2026 Volume 26 Issue 3 Pages 3_82-3_97
The 2016 Kumamoto earthquake resulted in significant casualties and building damage. To support safe and rapid assessment of building damage, this study develops a deep learning-based model for detecting completely collapsed buildings. The proposed approach uses a digital surface model (DSM) obtained from airborne LiDAR surveying together with post-earthquake aerial images. The model was trained using two different datasets. The results showed that the model achieved higher performance when using four-band images that combine the difference between pre- and post-earthquake DSM with post-earthquake aerial imagery. Based on this dataset, hyperparameter optimization using Optuna was further applied to improve the detection performance. The final model achieved an accuracy of approximately 81% and a recall of approximately 63%.
