Latest revision as of 13:17, 25 September 2026

Abstract

Timely and accurate building extraction from disaster imagery is important for emergency response, damage assessment, and post-disaster recovery. Staring satellite video provides stage-specific observations of the same affected area as a disaster evolves. However, building appearances vary substantially across stages, from intact structures to deformed buildings and fragmented remnants, resulting in pronounced stage-dependent within-class variation. Existing methods are commonly developed using single-date imagery, and their reliability across heterogeneous damage conditions remains insufficiently investigated. To address this problem, we propose PGDGNet, a Prior-Guided Detail–Global Fusion Network for frame-wise multi-stage building extraction. PGDGNet employs a dualbranch CNN–Transformer architecture to capture local structural details and global semantic context. Its Detail–Global Feature Fusion (DGFF) module applies branch-specific attention and spatial gating to reconcile original and attention-enhanced features before local–global integration. A joint color-and-boundary prior-guidance strategy provides complementary appearance and contour cues when structural information is weakened, while boundary-aware supervision facilitates the delineation of incomplete buildings. A patch-level dataset is constructed from frames representing the early, intermediate, and late stages of the 2018 Indonesian tsunami–earthquake event. Experiments show that PGDGNet achieves the highest Building IoU and recall across all three stages. Compared with the strongest baseline, CFFormer, PGDGNet improves Building IoU by 0.59, 0.63, and 0.81 percentage points and recall by 1.45, 0.80, and 1.08 percentage points in the early, intermediate, and late stages, respectively. The results demonstrate robust and reliable performance across disaster stages.

Document