram to observed data. To reduce the nonlinearity of FWI, a critical step is to exclude the noises and select seismic signals that are not cycle skipped for the inversion. Conventionally, the windows are created through a short-term-average-to-long-term-average ratio (STA/LTA) method and then selected by comparing the synthetic and observed data with fine-tuned controlling parameters. It works well in regional and global full-waveform adjoint tomographies, but the efficiency and accuracy become a problem when dealing with a tremendous amount of data in the exploration seismology. Here, we design a fully convolutional network (FCN) to select the time window. We manually select 24,000 high-quality training time windows from field data generated by the traditional STA/LTA-based algorithm (i.e., FLEXWIN). The FCN prediction results on test sets achieve good levels of accuracy, suggesting that the neural network provides results comparable to those manually selected ones. Our tests show that the FCN method is approximately 10,000 times faster than the conventional FLEXWIN algorithm. Moreover, the windowing results of the FCN method are more stable than those of FLEXWIN without any parameter tuning, showing the great potential for FWI. We validate the effectiveness of the window selection strategy in FWI through numerical experiments.