U-Net: A Binary Segmentation Architecture for Building Localization in Structural Health Monitoring using the xBD Dataset

Structural Health Monitoring (SHM) is essential in disaster scenarios, providing critical support for building localization, accurate damage assessment, and effective classification of damage severity. SHM refers to the process of monitoring and tracking the condition of engineering structures over a period of time to detect damages, deterioration or potential failures. Localization in such scenarios aids in identifying the buildings and generating ground truth. Traditional techniques of damage assessment are often manual and increases the labor cost due to onsite inspection. This paper aims to propose a methodology based on image segmentation which includes UNet, a Deep Neural Network(DNN) based architecture with a ResNet-50 encoder and custom decoder. It is used to detect and localize buildings from the predicted masks and classify them as Building Detected(1) and No Building(0) based on the probability scores. The proposed approach has achieved a test accuracy of 0.9727 and an Intersection over Union (IoU) score of 0.6289 and a Dice Coefficient score of 0.7721 on all 11 disaster types present in the xBD dataset. The proposed model is further compared with other segmentation architectures like Fully Convolutional Networks (FCNs) and Attention U-Net based on IoU and Dice Coefficient scores. The proposed model shows better performance as compared to the existing models.