Greedy strategies for data-adaptive shot selection

Reverse time migration (RTM) is a popular seismic imaging technique. Unfortunately, RTM computational cost scales linearly with the number of shots migrated, where that number can run into hundreds of thousands. One can mitigate this challenge computing a smaller number of linear combinations (or encodings) of shot records, and then migrate blended shot records. Although the shot-encoding approach can reduce RTM's cost significantly, the final image inevitably contains cross-talk effects introduced by the encoding process. Alternatively, blind decimation of the shots avoids this cross-talk issue. However, one could easily conceive of scenarios where decimation is not good enough. For instance, consider a model where most of the "interesting" features are localized in a small region of interest. In this paper, we introduce two adaptive, data-driven strategies for shot selection in RTM. Informally, given a budget parameter B, our high-level goal is to deduce the optimal subset of at most B shot locations that will produce an acceptable image of a given (user-specified) region in the subsurface. This problem is computationally intractable, and hence we propose two heuristics based on greedy optimization techniques. We report the results that demonstrate improvement over conventional decimation in terms of image quality and computing performance.