Seminar

Abstract: 

[nur in Englisch verfügbar]

In seismic processing Radon demultiple and trim statics are applied as post-migration gather conditioning and improve subsequent seismic inversion and interpretation. We present an end-to-end, deep learning-based alternative to popular classical algorithms for these demanding tasks. Our method is unified in the algorithmic layout, e.g., the neural network architecture or pre-processing steps, but uses tailored synthetic training data for the specific tasks. In the case of the demultiple process, the synthetic data models the conversion of Common Depth Point (CDP) gathers to their multiple-free counterparts. Similarly, for trim statics, our generated data models the transformation of CDP gathers to CDP gathers with aligned primary energy. We observe an excellent performance of our approach when applying the trained networks to complex field data. The application of our parameter-free method is particularly simple because the networks are solely trained on synthetics without the need for data-specific adjustments. Thus, the resulting deep learning-infused demultiple and trim statics operators can be interpreted as image-to-image transformations which directly produce the desired output.