Seismic Fault Segmentation Using Unsupervised Domain Adaptation

Seismic fault segmentation plays a crucial role in geophysics, especially in the exploration and extraction of oil and natural gas. However, identifying faults from seismic data is challenging and time-consuming due to the large volumes of data, complexity, noise, ambiguity, and reliance on subjective manual identification by specialists. To address these challenges, various automated methods based on Deep Learning have been proposed, but they require substantial amounts of labeled data. Synthetic data generation has been explored as a solution to generalize models for new real data, but it faces the domain shift problem, where different datasets may have different distributions and behaviour. In this paper, we propose a novel approach using Unsupervised Domain Adaptation (UDA) techniques to mitigate the domain shift issue in seismic fault detection. We utilize a synthetic dataset known for its use in the FaultSeg3D model and a labeled real dataset called Thebe. The selected techniques include Maximum Mean Discrepancy (MMD) and Fourier Domain Adaptation (FDA). MMD aims to reduce the discrepancy between internal layer outputs in the neural network (feature-level), while FDA transfers features between input images (image-level). For the experiments, we employed a smaller version of UNet and a variant called Atrous UNet, which has shown improvements in seismic images by utilizing Dilated Convolutional layers. The results demonstrated up to a 12% improvement in Intersection over Union and a 10% improvement in F1 by applying UDA techniques. Moreover, the proposed approach achieved more continuous and fewer false positive seismic fault detections.