COMPREHENSIVE ROAD BLOCKAGE ESTIMATION FOR URBAN RESILIENCE WITH ARTIFICIAL INTELLIGENCE SUPPORT

Abstract

This thesis explores road blockage detection during earthquakes, using GIS and AI to assess road accessibility and identify rubbles caused by collapsing buildings. The study incorporates buildings and rubble into a framework to develop a model for detecting and analyzing structural failures after earthquakes. Past researches done in this field oversimplify reality and they do not account for spatial layout and most of them focus on building vulnerability rather than road blockages. Moreover, studies are data-hungry and can not be generalized. Therefore, this thesis aims to incorporate buildings and rubble into framework to build a model that will be able to detect and analyze the building collapse after the earthquake. The study utilized satellite images taken before and after earthquake for cities of Antakya and Kahramanmaras. Images were annotated in ArcGIS Pro and were transformed to black and white to export and further use them in deep learning model. As the number of images was insufficient data augmentation was applied to artificially increase it. To allow model to learn both the presence and absence of collapsed structures background patches were added, and the Dice coefficient was used to assess segmentation quality, accounting for both positive and negative outcomes. With the help of ArcGIS and AI, the model achieved a detection accuracy of 0.9319 for the training set, 0.6913 for the validation set, and 0.4982 for the test set. The limitations of the study include insufficient number of images and mountainous characteristics of the Kahramanmaras that resulted in decrease of test accuracy. The research highlights the importance of road connectivity for social functionality and the challenges urban areas face during natural disasters. These insights can help improve emergency response plans and urban planning strategies, ultimately making cities safer and more accessible in the aftermath of earthquakes.