Seismic fault detection based on a multi-scale pyramid attention mechanism

Fault identification is a critical task in seismic structural interpretation and plays an essential role in oil and gas exploration and development. Despite the remarkable advances brought by deep learning in improving the efficiency and accuracy of fault detection, existing approaches still suffer from false positives and missed detections, particularly for small-scale faults and fault intersection zones under high-noise and geologically complex conditions.To address these challenges, this study proposes an improved three-dimensional network, MS-HPANet, built upon the UNet architecture. The model integrates multi-scale pyramid attention and hybrid pooling attention to enhance feature representation across multiple spatial scales. Specifically, multi-branch dilated convolution and channel attention mechanisms are introduced in both the encoder and decoder to capture fine-grained fault details and global structural information simultaneously, thereby strengthening the model′s sensitivity to small-scale faults. Moreover, directional pooling and grouped attention modules are incorporated into the skip connections to model directional fault features, guiding the network to focus on responses aligned with major fault orientations and effectively suppressing background noise interference. Experimental results demonstrate that the proposed optimization strategies enable MS-HPANet to extract fault structures accurately even under severe noise conditions, significantly improving the robustness, consistency, and reliability of fault identification.