Federated transfer learning for generalized salt body segmentation across multiple heterogeneous seismic datasets

Seismic segmentation is a key step in subsurface interpretation and hydrocarbon exploration, where precise delineation of features such as salt bodies guides reservoir modeling and drilling decisions. Progress in deep learning, particularly encoder-decoder architectures like U-Net, has improved segmentation accuracy. Yet, model development remains constrained by the scarcity of labeled seismic data and strict data privacy requirements. In this study, we propose a novel TransFed-SaltNet model that addresses these challenges by presenting a privacy-preserving framework that combines federated learning with transfer learning. The approach utilizes pre-trained EfficientNet-B7 encoders with U-Net decoders and fine-tunes them on distributed seismic datasets, enabling collaborative training without sharing raw data. Experiments on salt body segmentation tasks show that the TransFed-SaltNet achieves segmentation performance within 2–3% of a centrally trained model. It also demonstrated a 9.3% increase in the IoU compared to scratch-trained models. These results highlight the model's exceptional robustness and generalizability on unseen held-out data despite training on highly heterogeneous seismic datasets, offering a promising balance between performance and data confidentiality.