LEE- Afet Yönetimi Lisansüstü Programı

Bu topluluk için Kalıcı Uri

http://hdl.handle.net/11527/19190

Gözat

Yazar "Saloğlu Dertli, Didem" ile LEE- Afet Yönetimi Lisansüstü Programı'a göz atma
  • Öge
    Enhancing disaster management through deep learning: Building damage assessment using satellite imagery
    (ITU Graduate School, 2025-06-17) Mohammadiahani, Mohammadreza ; Saloğlu Dertli, Didem ; 801231003 ; Disaster Management
    In the face of increasingly frequent and severe natural disasters, the need for rapid, accurate, and scalable damage assessment tools has become more critical than ever. This study explores the integration of artificial intelligence (AI), specifically deep learning models, into the post-disaster response process to improve situational awareness and facilitate effective decision-making. Focusing on the case of Antakya, a city in Turkey severely impacted by the catastrophic February 6, 2023 earthquakes, the study presents a comprehensive framework that leverages satellite imagery and AI-driven segmentation to detect and analyze building damage. The primary goal of this research is to develop a deep learning-based pipeline capable of automatically identifying buildings in satellite imagery and assessing damage. A customized U-Net architecture was implemented due to its proven performance in semantic segmentation tasks. The model was trained on a curated subset of the xBD dataset, consisting of over 5,000 high-resolution satellite images labeled with building footprints and damage levels. Additional building footprint masks were generated using OpenCV and shapely operations on JSON annotations, with all images resized to 256x256 pixels and normalized for consistency. Visual evaluation played a significant role in validating the quality of segmentation outputs. Eight representative validation examples were analyzed, each showing the original satellite image, the ground truth mask, and the model's predicted binary mask. These visuals confirmed the model's ability to accurately detect most building structures, especially those with clear edges and uniform textures. Challenges remained in areas with heavy vegetation, shadows, or highly irregular building shapes, which occasionally led to false negatives or incomplete segmentation. To assess generalization beyond the training dataset, the model was tested on five pairs of pre- and post-disaster satellite images of Antakya, independently acquired using Google Earth Pro. These images were not part of the original training or validation data, offering an authentic test of the model's real-world applicability. Results demonstrated the model's robustness in detecting building footprints even in domains with different visual characteristics. The predicted masks and overlaid polygons on real-world imagery aligned well with visible structures, and heatmaps generated from the predictions successfully highlighted areas of dense or missing construction, providing visual insights into structural loss. The study also introduced building density heatmaps for both pre- and post-earthquake conditions in Antakya. These heatmaps offered an interpretable spatial visualization of predicted building footprints, highlighting the contrast in structural integrity before and after the disaster. A comparative analysis across five regions (Ant1–Ant5) revealed noticeable declines in building density, especially in urban cores, thereby validating the model's ability to capture large-scale damage patterns. Overall, the presented thesis confirms that AI-powered segmentation models such as U-Net can be effectively deployed for post-disaster damage mapping. While limitations exist, including dependency on high-quality imagery and the risk of under-segmentation in complex zones, the proposed system offers a scalable and efficient solution for disaster response. By automating the assessment process, the framework reduces response time and enhances decision-making in critical scenarios. The findings of this study contribute to the growing body of work advocating for the use of AI in humanitarian operations and establish a foundation for future improvements, including the integration of multi-modal data and advanced model architectures. This thesis ultimately demonstrates the transformative potential of AI in disaster management, emphasizing its ability to support faster, more informed, and more equitable emergency responses in the wake of large-scale catastrophes.