Bleg ...

I recently accepted an invitation to talk about the environmental implications of machine intelligence and automation.  This is a little out of the main stream for me (most of us?), but the good thing about an open-ended talk like this is that one can ask a lot of questions and simply suggest possible answers!  But still, I need help posing the right questions and what's a blog for if I can't use it to beg for help?

The way I see it, I can approach this from two angles:

Effect of machine intelligence and automation on the environment, i.e. things such as:

• the use of robots to enable easier and more routine clean up of environmental disasters
• the ability to clean up areas that are unsafe for humans
• reduction in the need for physical infrastructure if travel, commutes, etc. are replaced by virtual meetings

The effect of the use of machine intelligence and automation in environmental decision making i.e. things such as:

• Do the errors common to machine intelligence (the inability to extrapolate, for example) make it harder to predict extreme events when such statistical methods become very common?
• As forecasts improve in accuracy, do the customers of such forecasts make "hard" decisions more often, i.e. instead of hedging their bets?

What do you think? Is there material I should reference? Incidents that I should mention? Angles I should consider? Send me suggestions!

A great (but not timely) search engine for scientific papers

From looking at referring web sites, I discovered that something called the e-prints network had a link to my list of papers from their Geoscience category. The site appears to be run by the US Department of Energy.

It's a pretty neat resource -- searching the site searches all the PDFs linked from the site, so you get a full-text search of scientific papers. For example, I did a search for "storm AND (identification OR tracking)" on all the Geosciences papers and got back a pretty nice list of 31 papers, each of which would be worth reading. I wish I'd discovered this site when I was writing a paper on the topic a year ago!

I did notice that even though one of my earlier (2003) papers on the topic is listed, my recent (2009) paper is not. So, they must have stopped indexing the linked eprints. I hope that this is not a sign that the funding for this project has run out.

Using satellite data to predict onset of disease

An interesting article about how an outbreak of malaria was prevented:

What the researchers at Goddard had noticed at the time of the first outbreak was that in the months preceding it, surface temperatures in the equatorial part of the Indian Ocean had risen by half a degree. These higher temperatures brought heavy and sustained rains, cloud cover and warmer air to much of the Horn of Africa. Mosquitoes multiplied wildly—and lived long enough for the virus that causes the fever to develop to the point where it is easily transmissible. In September 2007 the researchers saw the same thing happening in the ocean, and suspected the same consequences would follow.

Later on, the article goes on to talk about how ponds and their salinity were measured to gauge whether the mosquitoes would have the right environment to multiply.

Machine intelligence as art

Washington's National airport has an art exhibit in the lobby. On a recent trip, one of the pieces displayed brought me short:
Damn if that's not what applying watershed segmentation to a storm image would look like (see last panel of figure from a paper in J. Atmos. Research):
Time to polish off the images in your AI papers and sending them off to art shows?

Machine intelligence to catch poachers

A cool use of machine intelligence, to track tiger populations:

Using a formula developed by renowned tiger expert Ullas Karanth of WCS, researchers accurately estimate local populations by how many times individual tigers are "recaptured" by the camera trap technique. It is expected that the new software will allow researchers to rapidly identify animals, which in turn could speed up tiger conservation efforts.

The "formula" is simply a way to do pattern recognition based on the stripe patterns -- apparently, tiger stripes, like human fingerprints, are unique. Because the stripes are unique, a nice side-effect is that the software can be used to identify cases where tiger pelts have been poached. By cross-labeling the poached pelt with camera images, it may be possible to find out where the tiger was last seen alive, so narrowing down the potential suspects.

Cloud-to-ground lightning initiation and cessation algorithm

Some real-time images from an algorithm we've developed for predicting cloud-to-ground lightning activity over the next 30 minutes. I presented it at the AMS meeting, but not in the AI session!

The current lightning density:
The current reflectivity composite:
Lightning Probability at a location over the next 30 minutes:

Regionalization

For me, the most interesting thing discussed in today's AI sessions was Jenny's regionalization idea. The fact that you could spatially cluster AI models based on the best predictors and then get very reasonable and intuitive results was quite cool. This was the image from her talk that sold it for me: