library
Advancing geospatial deep learning applications withopen source high resolution satellite imagery and training data
April 24, 2018
The SpaceNetTM collaborators announced today that they are the recipients of the United States Geospatial Intelligence Foundation (USGIF)'s prestigious Industry Achievement Award. Presented at the GEOINT 2018 Symposium, the award recognizes one team that has made an outstanding achievement using geospatial intelligence services, solutions, and technologies.
SpaceNet, launched in August 2016 as a collaboration between CosmiQ Works, Radiant Solutions, and NVIDIA, is an open innovation project offering a repository of freely available imagery with co-registered map features. Prior to SpaceNet there was minimal availability of free, precision-labeled, high-resolution satellite imagery available for computer vision research. Today, SpaceNet hosts datasets developed by its own team, along with data sets from projects like IARPA's Functional Map of the World (fMoW).
"SpaceNet is designed to lower the barrier to entry for developers, researchers, and startups to access high-quality geospatial data," said David Lindenbaum, Principal Engineer at CosmiQ Works. "This enables the entire research and development community to focus on developing solutions to challenging geospatial problems instead of spending time on data acquisition. Such solutions have the potential to be honed to power new commercial, academic, and government use cases."
SpaceNet has hosted three prize challenges, each focusing on a different aspect of applying machine learning to solve difficult mapping challenges, such as building footprint and road network extraction from imagery. The SpaceNet team created a training dataset with over 5,700 square kilometers of DigitalGlobe imagery, more than 685,000 building footprints, and more than 8,000 km of road networks hosted on Amazon Web Services (AWS). SpaceNet algorithms aim to help automate foundational mapping, which benefits a broad spectrum of end-users from first responders to intelligence analysts.
"One of our goals with SpaceNet is to collaborate with world-class partners to advance machine learning for geospatial applications," said Tony Frazier, President of Radiant Solutions, a Maxar Technologies company. "We see SpaceNet as an important effort to enable developers with the training data and tools to establish commercial benchmarks and create the next generation of innovative computer vision algorithms. Machine learning will continue to benefit a range of uses from foundational mapping to advanced geospatial analytics by helping humans automate certain tasks to gain faster speed and larger scale than otherwise possible. Such technology ultimately gives back critical time to end-users such as geospatial analysts and decision makers."
"SpaceNet leverages GPU-accelerated deep learning to solve important challenges in geospatial intelligence," said Kevin Berce, Senior Director of Federal at NVIDIA. "SpaceNet illustrates what'spossible when you apply AI to solving difficult ISR challenges like automatic road detection from satellite imagery. The results will empower disaster relief, humanitarian efforts, first responders and more, ultimately saving lives. Beyond that, we see SpaceNet as an opportunity to build a community of developers all working toward a shared goal — using open data to make the world a better and safer place."
The fourth SpaceNet challenge will focus on the identification of objects at high off-nadir angles of collection, as opposed to straight down views, to mimic imagery that may be collected by the growing number of planned next-generation constellations with higher revisit rates. The challenge will launch mid-summer 2018.