Some Thoughts on Mechanical Turk and Geo-Processing

We use Amazon Web Services (AWS) quite a bit. Mostly we use the EC2 and S3, but recently we have been using a limited bit of Mechanical Turk (MTurk) for some testing of the web site.

For those of you who don’t know what MTurk is, from the web site -

…The Mechanical Turk web service enables companies to programmatically access this marketplace and a diverse, on-demand workforce. Developers can leverage this service to build human intelligence directly into their applications.

Our use has been somewhat limited to testing of the web site only. However, there has been some image processing uses of MTurk, including the SAR efforts to find Jim Gray and Steve Fossett.

I wear two hats these days. We are still actively involved in the development of HyperSpectral Imaging (HSI) sensors and algorithms (see the Florida Environmental Research Institute). It was from these efforts that we developed the cataloging, discovery, and distribution systems that we spun out into WeoGeo.

The holy grail of imaging techniques is the automatic extraction of features and classification of materials within the raster data. It is something we have been trying to develop for over a decade. There are others who have been working at it longer.

After all these years, there are some problems that are still difficult to solve in processing imagery. They frequently require just looking at the images frame by frame to resolve features and classify stuff that just defies algorithmic development. It strikes me that there may be some parts of this processing that may not be easily solved using computer algorithms. Things like finding seam lines in overlapping aerial photographs.

Several major imaging vendors send a chunk of their current image processing to low cost countries like China and India to complete their large-scale projects. It seems that there might be a better way to accomplish such geo-processing tasks that still require eyes then to incur the time and expense of sending these tasks overseas. Perhaps Mturk and some smart programming might offer a different approach.

I also wonder what other sort of QC/QA tasks in geo-processing might be solved by MTurk. I might try to kick it around a bit at GeoWeb. Find me if you got some thoughts.

(Ford assembly line, 1913)