Post-Earthquake Candidate Building-Loss Screening from Bitemporal Google Earth Imagery: An Object-Level Framework for the 2023 T¨urkiye Earthquakes
Year: 2026
Type: preprint
Abstract: Rapid and reliable identification of building losses after major earthquakes is essential for post-disaster assessment and decision support; however, many existing approaches depend on accurately coregistered imagery, building footprints, geographic information system layers, or other auxiliary geospatial datasets that may not be consistently available across large affected regions. This study proposes an object-level bitemporal framework for automatically screening candidate building losses from historical Google Earth imagery acquired before and after the 6 February 2023 Türkiye earthquakes. A hybrid building-detection dataset was constructed by combining the Satellite-II Version 1 dataset with manually verified Google Earth images from earthquake-affected areas. Six representative object detectors—Faster R-CNN, RetinaNet, SSD, YOLOv10n, YOLOv11n, and YOLOv12n—were evaluated under a harmonized experimental protocol. YOLOv12n provided the most favorable accuracy–efficiency trade-off and was subsequently retrained using the complete hybrid dataset. On an independent Google Earth test set, the final model achieved 72.19% precision, 33.26% recall, and a 45.54% F1-score, indicating substantially improved precision after full-data training. For the temporal analysis, corresponding 2022 and 2024 images were spatially aligned using stable non-building references, including road and background structures, followed by object-level one-to-one matching based on geometric correspondence. A total of 145 bitemporal image pairs were processed. The detector identified 2,872 buildings in 2022 and 2,039 buildings in 2024; 842 buildings were successfully matched across acquisition dates, whereas 2,030 unmatched detections were retained as candidate building losses, yielding an initial automatic candidate-loss rate of 70.68%. The results demonstrate the potential of publicly accessible historical Google Earth imagery for scalable preliminary post-earthquake building-loss screening without mandatory reliance on building footprints, LiDAR, or cadastral data. The proposed framework is intended as a decision-support tool for prioritizing locations requiring detailed expert assessment rather than as a stand-alone system for definitive damage classification.
Source: Research Square
Author: perihan karaköse
Institution: Bartin University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204856624