Convolutional neural network-based reverse time migration with multiple energy
Shang Huang, Daniel O. Trad
Reverse time migration (RTM) with multiples has the advantage that it can handle steeply dipping structures and offer high-resolution images of the complex subsurface. However, there are some limitations on the initial model chosen, aperture illumination and computation efficiency. Reverse time migration has the dependency on reliable initial velocity models. If the input background velocity model is smoothed or not accurate, the RTM result image will have poor performance. RTM with multiple energy (RTMM) can help improve the illumination but will generate crosstalks because of the interference between different orders of multiples. One solution is to apply least-squares reverse time migration, which updates the reflectivity and suppresses artifacts through iterations. However, the accuracy still depends heavily on the input and the computation time is costly. We proposed a method based on a convolutional neural network (CNN) that behaves like the filter in LSRTM but is more efficient. This approach can learn patterns or features from geological structures and predict reflectivity by some smoothed velocity models through a modified residual U-Net, trained to improve the quality of RTM images. Numerical experiments show that RTMM-CNN can recover major structures and thin layers with high resolution and accuracy compared with RTM-CNN.