AUTOMATED BUILDING DAMAGE ASSESSMENT USING FEATURECONCATENATED SIAMESE NEURAL NETWORKS

Abstract

Building damage assessment is a crucial task in postdisaster management, enabling rapid response and resource allocation. Traditional methods relying on field surveys and manual remote sensing analysis are time-consuming, costly, and often impractical in disaster-stricken areas. To address these limitations, this research proposes an advanced approach utilizing a Feature-Concatenated Siamese Neural Network (FCSNN) for automated building damage classification. The model leverages pre-earthquake and post-earthquake data, implementing a novel concatenation mechanism that enriches feature extraction by aggregating low-, mid-, and high-level representations. This enhancement enables improved discrimination of structural changes, significantly boosting classification performance. Our study conducts three experimental scenarios: (1) binary classification of intact (G1) vs. collapsed (G5) buildings, (2) classification of non-collapsed (G1-G4) vs. collapsed (G5) buildings, and (3) multiclass classification based on the five EMS-98 building damage grades. The proposed FCSNN model achieves F1-scores of 79.47%, 54.09%, and 40.64% for the respective experiments, outperforming conventional Siamese Neural Networks (SNN) and state-of-the-art Support Vector Machine (SVM)-based methods. The model also attains accuracy scores of 87.24%, 95.28%, and 42.57% across the three scenarios, demonstrating its robustness in handling imbalanced datasets and complex damage patterns. The findings validate the effectiveness of feature concatenation in deep learning-based damage assessment, offering an efficient and scalable solution for post-disaster building evaluation. Future enhancements include the integration of voxelbased representations and alternative sub-network architectures to further refine classification accuracy. This study contributes to the advancement of AI-driven disaster response frameworks, facilitating faster and more reliable damage assessments in real-world scenarios.