Classificação e segmentação de danos em edifícios após desastres naturais utilizando dados tabulares e imagens registradas pré e pós-desastre

Abstract

Natural disasters that cause significant damage to urban structures, representing considerable risks to human life and substantial financial losses. The rapid identification of buildings with critical levels of damage is essential to mitigate further risks and accelerate recovery efforts. This work developed models to classify and segment building damage, using tabular data for classification and recorded images for segmentation. In the classification approach, we used tabular data from the Gorkha earthquake, applying feature engineering techniques to enhance the performance of machine learning models. The XGBoost model, trained with these techniques, achieved an F1-score of 0.7542, demonstrating its effectiveness in identifying building damage levels. This result highlights the importance of data preparation and optimization to improve the performance of classification models in disaster scenarios. For segmentation, we explored various deep neural network architectures, focusing on Unet with a BDANet backbone, in addition to ResNet18, VGG16, and ResNet50. Using satellite images, we applied mathematical morphology operations to refine the damage segmentation. This approach allowed for the segmentation of damage levels, including no damage, minor damage, and major damage. The Unet model with BDANet, combined with mathematical morphology techniques, achieved an F1-score of 0.7999, evidencing its superior performance in post-disaster damage segmentation. The classification approaches using tabular data and segmentation using recorded images proved effective in assessing building damage after natural disasters. Each methodology enhances the accuracy of assessments in its respective domain, providing valuable tools for authorities and rescue teams. These advanced models assist in prioritizing efforts and allocating resources more efficiently in disaster situations, significantly contributing to the effectiveness of emergency responses.