Cross-Domain Salt Body Segmentation via Dual-Branch Collaborative Modeling and Dual-Granularity Prototype Contrast

Accurate segmentation of seismic salt bodies is of great significance for oil and gas exploration. Although existing deep learning methods have achieved remarkable progress within a single survey, segmentation performance degrades significantly when deployed to target surveys that exhibit systematic differences in geological settings and salt morphology, owing to a distributional shift in feature space. To address the cross-domain generalization problem, this paper first designs a CNN-Transformer dual-branch fusion network, DBF-CTSaltNet, as the backbone, which enhances the collaborative modeling of local boundaries and global morphology through the synergy of a Morphology-Adaptive Residual Unit, a Structured Global Guidance Unit, and a Bidirectional Cross-aware Fusion Unit. Building upon this, an unsupervised domain adaptation framework, UDA-DBF-CTSaltNet, is proposed, which independently constructs class prototypes in the respective feature spaces of the two branches and simultaneously drives cross-domain alignment at both local-boundary and global-morphology semantic granularities via dual-granularity prototype contrast. Experiments demonstrate that DBF-CTSaltNet outperforms mainstream models on the TGS source domain, and UDA-DBF-CTSaltNet significantly improves cross-domain segmentation performance on both the F3 and SEAM target domains.