PlanetScope imagery of Houston, Texas following heavy floods in 2017 © 2017, Planet Labs Inc. All Rights Reserved.
Curious Planeteer working to make the Earth's changes visible, accessible and actionable.

Slingshot Aerospace Maps Flooded Areas following Hurricane Harvey


When Hurricane Harvey made landfall in Houston earlier this month, Slingshot Aerospace, a Planet Application Developer Program partner, leapt into action. Working with Team Rubicon and the BAE Systems Geospatial eXploitation Products™ (GXP®), Slingshot was able to use its flood detection algorithms on up-to-date satellite imagery to determine the extent of the flooding and its severity. With this knowledge, they were able to highlight dry areas to assess search and rescue efforts to enable first responders and medical crews to safely reach those in need. Slingshot also used its algorithms to quickly identify flood relief and triage staging areas close to non-flooded roadways and optimal hospital routes. BAE Systems leveraged their cloud-hosted GXP solutions to provide Team Rubicon’s disaster response teams with the ability to visualize and analyze Planet imagery and Slingshot products.

Take a look at the mapping they’ve done in the storm’s aftermath:


Slingshot’s flood mask of West Houston overlaid on RGB (data source: Planet)

Slingshot’s flood depth mask highlighting varying degrees of flooding (data source: Planet)

Map of observed flood extent using cloud penetrating SAR (data source: Sentinel-1)
GXP-Slingshot overlay3_small

Slingshot’s Flood Mask layered on BAE’s GXP combined with geocoded imagery (source: Planet) of Memorial Dr. confirming mask accuracy (source: BAE)

Identification of low risk staging areas with easy access to highways and hospitals.
Given the appropriate raw imagery, labels, and ground truth images, Slingshot’s building extraction model makes a pixel-by-pixel classification and creates a semantically segmented mask. Slingshot ran tests over two datasets for Los Angeles, CA, and Corpus Christi, TX. Each dataset consisted of 200 different RGB, 15cm resolution, 2500x2500px images with an associated ground truth. Each image dataset was split into two components with 80% of the images reserved for testing and 20% for validation. Using only two datasets covering Los Angeles, CA and Corpus Christi, TX, the model has reached an F1 score of 92.3% accuracy.

This is powerful example of how geospatial analytics and up-to-date imagery of anywhere can affect the speed and accuracy of disaster response and long-term recovery.