A hybrid network for three-dimensional seismic fault segmentation based on nested residual attention and self-attention mechanism

Fault detection is a crucial step in seismotectonic interpretation and oil-gas exploration. In recent years, deep learning has gradually proven to be an effective approach for detecting faults. Due to complex geological structures and seismic noise, detection results of such approaches remain unsatisfactory. In this study, we propose a hybrid network (NRA-SANet) that integrates a self-attention mechanism into a nested residual attention network for a three-dimensional seismic fault segmentation task. In NRA-SANet, the nested residual coding structure is designed to fuse multi-scale fault features, which can fully mine fine-grained fault information. The two-head self-attention decoding structure is designed to construct long-distance fault dependencies from different feature representation subspaces, which can enhance the understanding of the model regarding the global fault distribution. In order to suppress the interference of seismic noise, we propose a fault-attention module and embed it into the model. It utilizes the weighted and the separate-and-reconstruct channel strategy to improve the model sensitivity to fault areas. Experiments demonstrate that NRA-SANet exhibits strong noise robustness, while it can also detect more continuous and more small-scale faults than other approaches on field seismic data. This study provides a new idea to promote the development of seismic interpretation.