Drone damage assessment

It hasn’t been long since Hurricane Laura wreaked havoc in America – many have recorded the damage with drones, and scientists say it’s a very good and precise method that, combined with artificial intelligence, could benefit us. A hurricane struck the United States again – this time on Laura, Aug. 27, severely hitting the states of Texas, Louisiana and Arkansas. People used their drones flying in several places to try to assess the damage, and they also posted videos on social media sites. These videos are precise sources, researchers at Carnegie Mellon University have noted. They are currently working on drones that can be used for rapid damage assessment. Using artificial intelligence, researchers are developing a system that can automatically identify buildings and, in the first instance, determine if they are damaged and how severe these damages can be.

“Current damage assessments are mostly based on individuals detecting and documenting damage to a building,” said Junwei Liang, a PhD student at the CMU Institute of Language Technology (LTI). “But it’s a slow, expensive and labor-intensive task.”

Satellite images do not provide sufficiently detailed images and only show damage from a single – vertical – perspective. At the same time, drones can provide close information from many angles and perspectives. Of course, it is possible for anyone to send drones into the air as a first reaction, but today many of the residents have it and are already routinely sending it out after a natural disaster.

“Many drone videos are available on social media after one such incident, meaning these devices can be valuable resources in damage assessment,” Liang said.

Xiaoyu Zhu, a student of artificial intelligence and innovation at LTI, remarked that the rudimentary system can cover parts of buildings that appear damaged and determine if the damage is severe or less – or whether the building has been destroyed. The results of the research group will be presented at the Winter Conference on the Application of Computer Vision in 2021. The researchers, led by Alexander Hauptmann, a research professor at LTI, have downloaded drone videos of hurricane and tornado damage in Florida, Missouri, Illinois, Texas, Alabama and North Carolina. The videos were then recorded to identify the building damage and assess its severity. The resulting data set, which was the first to use drone videos to assess building damage caused by natural disasters, was used to train an artificial intelligence system called MSNet to be able to detect building damage. The data set is also available to other research groups through Github. The videos do not yet include GPS coordinates, but researchers are working on a geolocation scheme that would allow users to quickly identify the location of damaged buildings, Liang highlighted. This would require training the system using Google Streetview imagery. MSNet could then adapt the positioning signals learned from Streetview to the characteristics of the video, the expert added.