Research Information System
2026 International Conference on Electrical Engineering, Intelligent Control and Artificial Intelligence, EEICAI 2026, Bangkok, Thailand, 9 - 11 January 2026, pp.363-368, (Full Text)
Rapid and accurate assessment of building damage is paramount for effective disaster response operations. This paper addresses the semantic segmentation challenges in detecting structural destruction following the devastating earthquakes in Turkey on February 6, 2023. Traditional Convolutional Neural Networks (CNNs) often struggle to distinguish between complex damage levels due to their limited global context understanding. To overcome this, we propose a Hierarchical Swin U-Net architecture that leverages the long-range dependency modeling capabilities of Swin Transformers. Our primary contribution is the introduction of a novel, manually annotated dataset comprising 1,260 high-resolution images with over 10,000 building instances, classified into six granular damage levels. We present a two-stage hierarchical strategy: first localizing building footprints to filter out background noise, and subsequently classifying damage severity within these regions. Furthermore, we employ a transfer learning approach using a curated version of the xView2 dataset to address domain mismatch. Experimental results demonstrate that the proposed method, achieving a state-of-the-art 96.28% mIoU, significantly outperforms baseline models like U-Net (30.18%) and DeepLabV3+ (82.33%), particularly in distinguishing visually similar categories such as “Heavy Damage” and “Destruction.”