“There is a need to take advantage of the vast repositories of satellite imagery that we already have so that new breakthroughs in geospatial disciplines and applications can be facilitated,” says Tod Bacastow, director of strategic alliances for DigitalGlobe, which owns and operates a constellation of in-orbit commercial satellites, and has under management a large body of satellite imagery that can be further plumbed and exploited for commercial and non-profit applications.
Bacastow references a time seven or eight years ago when researchers at Stanford University expressed a need for more advanced computer visualization technology that could be applied to geospatial data. “They wanted the technology to develop smarter geospatial information that could tell us more, and they also recognized that there was a machine language challenge that they were bumping up against,” Bacastow says. “What was needed was a body of raw satellite imagery that researchers could work with so that machine-based intelligence could be trained to interpret this data and provide automation for certain geospatial mapping and analytics functions that users could benefit from.”
A group of companies that included DigitalGlobe and Nvidia created a satellite imagery data repository for researchers so they could work on machine training and automation — and the SpaceNet competition was launched.
“For the first competition, we wanted to focus on urban infrastructure mapping,” Bacastow says. “So we started with a data repository that contained 200 square kilometers of imagery with over 200,000 building footprints for Rio de Janeiro. What we wanted to do in training the machines was to take a lot of the manual mapping steps out of urban infrastructure mapping.”
Bacastow says that when maps are created today, the process begins with taking satellite imagery and creating map vectors, which collect geographic information at various levels of detail. “The individual doing the mapping then traces against this satellite imagery, but he also discovers that it doesn’t scale well and that its accuracy could be questioned. … This is a process which could be improved with the map drawer creating a polygon and attributing certain attributes to it, like the type of building that is being mapped, and what its various functions are. After this step, potentially a machine automation process could follow this by drawing the vectors.”
Equally important is the ability to analyze the vast numbers of satellite images that are collected on a daily basis. It is humanly impossible to do this, but machine-based intelligence and automation could be inserted into the process to help.
As machine automation techniques get refined and the resulting data becomes more accurate, more practitioners in disciplines that use geospatial technology will trust and rely on it. This is where installing metrics into the process that can measure results for accuracy becomes very important.
“We score for map accuracy and we use human checkers and determine how well the map aligns with the on-the-ground physical truth,” Bacastow says. “We also score each vector that we evaluate.”
It is friendly competitions like the SpaceNet Challenge that will energize researchers from around the world to see just how they can use machine automation to not take geospatial satellite imagery and mapping to the next level. The possibilities are unlimited. The mapping of urban and remote areas of the world can be automated with intelligence-machine-based technology. In areas like humanitarian aid and disaster mitigation, resources can be expeditiously directed to where they are needed. In autonomous vehicle navigation and monitoring, more accurate mapping and higher precision results can be obtained.
“By training machines to analyze and work with satellite imagery, and then opening it up to research, we are also opening up new opportunities in innovation,” Bacastow says.