An interactive feature-fusion neural network for 3D fault seismic recognition

Abstract Currently, most deep-learning-based fault-recogni-tion methods rely on traditional networks, such as U-Net and convolutional neural networks (CNNs). Traditional networks excel in feature extraction because they use a large number of traditional convolution operations. However, they suffer from issues such as feature redundancy, slow model inference speed, and overfitting. This study proposes an interactive feature-fusion neural network (IFFN) for three-dimensional (3D) fault recognition. First, the IFFN adopts a U-shaped network structure, leveraging techniques such as three-dimensional (3D) depth-separable convolution, mixed-hole convolution, and upsampling to address feature redundancy caused by conventional con-volutions. Dropout layers are incorporated into each convolution block to suppress overfitting and enhance the generalization ability of the network. Second, the IFFN introduces an attention mechanism to emphasize the relationship between features, considering the inherent relationships among fault structures. Finally, the IFFN employs an interactive feature-fusion layer that effectively combines low- and high-level features to mitigate feature loss during the convolution process, ensuring accurate feature extraction. The flexibility of the interactive feature-fusion layer allows customization according to specific task require-ments by adjusting the number of layers in the encoder and decoder. Experimental evaluations conducted on the FaultSeg3D dataset, which is enhanced to address overfitting, demonstrates that the IFFN achieves superior performance. The IFFN exhibits a 10% reduction in network complexity with only 2.76 million parameters compared with that of traditional U-shaped networks. Furthermore, it attains an impressive intersection-union ratio of 86.4%. The proposed 3D fault identification method in this study provides a valuable approach for improving fault-recognition for earthquake interpretation.