Post-disaster structural damage assessment based on semantic segmentation using remote sensing images

In the aftermath of natural disasters, integrating aerial imagery with interior building images is considered one of the most effective and reliable methods for remote sensing-based damage assessment, as it enables the extraction of structural damage information from roof, façade and interior views. While interior damage analysis focuses on inspecting structural elements, exterior assessments are typically limited to roofs and façades. The first part of this dissertation addresses deep learning-based semantic segmentation approaches developed to overcome the limitations of conventional machine learning techniques in post-earthquake damage detection and crack segmentation in reinforced concrete structures. Specifically, a DeepLabV3+ based semantic segmentation architecture was proposed for pixel-level detection of surface cracks in concrete. Within this framework, the performance of various pre-trained backbone architectures was comparatively analyzed in the encoder block. To evaluate the effectiveness of low and high-level feature extraction capabilities of these architectures, publicly available datasets, namely DeepCrack and CrackForest, were employed. experimental results demonstrated that the MobileNetV2 model outperformed the others in terms of both parameter efficiency and segmentation accuracy. The second part of this study presents a comprehensive analysis of deep learning-based semantic segmentation models for post-disaster structural damage assessment in terms of accuracy, generalization ability and computational efficiency. Using the custom-built Kahramanmaraş Earthquake dataset, along with high-resolution drone and satellite imagery from the Joplin Tornado and xBD datasets, the objective was to detect building-level damage grades. The DeepLabV3+ segmentation architecture was evaluated using various backbone networks. For building-level classification, segmentation outputs were merged with polygon boundaries extracted from pre-disaster imagery to determine the structural damage status of individual buildings. experimental findings revealed that ResNet50 and Xception provided the highest segmentation accuracy overall, whereas MobileNetV2 offered advantages in computational efficiency, making it preferable for lightweight and rapid applications. On the other hand, due to class imbalance and the complexity of debris patterns, all models exhibited difficulty in accurately detecting fully collapsed buildings an issue most prominent in architectures with shallower depth. This study proposes a deep learning-based framework for rapid and accurate post-disaster damage assessment.