Adaptive Multiple Subtraction Using a U-Net Network Applied to the Double-Focusing Method

Abstract

Summary The source-receiver Marchenko redatuming called adaptive double-focusing method is a technique that allows removing multiple reflection events in seismic data. Its effectiveness strongly depends on the method used for adaptive subtraction. This is because the primary reflections have much stronger amplitudes than the modeled multiples, then with conventional adaptive multiple subtraction approaches, this does not works, being able to lose the signal of primary reflections. On the other hand, methods based on Deep Learning (DL) have emerged as a good alternative to use in problems like adaptive subtraction. This is showing themselves as a promising tool to remove the multiples without degrading the primary amplitudes in methods based on adaptive subtractions. In this work, we test the applicability of the adaptive subtraction based on U-Net convolutional neural networks in the double-focusing method, which treats adaptive subtraction as a supervised learning problem. We use a synthetic numerical example to evaluate its effectiveness and compare it with the result of the adaptive subtraction based on nonstationary regression. Results show that the U-net method offers better protection of primaries and also attenuates efficiently the multiple reflections.