Looks to me like an idea very easy to put in practise :

- Buy some recent high-res data, maybe from DMCII (http://catalogue.dmcii.com/). You can get relatively recent (month old) data for very cheap. Only issue is to find high-res enough data. - Link the image to a gg map / OSM equivalent for geocoding - Create mechanical turk or taskrabbit to search for pools / houses without pools

Do that for one neighborhood and see if you can make a sell. No automation needed, or at least no complex CV involved.

In Arizona you could also do something for solar panels. Solar panel companies are always trying to sell panels, but they can't tell you right away (a) what size/shape panels you can fit on your roof and (b) how much energy you're likely to generate, and what that's worth. I bet if they could give us that info (and it came from an impartial 3rd party like your startup) specifically customized for my house, they'd have a much better chance of selling me solar panels.

I had a similar idea for pool builders. Find streets where the majority of houses have pools and sell lists of the people without pools.

These guys have a similar concept. Except they really only seek to estimate materials quantities.

http://www.eagleview.com/

I have used their xml files previously to automatically tile roofs with solar panels.

Aerial drone imaging? It would be much more accurate and give you better resolution. Downside is harder to scale.

How do you identify a damaged roof? Is is super obvious?

But satellite imagery can be several years old.

Match the list against credit ratings?

Same could go for lawn care.