Three-Dimensional Seismic Identification of Fault-Karst Bodies in the Shunbei Area Using a Geologically Constrained Synthetic–Field Mixed-Supervision Cascade nnU-Net Workflow

Abstract

Fault-karst bodies are an important ultra-deep carbonate reservoir type in the Shunbei area, Tarim Basin, Northwest China. Their seismic responses commonly include localized high-amplitude anomalies and bead-like reflections. Automated identification remains difficult because the targets are sparse, their boundaries are poorly defined, and complete voxel-level labels are unavailable for field data. We developed a geologically constrained synthetic–field mixed-supervision workflow based on the standard low-resolution-to-full-resolution cascade of nnU-Net. Fully labeled synthetic seismic samples provide complete three-dimensional geometric supervision. A partially labeled field-seismic region of interest supplies trusted foreground and background constraints, while uncertain voxels are excluded from both Dice and cross-entropy losses. The baseline cascade achieved a mean Dice of 0.92364 on the synthetic validation set. Combining field partial-label supervision with valid-voxel boundary weighting increased this value to 0.94642 in a 2 × 2 ablation. On the independent synthetic test set, the workflow achieved Dice = 0.92540, IoU = 0.86116, precision = 0.93077, and recall = 0.92010, averaged over three random seeds. Field predictions were evaluated against seismic-reflection characteristics and local well–seismic consistency. The workflow may support three-dimensional screening and characterization of candidate fault-karst bodies in the Shunbei area.