<p>The identification of seismic faults plays a crucial role in geological exploration and resource development. However, faults typically manifest as small targets that occupy only a minor proportion of the data. This small-target characteristic, coupled with data imbalance, presents significant challenges for accurately identifying seismic faults. The proposed SCAFaultNet model introduces a SCA module to address fault identification, effectively enhancing the model’s ability to recognize small fault targets through the fusion of spatial, channel, and axial attention mechanisms. The MLFF block combines feature maps at different scales with high-level semantic features, guiding the network to focus on both shallow and deep features for more comprehensive feature perception. To further address data imbalance, an optimization strategy based on a combined loss function is designed, integrating BCE loss with Dice loss to effectively improve the model’s accuracy in recognizing small fault targets. Experimental results demonstrate that SCAFaultNet achieves a precision of 0.8782, an intersection over union (IoU) of 0.7748, a Dice coefficient of 0.8724, and a recall rate of 0.8676 on the synthetic dataset, outperforming other models. On the field seismic datasets, SCAFaultNet successfully captures more small fault features, further validating its advantages in small target fault identification tasks. These findings advance seismic interpretation methodologies by providing novel insights into small target characterization in geophysical data.</p>