Automated Damage Detection After Earthquakes: Algorithms and Image Catalogs

Session: Earthquake Ground Motion and Impacts [Poster]
Type: Poster
Date: 4/30/2020
Time: 08:00 AM
Room: Ballroom
Description: 

When an earthquake occurs, rapid response efforts are focused on human life and casualties and therefore, rapid information collection is crucial. Collapsed and damaged buildings and infrastructure are usually the location for the highest number of human casualties and the highest loss. Therefore, detecting the collapsed and damaged infrastructure immediately after an earthquake and providing information to first responders and loss estimation is critical. The most efficient way to get this information immediately and in large scale is to use earth observing satellites. Today’s commercial satellites can capture images from the earth’s surface with resolution as high as 30 cm. In recent years, two significant databases of demolished and intact buildings across the world have been made publicly available through Defense Innovation Unit Experimental (DIUx) and National Geospatial –Intelligence Agency (NGA) known as the “xView” and “xView2” datasets. The “xView2” dataset provides damage labels for buildings across multi-hazards. The labels range from intact to destroyed: (no damage, minor damage, major damage, destroyed). We have developed and tested a rapid damage classification framework using convolutional neural networks (CNNs), a popular Machine Learning algorithm, which uses labeled images of damaged buildings as inputs and provides an automated classification of the four levels of building damage affected by the earthquake. Through evaluation of the “xView” and “xView2” datasets and automated classification algorithms, we are able to provide lessons learned in building effective image catalogs of damaged buildings and in algorithm use for rapid damage detection using CNNs. Our goal is to leverage these lessons to build additional image catalogs for ground failure damage and infrastructure damage (beyond buildings).

Presenting Author: Laurie G. Baise

Authors