I recently got a call from a humanitarian colleague in the field who asked whether it would be possible to crowdsource the basic analysis of satellite imagery. They wanted to know because their team was sitting on a pile of satellite imagery but did not have the time or staff to go through the high-resolution pictures. They wanted to use the imagery to identify where IDPs were located in order to know where to send aid via helicopters.
My colleague’s question reminded me of the search for Steve Fossett, a famous adventurer who went missing in September 2007 after taking off from a small airport in Nevada in a small single-engine airplane. The area where Steve went missing is particularly rugged terrain. The search and rescue aircraft were not able to find any sign of wreckage. However, high-resolution satellite imagery from GeoEye enabled Amazon to produce a Help Find Steve Fossett website, allowing volunteers to search small sections of the available imagery.
“This is an approach to more rapidly search a large area of imagery using many eyeballs of people around the world. A similar technique was used to search for Jim Gray, a Microsoft scientist who went missing on his sailboat off the coast of California.”
Micro-tasking the analysis of satellite imagery has already been done. So why not in the context of disaster response? One could add this feature to a platform like Crowdflower, which is already being used as a plugin to micro-task the processing of text messages from disaster affected areas. Instead of text, volunteers would see a small subsection of satellite imagery. They’d be asked whether they could see any evidence of individuals in the imagery and if so how many approximately they can make out. A simple 5-minute guide on how to identify people and approximate population size using satellite imagery could be put on YouTube for volunteers to watch before getting started.
Like any type of micro-tasking approach (a.k.a. mechanical turk service), one could triangulate answers to maintain some level of quality control. For example, only when 10 volunteers each tag an image as having individuals in it would the picture be processed as such. The same would apply to the population ranged estimated in a given image. This wouldn’t necessarily produce perfect results, but it would take the bulk of the load off the shoulders of humanitarian on the ground. It would act as a first filter.
Of course the obvious question that arises is security and privacy. There are several ways this could be addressed. First, images would be stripped of any GPS coordinates. Second, images would be sliced up in small bits to prevent easy recognition of the territory. Third, a volunteer would not be given contiguous slices so they couldn’t piece together more information from the satellite imagery. These measures won’t provide 100% security and privacy. The only way to achieve that would be to use bounded crowdsourcing, i.e., only have trusted individuals analyze the imagery.