Deep Ensemble Learning for Rapid Large-Scale Postearthquake Damage Assessment: Application to Satellite Images from the 2023 Türkiye Earthquakes

Abstract

Extensive field reconnaissance damage surveys, publicly available after the Türkiye earthquake sequence of 2023, provided a unique opportunity to devise and validate a rapid postevent damage assessment framework that uses Artificial Intelligence (AI) techniques to overcome typical challenges encountered in rapid regional damage assessment efforts. By analyzing publicly available satellite images of the significantly impacted city of Antakya, we manually identified and segmented fully or partially collapsed buildings and buildings with visible damage. These were then paired with the various damage state labels in the government data. An AI-based framework was subsequently developed to automate segmentation and damage assessment processes, delivering damage state estimates for other affected regions. The resulting tool, dubbed rapid postearthquake aerial imagery damage assessment (RAPID-A), is an ensemble of various deep segmentation models that work on satellite image channels with a resolution of 30 cm per pixel, augmented with damage proxy maps, such as NASA ARIA maps with a resolution of 30 m per pixel. Using an image segmentation strategy over object detection, RAPID-A provides an uncertainty-aware estimate of different intensities of collapsed and damaged-but-not-collapsed buildings as log-normal distributions. Test case studies carried out over Gaziantep and Kahramanmaras—two of the heavily impacted cities—demonstrate that RAPID-A is generalizable, accurate, and efficient. Such tools can offer significant aid in rapid initial damage evaluations before first responders and damage assessment teams are dispatched. It could also be a main tool for countries that might be unable to launch large-scale regional assessment campaigns. The latter is highly important as it helps other nations focus their aid. The results further suggest that RAPID-A is generalizable and can be applied to future affected areas with an unseen urban fabric.