Building Damage Prediction with U-Net and Res-Net Architectures

Precise prediction of earthquake-induced building damage is essential for efficient disaster response, recovery planning, and resource allocation. Conventional techniques, which are mostly CNN-based, have been extensively employed for image classification tasks. These methods tend to fail to capture fine-grained structural information—like building edges and accurate damage extents—particularly in high-resolution aerial or satellite images.To overcome these shortcomings, we present a new Building Damage Prediction Framework that combines CNN with U-Net and ResNet architectures to enhance the accuracy and reliability of damage evaluation. In this work, we use U-Net to accurately segment buildings from pre- and post-disaster images, effectively outlining their perimeters. ResNet, a deeper, stronger network, is utilized next to extract high-level features that help in more accurate prediction of the severity of damage.Through the synergistic exploitation of these architectures' complementary strengths, our framework provides more fine-grained and accurate evaluations than traditional CNN-based approaches. Experimental assessment on high-resolution remote sensing data from recent earthquakes shows that the proposed model outperforms current methods by a large margin in estimating the extent and severity of building damage.This work identifies the capacity of sophisticated deep learning models in facilitating post-disaster damage evaluation, especially within densely constructed cityscapes.