Transformer-Based Segmentation of Post-Earthquake Damage

Rapid and reliable mapping of earthquake-induced building damage supports emergency response and recovery planning. This study investigates post-event-only damage segmentation from high-resolution optical imagery using transformer-based semantic segmentation. We adopt SegFormer with a MiT encoder and an all-MLP decoder head, and compare it against a strong baseline (Swin Transformer with UPerNet). Using the KATE-PD benchmark, we evaluate the effects of loss design (cross-entropy vs. cross-entropy + Lovász), data augmentation, test-time augmentation (TTA), and validation-based threshold selection. Results show that the best-performing configuration (SegFormer + Lovász + Augmentation + TTA) achieves improved performance over the baseline SegFormer, while maintaining a simple post-event-only inference pipeline. Qualitative examples highlight cleaner delineation of damaged regions under augmentation and TTA. Keywords: post-earthquake damage, semantic segmentation, transformer, SegFormer, Swin, KATE-PD, test-time augmentation.