LAS VEGAS — Intel has developed AI models to identify geographical features from satellite imagery for the creation of accurate, up-to-date maps. The company has been working closely with the Red Cross on its Missing Maps project, which aims to create maps for areas of the developing world to improve disaster preparedness. Many areas of the developing world do not have up-to-date maps, which means that aid organizations can struggle to work efficiently in the event of natural disasters or epidemics.

“As someone who’s been on the ground with the Red Cross, having access to accurate maps is extremely important in disaster planning and emergency response,” said Dale Kunce, co-founder of Missing Maps and CEO of American Red Cross Cascades Region. “But there are entire parts of the world that are unmapped, which makes planning and responding to disasters much more difficult. This is why we’re collaborating with Intel to use AI to map vulnerable areas and identify roads, bridges, buildings, and cities.”

“If you don’t know where all the roads are before a hurricane hits, after it hits, you have no idea where flooding has occurred or which roads are washed out and which aren’t,” said Alexei Bastidas, deep-learning data scientist at Intel AI Lab, in an Intel podcast on the subject. “If you don’t have an accurate enough map of what was there beforehand, it really prevents you from responding to the disaster as it’s ongoing. The other thing to consider is that a lot of these disasters … are weather events — cyclones, typhoons, hurricanes, even volcanic eruptions. These weather events can occlude the satellite sensor; they create clouds … It makes it extremely challenging for somebody like the Red Cross to respond to an event.”

At present, Missing Maps uses a team of volunteers to go though satellite images and identify roads, towns, bridges, and other infrastructure. The volunteers manually update an open-source map called Open Street Map, which is laborious and time-consuming.

Intel’s AI Lab, in collaboration with Mila and CrowdAI, developed an image-segmentation model and used it to identify unmapped bridges in Uganda from satellite pictures. Object-detection approaches were discounted due to performance in favor of segmentation. Bridges were selected as a trial feature because they are critical infrastructure and are particularly vulnerable to natural disasters such as floods. Seventy previously unmapped bridges were discovered by the system; the Ugandan National Society can use this data to better plan evacuation and aid-delivery routes.

Uganda Map

The system identified 70 bridges across Uganda that were previously unmapped by either Open Street Map or the Ugandan Bureau of Statistics. (Image: Intel)

Satellite imagery can be particularly challenging to work with. The lack of an obvious frame of reference for up and down is challenging, said Bastidas. Also, images are not always taken from directly above, meaning the same feature may be seen from different angles. Differences in the local terrain as well as styles of infrastructure and architecture make it hard to train models on labelled data from other parts of the world. Even in images from the same country, terrain may look very different in summer and winter, and features such as bridges show huge variation in size and style.

Intel’s training dataset therefore had to come exclusively from Uganda. In fact, a section of Northern Uganda was used, which includes multiple views of the same bridges to enable models to learn about seasonal and nadir-angle changes.

The models started by looking for waterways and highway features, and any areas where a highway crossed a waterway was marked as a candidate point for a bridge. Known bridge locations within 30 m of any candidate points were discarded. Bounding boxes were added around these intersections, and then satellite images from areas in the bounding boxes were pulled. The models could then interpret the images to see whether they contained a bridge.

The models ran on second-generation Intel Xeon scalable processors (Cascade Lake) with DL Boost and nGraph. Bastidas said that these processors were chosen for their giant size; satellite images are often 1,024 square pixels, and it was desirable for the chip to process an entire image at once.

According to Bastidas, the next steps for the project may include the generation of models that can aid human mapping volunteers, perhaps predicting bridge locations but leaving the final decision to human eyes.

“We are also interested in trying to come up with ways to leverage existing open-source data to make models that are more robust, more generalizable, and can [work] with more tolerance for this geographically distinct area,” he said.