Gabor CNN-based improvement of tunnel seismic migration imaging and field application with domain adaptation assistance

Abstract

In tunnel seismic forward-prospecting, the accuracy of migration imaging impacts the geological interpretation of the area ahead of the tunnel face. However, the traditional reverse time migration (RTM) method, which is the adjoint of the Born forward modeling, often yields approximate estimations of reflectivity. This approximation error becomes even more pronounced in the context of small offset tunnel conditions. To address this issue, we propose a novel method for enhancing tunnel RTM imaging by leveraging Gabor Convolutional Neural Networks (CNN). In our approach, we employ a Gabor CNN that incorporates learnable parameters within the Gabor filters to extract pertinent features from tunnel RTM imaging results. By training the network with RTM images as input and the true reflectivity as labels, we enable the network to learn underlying patterns and improve the quality of the imaging. Notably, we tackle the challenge of limited labeled field data by introducing MLReal, a domain adaptation method. MLReal enhances the generalizability of the proposed network to field data by employing an inter-processing and transformation approach that aligns the target data with the synthetic dataset. This alignment allows the network to adapt to real-world field conditions, bridging the gap between synthetic training data and field applications. Extensive numerical experiments validated the superiority of the Gabor CNN, showcasing its ability to generate results closely resembling true reflectivity while outperforming LSRTM. Furthermore, a field case study is conducted in a water transmission tunnel as a practical application to verify the potential of the MLReal-assisted Gabor CNN.