SpaceNet and Intel: Remote Sensing Data for Deep Learning

SpaceNet is an organization dedicated to giving developers, researchers, and startups access to high-quality geospatial data. Over the last two and a half years, SpaceNet has open sourced a large, curated data set with over 6500 sq km of high-resolution satellite imagery, w ~800,000 building footprint labels and 8000 sq km of road network labels. SpaceNet also develops and administers data science challenges to solve the problem of extracting building footprints and road networks from satellite imagery at scale. We are excited to share our work with SpaceNet at the AWS Public Sector Summit. In our session, we discuss the challenges of deploying these machine learning algorithms in operational timelines, and how to accelerate delivery of relevant information derived from satellite imagery after a natural disaster. We are incredibly grateful for SpaceNet’s commitment to becoming the prominent open-source repository for remote sensing data and derived algorithms. We see remote sensing data as a source of future compute demand that requires very large memory models, lending itself well to both our Intel Xeon processors as well as our upcoming accelerators.

At Intel, we appreciate the challenges and benefits of working with remote sensing data:

  1. Unlike traditional computer vision imagery found in ImageNet or PASCAL Visual Object Classes, satellite imagery features varying densities and sizes of objects within the same class. Consider, for example, buildings — a massive warehouse can be labeled as a building, just as a small hut might be.
  2. Satellite imagery can feature more than just 8-bit RGB color values. Multispectral SpaceNet datasets include bands that highlight vegetation or water, as well as bands in the near-infrared spectrum. This opens up paths for richer spectral feature extraction, by trading off with transferability of traditionally pre-trained models.
  3. Unlike natural images, gravity doesn’t affect overhead imagery. Instead, models must be rotationally invariant, something that models trained on traditional datasets struggle with.
  4. Task difficulty is compounded by several factors, including the angle between the satellite and the target area, amount of cloud cover, and weather patterns. We recently released a dataset, joint with Spacenet, to encourage new algorithms that are robust to different viewing angles.

Monitoring our planet has been a long-standing mission, used by farmers for agricultural planning, cities for urban planning, disaster relief organizations for response planning, scientists for environmental research, and much more. As satellite constellations both become ubiquitous and improve in their ability to gather higher fidelity data, we must develop deep learning-based tooling and algorithms that work with this type of data. We believe this challenge is sufficiently distinct from traditional computer vision tasks to require special attention by researchers and developers.

For these reasons, we are proud to become members of the SpaceNet Foundation, and look forward to collaborating on the release of innovative datasets. To learn more about these challenges, and how to leverage Intel hardware and software to solve some of these issues, you can check out our workshop at last year’s Intel AI DevCon, where we showcased deep learning on SpaceNet’s Las Vegas dataset using Intel Xeon Processors.