Researchers with the University of Washington worked in tandem with the University’s Disaster Data Science Lab to construct a convolutional neural network that identifies damaged buildings. Satellite imagery has long been used to detect the presence of infrastructure, vehicles, and natural phenomena but has yet to be extensively applied to post-hurricane damage detection.
When determining a natural disaster response plan, officials need to know where the most severe damage is concentrated. In other words, they require situational awareness. The typical form of assessment used today is known as a windshield survey — responders drive around the impacted area to gain firsthand knowledge of the emergency terrain. While aerial images captured by drones can give emergency response crews a bird’s-eye view, this fails to eliminate the need for a human to inspect the visual information. As a result, these procedures tend to be tedious and far from foolproof.
To automate the damage detection process, the researchers devised a machine learning algorithm that can classify buildings as either “flooded/damaged” or “undamaged” using satellite imagery. They trained a neural network on volunteer-annotated data that consisted of before and after scenes from the Greater Houston area in the wake of Hurricane Harvey. In spite of the lower resolution of some satellite images and inconsistencies in image quality, the best performing model achieved 97.08% accuracy on the unbalanced test set.
Faster response rates and better resource allocation in the event of a disaster could save countless lives and ensure victims get the care they need. Going forward, the researchers “wish to expand the current research to road damage annotation which could help plan effective transportation routes of food, medical aid, or energy to the disaster victims.”
Find out more here.
