Re-mapping after natural disasters is critical for assessing damage, optimizing aid distribution, and more. Creating new maps is thus is a central element of any disaster response effort. At present, re-mapping involves time-intensive, manual approaches where mappers spend significant time labeling buildings and roads in imagery. Current efforts are too slow to aid immediate response efforts: for example, re-mapping of Puerto Rico by the Humanitarian OpenStreetMap Team and 5,300 volunteers took approximately two months. Automation or partial automation of mapping using artificial intelligence presents an exciting opportunity to accelerate the process.
To acquire images more quickly, early satellite imagery after disasters is often taken at a substantial angle away from directly overhead - often over 40 degrees “off-nadir” - but past studies have not answered a fundamental question for using AI for disaster response mapping: can algorithms identify features (e.g. buildings) in images taken at a significant angle? No existing public datasets contain many images acquired significantly off-nadir angles, and those datasets generally do not contain multiple looks of single geographies to evaluate the effect of look angle. To enable exploration of look angle’s impact on automated foundational mapping, we open sourced a dataset of 27 views of Atlanta, GA taken from 7 to 54 degrees off-nadir during a single pass of a WorldView-2 satellite, along with high-quality, geographically accurate building footprint labels of the imaged area. These data, provided to the public free of charge on Amazon Web Services (AWS) S3 with a CC-BY 4.0 license, were used in the SpaceNet Challenge Round 4 run on TopCoder. In this challenge, participants competed to develop the best building footprint extraction algorithm from the many different look angles provided. We will discuss the dataset, the barriers to applying machine learning algorithms to off-nadir images, and the lessons learned about the limits of current state-of-the-art algorithms through this challenge. These lessons will inform the viability of machine learning as a component of disaster mapping.
