Comparison of CNN Architectures for Earthquake Damage State Classification of Different Building Types

Abstract

Rapid and reliable post-earthquake damage assessment of the built environment is critical for prioritising response and recovery efforts. This study investigates deep-learning strategies for classifying different building types into three damage states: slight, moderate and severe, using photos that capture both the structure and its surrounding context. Unlike many prior works that rely on tightly cropped images of specific cracks, our dataset intentionally preserves a wider field of view, making the task more realistic yet significantly more challenging. Two convolutional neural network (CNN) architectures are examined: ResNet-18 and ConvNeXt-Tiny. A two-stage training protocol is employed. In the first stage, before training on the original photos, a ResNet-18 model is pre-trained on a subset composed exclusively of close-up patches, providing a strong damage sensitivity. Second stage uses the original training set of the photos, extended using horizontal-flip augmentation. The same protocol is then replicated for ConvNeXt-Tiny architecture. Results show that dataset expansion and simple augmentation improve ResNet-18 recall, confirming the benefit of contextual diversity. Nevertheless, ConvNeXt-Tiny consistently outperforms ResNet-18, achieving a recall of 72% and an F-score of 0.85. The findings highlight the importance of appropriate selection of CNN architectures, especially for demanding tasks. The proposed workflow offers a pragmatic route toward scalable, context-aware earthquake damage assessment tools, and provides a benchmark for future research on holistic post-disaster image analytics.