Fault-GSA: High Generalization 3-D Fault Detection Method Based on Sparse Annotations

Current methods for the intelligent interpretation of seismic faults rely heavily on complete synthetic annotations. However, the complexity and non-uniformity of field seismic data often limit the generalization of models trained solely on synthetic data. Given the difficulty in obtaining complete field annotations, using sparse annotations to guide the learning of fault networks has become essential. We introduce a novel 3D fault detection method named Fault-GSA, which integrates three key training components: Supervised Learning with Synthetic Data (SLSD), Noise Learning with Sparse Annotations (NLSA), and Semi-Supervised Learning with Unlabeled Data (SSLUD). Specifically, SLSD enhances the ability to process various scales of faults in geological data by integrating of a Multi-Scale Self-Attention Fusion module (MSA). NLSA improves the model’s performance on sparse data by addressing false negatives in limited labeled data. SSLUD uses a dual-teacher model to enhance the model’s generalization ability in unknown geological environments by learning from unlabeled data. Experiments show that Fault-GSA significantly improves the accuracy and continuity of fault detection, achieving higher detection rate and better adaptability across multiple work zones compared to existing methods.