ISF-Unet: A Lightweight 3D Seismic Fault Segmentation Network with Inception Depthwise Convolution and Channel Attention Fusion

Abstract

Accurate detection of seismic faults is critical for geological exploration, reservoir characterization, and hazard assessment, yet traditional approaches remain vulnerable to noise and often rely heavily on manual interpretation. To address these limitations, a lightweight 3D seismic fault segmentation network, ISF-UNet, is developed with a central innovation in the ISF module, which integrates an Inception Depthwise Separable Convolution sub-module and a Squeeze-and-Excite Fusion sub-module. The ISF module employs a multi-branch parallel structure with a shared channel attention mechanism to capture multi-scale spatial features while maintaining efficiency with minimal parameter overhead. A composite Dice-Focal loss function is further incorporated to mitigate class imbalance and enhance segmentation performance. The network is trained with multiple data augmentation strategies to improve generalization. Evaluation on both synthetic and real seismic datasets, including F3, Kerry3D, and field data from an oilfield in the Junggar Basin, China, demonstrates that ISF-UNet achieves superior recognition accuracy and structural continuity of faults compared with conventional architectures. The lightweight design not only ensures robust feature extraction but also provides practical advantages for deployment in resource-constrained environments. These results highlight the effectiveness of integrating multi-branch convolutional operations with channel attention and tailored loss functions, offering a reliable and scalable solution for seismic fault detection.