3D-Xception-UNet: An Improved Lightweight U-Net Variant for 3D Seismic Fault Segmentation

Fault segmentation in seismic samples is a major task of structural interpretation, which is manually performed by experts. Recent deep learning based methods consider seismic samples as 3D images and then extract valuable patterns using variants of convolutional neural networks. The authors, in the study, suggest a new lightweight variant of U-Net (3D Xception UNet) for 3D samples, consisting of Xception-like blocks, to segment seismic faults. Experimental results record a remarkable accuracy, 97.31%, outperforming recent studies. Additionally, 3D Xception UNet is a potential solution for practical applications with low resources in seismic data analysis because its complexity is approximately reduced 15 times, compared to the baseline model.