AI Use in Disasters: A Work in Progress with a Promising Future

Natural disasters like earthquakes, hurricanes, floods, wildfires, tornados, tsunamis, and volcanoes feel outside of human control. Still, AI disaster emergency response may be able to change the way we predict, prepare for, and respond to such emergency situations.

As seen with the recent earthquake in Turkey and Syria, being able to predict the potential magnitude, patterns, and aftershocks of an earthquake could save tens of thousands of lives.

Automation in the emergency response arena can potentially predict events and detect building structural issues, evaluate bridges and roads, run flood simulations, issue warnings via Google Maps and Google Search, coordinate relief efforts, and eventually direct resources for repairs.

How AI Disaster Emergency Response Automation Works

AI can be used to analyze large amounts of data from multiple sources far faster than any human could, creating useful predictions and conclusions in near real time. In the context of disaster response, the predictions are created using data such as seismic sensors, geological data, images from drones and satellites, rainfall measurements, and climate records.

The speed with which automation operates and the scale it can handle makes it far more promising than traditional efforts. In recent events, automation technologies were able to find areas of damage that workers on the ground were previously unaware of, allowing for redirection of efforts.

Automation can build on other technology, too.

For example, drones can capture otherwise unobtainable footage but can’t analyze it. AI disaster emergency response technology such as xView2 can provide this higher-level analysis.

This open-source technology was developed by the Pentagon’s Defense Innovation Unit, Carnegie Mellon University’s Software Engineering Institute, Microsoft, the University of California, Berkeley, and others. It evaluates images at a pixel level in a process called semantic segmentation.

Limitations in Automation

As is true in many AI use cases, there are limitations to current AI disaster emergency response technology. Much of the data used in predicting future events comes from past events, and current measurements and situations in real time may not be factored in correctly or quickly enough.

Also, human error could still be an issue because the technology relies on humans both feeding and interpreting the data. And since a mistake in this context could mean thousands of deaths, getting predictions right is vital.

Adoption among workers on the ground is another challenge. The good news is that researchers using this technology have already taken data from more than 131,000 earthquakes (not to mention other natural disasters) and tested models in over 30,000 events.

When strategizing rescue efforts using satellite images, technology is limited by the positions of satellites within their orbits, available daylight at the time of the event, and even weather conditions.

Here’s a recent real-world example: the recent earthquake in Syria and Turkey occurred on February 6th, but useful images were not acquired via satellite for another three days.

Solutions are being proposed to use alternatives such as synthetic aperture radar, which uses microwave pulses and thus doesn’t need light to function.

Satellites also only give an overhead view of buildings: they cannot see inside structures or reveal what other perspectives may show. Even so, AI disaster emergency response has been found to have about an 85-90% accuracy rate in evaluating and assessing the damage.