Geospatial artificial intelligence, or GeoAI, researchers at the Department of Energy's Oak Ridge National Laboratory contributed to this year's SpaceNet 8 challenge, a data science competition to create the best algorithm for infrastructure and flood mapping in a pre-defined region. SpaceNet 8 is using a data pre-processing pipeline and a baseline algorithm developed by ORNL, a first for the team to support an open-source benchmarking challenge.
SpaceNet is an informal collaboration among digital industry leaders such as Maxar, Amazon Web Services, Topcoder, IEEE Geoscience and Remote Sensing Society and ORNL that is intended to accelerate open-source machine learning capabilities.
SpaceNet challenges are opportunities for geospatial professionals and data scientists to interact with data and analytics and apply them to a relevant geospatial use case. This year, as news headlines highlight devastating floods around the world, ORNL researchers joined their SpaceNet counterparts to organize and scope a new challenge toward infrastructure and flood mapping.
"It's quite rewarding to be in this partnership where we are teaming with other SpaceNet partners to shape priority benchmark challenges and develop tools to accelerate open-source geospatial machine learning capabilities. Creating benchmarks is critical for advancing research and lowering the barrier of entry for many in the community," said Dalton Lunga, a senior scientist and group lead for GeoAI at ORNL.
At the start of the competition, participants gained access to a baseline algorithm as a reference point. From there, they customized the provided algorithm and developed their own algorithm to find flooded roads or buildings on satellite images. When the challenge concluded on Aug. 23, the winner was selected based on their algorithm's accuracy on a testing dataset.
When floods devastate communities, emergency management organizations are in a race against time to deploy responders and supplies. These organizations increasingly lean on data scientists using satellite images and computers to make the best plan for saving lives. As thousands of satellite images are taken after flooding, data scientists create algorithms to sort through the images and detect buildings, houses and other features. For the SpaceNet 8 challenge, the algorithms are looking for flood-damaged buildings and roads.
Most often, communities lack access to high-quality labeled geospatial data sets, customizable algorithms or large computers to run the computations. ORNL's participation in this challenge provides thought leadership to the community, said Lunga. "We give them an example of an algorithm that works, and then they can just fine-tune to improve upon that."
Data is a critical component not only during the competition, but also in real scenarios. ORNL was able for the first time to share a set of images with the geospatial community. Jacob Arndt, a geospatial imagery data scientist, led the development of the baseline algorithm and pre-processed the data to make it usable for the challenge. Lunga and Arndt hope this experience will encourage data scientists and geospatial professionals across the world to participate in the SpaceNet competitions or customize the algorithms further for other geospatial applications.
Arndt specializes in assessing infrastructure damage for a variety of disasters. "The type of framework that's being used for flood detection in the SpaceNet 8 challenge is applicable to different types of damages. One of which is identifying damage due to conflicts, like what we're seeing in Ukraine, for example, or damage due to other natural disasters such as fires and tornados."
Challenges such as SpaceNet, though a competition in spirit, demonstrate how a community of data scientists are crowdsourcing advances in machine learning. "As data and AI resources are shared among more professionals, such actions afford the community huge strides forward in responding to impacts due to anthropogenic threats and natural disasters," Lunga said.
