Post-Disaster Building Damage Segmentation Using Convolutional Neural Networks

Natural disasters are events caused by nature such as earthquakes, tornadoes, tsunamis, forest fires, and others. The impacts of natural disasters are significant and varied across various sectors, including the economy, health, and primarily, infrastructure. Effective and efficient actions are needed to assist in the recovery following natural disasters, one of which is aiding in the identification of building damage levels post-disaster. To address this issue, this research proposes a system capable of performing segmentation to determine the level of building damage post-natural disaster using convolutional neural network methods. The data utilized consists of aerial images sourced from xView2: Assess Building Damage, comprising 50 aerial images with 5 classes: no-damage, minor-damage, major-damage, destroyed, and unlabeled. The steps undertaken in this research include data preprocessing using patchify and data augmentation. Subsequently, feature extraction is performed using convolution, followed by the training process using a neural network with the proposed architecture. This study proposes an architecture with 27 hidden layers, with feature extraction utilizing average pooling. The model evaluation process will employ Mean Intersection over Union (MIoU) to assess how closely the segmentation prediction results resemble the original data. The proposed architecture demonstrates the best MIoU result with a value of 0.31 and an accuracy of 0.9577.