RTM Image Conditioning Using Deep Learning

Summary

Reverse time migration (RTM) is a widely used method in seismic imaging, particularly in situations where other migration methods may not be effective, such as in complex geologies. However, traditional RTM images are affected by a strong low wavenumber noise, which tends to appear mainly at shallow depths and above strong reflectors and can mask migrated structures, making it difficult to interpret the resulting image. To reduce this noise, Laplacian filtering is often used, which adjusts the filtering parameters to preserve the characteristics of useful better signals while still effectively removing low-wavenumber noise. In this work, an alternative approach is proposed, which utilizes deep learning. Here, a residual neural network is trained on a dataset of input-output pairs to replicate the functionality of a modified Laplacian filter. This approach results in outcomes that are similar or potentially better than the conventional method.