SpaceNet 5 Dataset Release

Adam Van Etten & Ryan Lewis

Preface: SpaceNet LLC is a nonprofit organization dedicated to accelerating open source, artificial intelligence applied research for geospatial applications, specifically foundational mapping (i.e. building footprint & road network detection). SpaceNet is run in collaboration with CosmiQ Works, Maxar Technologies, Intel AI, Amazon Web Services (AWS), and Capella Space.

The breadth of challenges that can be addressed by overhead imagery is impressively broad, and continues to grow as new and improved systems are deployed. Yet a lack of high quality labeled training data continues to impede progress in many areas of remote sensing analytics. To this end, the SpaceNet partners are proud to report the release of the SpaceNet 5 dataset, which adds significantly to the SpaceNet data corpus.

SpaceNet 5 focuses on automated road network extraction and route travel time estimation from satellite imagery. For this challenge we label four new cities in diverse locales:

• Moscow, Russia
• Mumbai, India
• San Juan, Puerto Rico
• Mystery City

This brings the total number of SpaceNet cities up to 10. The six previous SpaceNet cities are: Rio De Janeiro, Las Vegas, Paris, Shanghai, Khartoum, and Atlanta. For the first time, we withhold one city (our “Mystery City”) for final testing, which should improve model robustness and deployability to new regions by ensuring that challenge participants don’t overtrain their model for the existing cities. We will release the imagery and labels for the Mystery City after the challenge has concluded and the winners announced. Details of the new dataset are shown below:

Figure 3. Puerto Rico test image chip. Left: Panchromatic image, center: 8-band multispectral image, right: RGB image.

Road centerline and metadata tags were hand-labelled by a team of experts for each roadway in the areas of interest. These metadata tags allow us to infer a reasonable speed limit for each roadway (for example, a one-lane residential road likely can be traversed at 25 mph). The precise vector labels and speed estimates for each roadway permit a unique temporal aspect to be included in SpaceNet 5: we challenge competitors to extract both roadways and travel time estimates for each road (see our blog for an algorithmic baseline). Such optimized routing is fundamentally crucial to a great number of challenges.

Figure 1. Sample image chips over Moscow, Russia. Left: raw image demonstrating some of the challenges of this area of interest: snow and shadows. Right: SpaceNet labels overlaid, with a speed limit color scale of yellow (20 mph) to red (65 mph).

Figure 2. Sample image chips over Mumbai, India. Left: 8-band multispectral image. Right: The same location in RGB with SpaceNet labels overlaid, with a speed limit color scale of yellow (20 mph) to red (65 mph).

What’s Next

The challenge will kick off on Tuesday, September 3rd and the SpaceNet 5 challenge registration site is live on the Topcoder SpaceNet Challenge site. If you are new to Topcoder, then you are welcome to register here. As always, stay tuned to The DownLinQ and check out https://spacenet.ai for more details on the dataset and challenge.

Special thanks to Nick Weir, Jake Shermeyer, and Daniel Hogan for their assistance in preparing the dataset.