Modern environmental disasters demand faster, more efficient, and more dynamic responses than public systems were initially built to manage. With the potential dissolution of the Federal Emergency Management Agency, AI and drone technology will likely be filling the gaps at some point.
Stateline, the editorial arm of The Pew Charitable Trusts recently published an article discussing the major concerns states and local governments have raised regarding the potential shut down of the Federal Emergency Management Agency (FEMA). This comes at a time when storms continue to inflict damage beyond the capacity of existing systems. A recent example is the off-peak rainstorm that triggered flooding that took the lives of nearly 140 residents of Kerr County in the Texas Hill country. With another historic hurricane season on the horizon, local agencies are looking for ways to do more with less, and be as nimble as possible.
Disaster specialists have been sharing thoughts on how AI and drone technology can be used to essentially cover more ground with less human power. A Route-Fifty article in the commentary section discussed, “A new path to recovery: How AI can transform disaster response.” An expert writing for Govtech questioned, “When Disaster Strikes, Can AI Deliver Where FEMA Doesn’t?” These are vital questions and concerns local and state governments are needing to address.
While exploring expanded options for emergency response is becoming more necessary, some experts say the specific utility of AI and drones as first responders, as well as a search and rescue tool, is still limited. Two professors from Texas A&M University recently wrote an article about the potential of an AI-integrated system to survey drone footage of flood zones and actually help first responders locate victims by sending the teams on the ground an analysis from the aerial view. The actual technology involved applies computer vision and machine learning to the live or recorded drone footage of a disaster area. This has the potential to help search and rescue teams prioritize areas to focus on rather than set eyes on every area, only to rule out say 50 percent of them as not viable after the fact. But this is the exact strategy researchers are saying is still producing a high error rate.
The main challenge is the breadth of available datasets that show the type of of debris specific to a certain disaster in order to train the program effectively i.e. hurricane debris and flood debris from a rainstorm are different, particularly when predicting the location of victims. So without, at the very least, hundreds of thousands of aerial images and videos of flood zones specific to rainstorms, with potential victims visible, the error rate will remain high. Most AI-integrated systems are currently trained for locating the full frame of an actual human being rather than signs of the potential presence of a human being, like clothing obscured by water or mud. Another issue is the oblique view that drone cameras produce make it difficult to precisely pinpoint and transmit the location of the flagged areas relative to the drones position when the visual was produced.
