AttentionFaultFormer: An attention-enhanced 3D CNN & transformer model for seismic fault detection

In seismic exploration, accurately identifying faults is fundamental for interpreting seismic data. Due to the highly sparse distribution of faults in the seismic volume and the heavy reliance on local features to reconstruct fault segmentation masks, predicting faults using 3D Vision Transformers (ViT) remains challenging. We propose a novel attention-enhanced 3D U-shaped CNN and ViT hybrid model, called AttentionFaultFormer, specifically for fault detection tasks. AttentionFaultFormer includes a transformer-based encoder, a residual-based decoder, and attention-enhanced skip connection. The skip connection integrates local feature extraction from convolutional layers with channel and spatial attention mechanisms to enhance the representation of detailed fault features. Furthermore, based on the geometric structure of faults, we designed the Multi-Axis Striped Convolution Attention (MASCA) module and incorporated it into the shallow skip connections to help the model delineate continuous long faults. We applied our model to three field seismic datasets (F3, Kerry3D, and Thebe). Comparative analysis against two CNN models and two ViT models reveals that our model predicts more continuous and interpretable faults. • We devised the AttentionFaultFormer, an attention-enhanced 3D U-shaped CNN&ViT model. • AttentionFaultFormer is specifically for the fault detection task. • Attention module with striped convolution benefits to delineate continuous faults. • Attention-enhanced skip connection boosts fault feature reconstruction. • AttentionFaultFormer effectively predicts interpretable faults in three field data.