Shot encoding with random projections

Abstract

In order to image complex geological structures, seismic surveys acquire an increasingly large amount of data. While the resulting data sets enable higher-resolution images of the subsurface, they also contain redundant information and require large computational resources for processing. One approach for mitigating this trend is blended imaging, which combines the original shot records into a smaller number of blended shots at the expense of crosstalk in the final image. Since the cost of imaging is roughly proportional to the number of shots, blended imaging directly leads to a faster imaging process. In contrast to the existing shot encoding schemes, we establish a novel connection between blended imaging and dimensionality reduction using the Johnson-Lindenstrauss lemma. We introduce three new shot encoding schemes based on random projections and evaluate their performance. Our experiments on three data sets show that our random shot encoding schemes are competitive with existing shot encoding schemes and outperform decimated shot encoding for small numbers of shots.