Deep Learning in Seismic Interpretation: Federated Advances in Salt Dome Segmentation
Researchers have developed FedSaltNet, a federated learning framework designed for salt dome segmentation in seismic data. This framework addresses data sovereignty and scarcity issues by enabling collaborative interpretation without sharing raw data. A key finding is that a simple U-Net architecture significantly outperforms a more complex ResNet-18 U-Net in data-constrained federated settings, achieving a 166% higher average Intersection over Union (IoU). Additionally, a novel Foreground-Weighted aggregation strategy improved IoU by 4.0% over conventional federated methods, demonstrating its effectiveness in handling class imbalance and data heterogeneity. AI
IMPACT Establishes the viability of federated deep learning for collaborative subsurface interpretation, potentially accelerating exploration and safety decisions.