A Building-Level Earthquake Damage Estimation Method Using Segmentation Models Pretrained on Open Land-Cover Data

Abstract

A common challenge in earthquake damage detection using deep learning is that post-disaster labels are costly to create and are rarely available immediately after a new event. This study therefore investigates a building-level damage estimation method that relies solely on segmentation results produced by a model pretrained on the open land-cover dataset OpenEarthMap, without any damage labels. Two indicators are proposed: one based on the post-event building-class ratio (Indicator A) and one based on the pre/post change of that ratio (Indicator B). The approach is evaluated on the 2024 Noto Peninsula earthquake (Wajima City, Japan) against visual-inspection labels, and is further analyzed by (i) per-class performance for total-collapse and partial-damage buildings, (ii) comparison against label-free alternative decision rules, and (iii) a sensitivity analysis on the polygon buffer size. The results demonstrate that the proposed method achieves an overall Precision of 0.85 and F1 of 0.65 on the binary damage classification, performing well on large-scale damage where the building outline is substantially lost (typically fire damage), whereas accurately identifying partial damage with remaining roofs (semi-collapse recall 0.05) remains challenging.