We’re excited to partner with SpaceNet LLC, a nonprofit focused on machine learning techniques for geospatial applications, to support the SpaceNet7 Multi-Temporal Urban Development Challenge which was just recently announced. The challenge focuses on developing better methods to track building construction over time using Planet imagery mosaics. Rapid and accurate remote-sensing of infrastructure change can aid in a variety of efforts, from infrastructure development to disaster preparedness to epidemic prevention.
Established in 2016 by In-Q-Tel’s CosmiQ Works, and DigitalGlobe (now part of Maxar Technologies), SpaceNet is dedicated to accelerating the research and application of open source AI technology for geospatial applications. SpaceNet offers free, precision-labeled, electro-optical and synthetic aperture radar satellite imagery data sets and runs challenges with prizes to foster emerging analytical frameworks. Previous competitions led to a dramatic expansion of the availability of open source data of building footprints and road networks for the geospatial machine learning community.
Whereas all six previous SpaceNet challenges were based on static road and building detection, the partnership between SpaceNet and Planet will allow the seventh competition to focus on discovery of change events directly. This technical challenge has been unlocked by access to a spatio-temporal dataset at a scale and cadence that was previously unavailable to the broader research community. Using dense time-series imagery of over 100 regions around the globe, this competition will allow researchers to experiment with novel change detection methods that would be infeasible on smaller datasets.
Planet images the entire landmass of the Earth on a daily basis at 3-5 meter resolution. The high temporal cadence and planetary coverage of our Dove constellation enables an entirely new class of remote-sensing applications, including detecting rapid urbanization, updating maps in a diverse set of geographical areas (not just dense urban areas), detecting unplanned infrastructure development, finding illegal activities in protected areas and uncovering precursors to deforestation.
The algorithms developed in this competition will also help address quantifying population statistics in regions of the world where Civil Registration Systems are not up-to-date, for which building footprints are a direct predictor. This will have an impact on disaster preparedness, the environment, infrastructure development and epidemic prevention.
In addition to the humanitarian benefits and practical applications, the data set presents novel challenges in the fields of remote sensing and machine learning. In contrast with common computer vision tasks, patterns are much more textural and occur at small scales. Combined with the strong background variability between images, object matching and tracking becomes trickier than in typical video deep learning models. We hope that challenges like this one will serve as catalysts for the development of novel techniques to extract more meaningful information from spatio-temporal datasets.
The Multi-Temporal Urban Development Challenge will start on August 31st and will last two months. The winners and the results will be announced at the AI conference Neural Information Processing Systems (NeurIPS) in December 2020.
As an integrated aerospace and data company, Planet’s goal is to image the entire world every day, making change visible, accessible and actionable. Having the right algorithms and tools to turn imagery into useful insights is key to realizing this promise.
This blog was co-authored by Planet’s analytics and modeling product manager Jonathan Evens and Planet’s analytics and modeling engineering manager Jesus Martinez Manso.