Improving Seismic Fault Detection through Fine-Tuning of Pre-Trained VGG۱۹-UNet

Fault detection is a critical step in seismic interpretation, as faults influence hydrocarbon migration, reservoir compartmentalization, and geohazard assessment. Traditional methods such as coherence, semblance, and curvature attributes provide valuable insights but are often sensitive to noise, require extensive parameter tuning, and may fail in geologically complex settings. Recent advances in machine learning (ML), particularly convolutional neural networks (CNNs), have improved fault detection but typically require large labeled datasets, which are scarce in seismic studies. To address this limitation, we propose a transfer learning approach using a pre-trained VGG۱۹-UNet model fine-tuned for seismic fault detection. The workflow integrates Fault Enhancement Filtering (FEF) to improve reflector continuity and fault visibility, followed by the preparation of fault masks and extraction of ۱۲۸×۱۲۸ pixel seismic–mask pairs. Data augmentation, including flipping and Gaussian noise addition, was applied to improve generalization. Training was performed for up to ۱۰۰ epochs with a batch size of ۸, reserving ۲۰% of the dataset for validation and employing early stopping to prevent overfitting. The fine-tuned model demonstrates substantial improvements compared to the original pre-trained VGG۱۹-UNet, yielding enhanced fault continuity and fewer false detections. Results confirm that transfer learning and fine-tuning offer an efficient and reliable approach for seismic fault detection, significantly reducing the reliance on large labeled datasets and outperforming conventional attribute-based methods.

Keywords:

Seismic fault detection , Transfer learning , Fine-tuning , VGG۱۹-UNet , Pre-trained models