Seismic Overthrust Fault Detection via Weakly Supervised Active Learning

Complex overthrust faults in deep seismic data often show weak boundaries and discontinuous spatial extensions. Under sparse annotation, three problems arise: weak-boundary recognition, insufficient three-dimensional connectivity supervision, and uncontrolled pseudo-label propagation. This study proposes a weakly supervised active learning framework with three modules. FG-ResACE U-Net is a two-dimensional segmentation network that outputs section-wise probability maps from amplitude, gradient, and continuity inputs, with wavelet gating, Atrous Spatial Pyramid Pooling (ASPP), and attention. Slice predictions are stacked into a three-dimensional probability volume; Top-K windows are selected from this reconstructed volume, not from a three-dimensional CNN, where K=0.02 means the top 2% of candidate windows. Geological-topology active learning ranks windows by prediction entropy, connectivity defects, and strike-direction deviation. Corridor-constrained Mean Teacher accepts positive pseudo-labels only at the intersection of expert candidate corridors and high-confidence teacher predictions. On synthetic overthrust models and Keshen three-dimensional seismic data, the complete framework, under the evaluated settings, shows more continuous main-fault responses and fewer fence-like fragments than the reported architectural variants.