FaultEdgeFormer: An Edge Enhanced Transformer Model for 3D Seismic Fault Detection

Fault detection plays a critical role in interpreting seismic data. Deep learning methods, particularly convolutional neural networks (CNNs), have demonstrated strong potential in fault detection tasks. However, current CNN-based methods face two key challenges. One is that the optimization of CNN kernels begins with randomly initialized weights, which restricts these kernels’ capability to emphasize fault related features during the early training process. The other is that existing CNN-based architectures are unable to learn global contextual information due to the fixed kernel size and limited receptive fields of convolutional operations. To overcome these challenges, we propose an edge enhanced transformer model to improve 3D fault detection performance (FaultEdgeFormer). First, faults can be characterized as edges within the seismic data. Traditional edge detection operators, like Sobel kernels, explicitly calculate gradient information for edge features. Based on this observation, a trainable Sobel convolution is proposed for the first layer of the network. The predefined weights in the trainable Sobel convolution can adaptively enhance edge information, constraining the network to focus on fault related features. Second, an improved Swin transformer block is proposed to extract long-range dependencies through the attention mechanisms, while integrates the locality of convolutional operations. This improvement is achieved through two primary modifications: introducing convolutional projections into the window-based self-attention mechanism (W-MSA) and incorporating depthwise convolutions within the multilayer perceptron (MLP). Experimental results demonstrate that the proposed FaultEdgeFormer outperforms both CNN-based methods and Transformer-based methods on several synthetic and field seismic datasets, delivering more accurate fault predictions with better continuity and resolution.