A novel post-disaster damage assessment using generative AI: ground-level damage inference from aerial data after the 2023 Turkey Earthquakes
International Journal of Remote Sensing, cilt.47, sa.7, ss.3161-3179, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 47 Sayı: 7
- Basım Tarihi: 2026
- Doi Numarası: 10.1080/01431161.2026.2628294
- Dergi Adı: International Journal of Remote Sensing
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Artic & Antarctic Regions, BIOSIS, Compendex, Environment Index, Geobase, INSPEC, Public Affairs Index
- Sayfa Sayıları: ss.3161-3179
- Anahtar Kelimeler: AI-based methods, building damage, deep learning, diffusion models, Earthquakes, remote sensing, structural engineering
- Ankara Yıldırım Beyazıt Üniversitesi Adresli: Evet
Özet
Post-disaster assessment is pivotal for rescue operations. However, ground-level data are often not available immediately after an earthquake. This study proposes a novel approach to use generative AI (Artificial Intelligence) for assisting with post-earthquake assessment or disaster preparedness, by reasoning visual ground-level damage from rapidly collected aerial data. Our approach takes Unmanned Aerial Vehicle (UAV) orthophotos and the potential damage map as inputs to predict the ground view at every point of the map, subject to a presumed damage level. Using generative AI, specifically, a conditional diffusion model trained from real post-earthquake ground images, we generate a reasonably realistic view of a damaged building. This novel technique can enhance first responders’ visual perception of ground-level building damage when data are scarce. This allows quick building damage assessment, failure identification and understanding damage extent for effective evaluation and emergency response. Our proposed diffusion model, termed LEGG (LoRA-Enhanced Ground-view Generation), produces images guided by two textual prompts: ‘damaged’ and ‘undamaged’. A pre-trained diffusion network is fine-tuned using Low-Rank Adaptation (LoRA) on ground-level photographs of damaged structures, enabling the model to capture and reproduce detailed damage characteristics. Overhead imagery and learned ground priors are then fed into the adapted diffusion model to generate the high-resolution ground-level views. Our model achieved significant performance improvement on a dataset of approximately 3000 buildings from the 2023 Kahramanmaras, Turkey earthquake, reaching a Peak Signal-to-Noise Ratio (PSNR) of 19.82 and Fréchet inception distance (FID) of 66.06. Compared to previous examples, we demonstrated better detection of partial collapse and subtle façade cracks. While recent earthquake studies have mainly relied on UAV or LiDAR-based detection of collapsed buildings, none have addressed the synthesis of realistic ground-level views, making this work unique and distinct in bridging that gap.