Using AI for good: a new data challenge to use AI to triage natural disaster aerial imagery
By Kalev Leetaru
Deep learning has revolutionized how we process the vast firehoses of data that define modern life. Yet, the daily drumbeat of AI headlines tends to center on the commercial applications of AI and how it is reshaping how companies do business. In a refreshing twist, a new open AI challengeby the World Bank, in collaboration with WeRobotics and OpenAerialMap, illustrates the incredible potential of deep learning for humanitarian applications, especially in the critical hours and days after a major natural disaster.
From satellite imagery to aircraft and UAV aerial imagery, the ability to gain a birds-eye-view of a disaster is critical to rapidly triage the level of civilian and infrastructure damage, the condition of transportation corridors and to prioritize relief efforts. A key obstacle is that the volume of available imagery typically far outstrips the ability of humans to fully assess it all at the time scales and resolution needed.
This new AI challenge looks to address this by focusing on two initial areas: identifying trees and roads from aerial imagery.
Food producing trees are often a critical source of economic and food security in the South Pacific Islands and widespread damage to them from disasters can have long term impacts. The first task of the Challenge therefore focuses on building an image classifier that can take an aerial image and return an annotation layer that identifies all coconut, banana, papaya and mango trees and their locations in the photograph with at least 80% accuracy. In a production scenario, one could imagine being able to compare before and after imagery in realtime to generate precise counts of tree loss and the specific species lost. Given rapid advances in mobile low resource execution of neural networks, one could even imagine such algorithms eventually operating in realtime, perhaps even on the UAV itself in flight, allowing it to autonomously adjust course to image the most heavily damaged areas in greater detail.
The second task revolves around identifying roads from imagery and determining whether they are one or two lanes and whether they are paved or dirt roads. Deployed eventually in the field, such algorithms could provide aid planners near realtime road condition reports for the affected area, allowing them to immediately plan aid transportation routes and prioritize road rebuilding efforts.