News for hackers. Stuff that matters.
Re:Counting solar panels in the U.S. with machine learning a... by dmurray on Friday February 01, 2019 @03:36AM (#19052582)
Why spend a month of computer time to evaluate all one billion images? A randomly chosen million would seem to give you excellent data, more than enough to make the statistical conclusions the article talks about. Maybe you'd run the chance of misrepresenting large-scale commercial installations, but you could correct for that manually, and that's already a risk since the methodology skips some rural areas. Maybe it's simply good advertising to say you did the whole thing. But more likely the real purpose of the survey is to sell the gathered information at a much more granular level: zip codes, streets, or better still, individual households.
Re:Counting solar panels in the U.S. with machine learning a... by AznHisoka on Friday February 01, 2019 @09:15AM (#19054199)
What's the usefulness of analyzing just a random sample? They need to analyze 1 billion images so they have granular data on what % of homes in specific zip codes, towns, etc have solar panels installed. This is actionable data. IE. if there's lower % usage in a specific zip code, they might implement some action to improve it in that specific zip code.

So of course, the value is all in the granularity. They already trained the model on random images.

Re:Counting solar panels in the U.S. with machine learning a... by Scoundreller on Friday February 01, 2019 @10:58AM (#19055128)
Agreed.

If you can figure out exactly where the residential rooftop panels are, you’ve found a goldmine of people susceptible to door-to-door sales of “get 10% returns guaranteed per year by letting us sell you/install/use your ______”.

Re:Counting solar panels in the U.S. with machine learning a... by yorwba on Friday February 01, 2019 @04:21AM (#19052757)
You can download data at the census tract level from their project website [1]. Their code is on GitHub [2]. You can download their pretrained models.

I think your speculation about monetization is unfounded.

[1] https://web.stanford.edu/group/deepsolar/home

[2] https://github.com/wangzhecheng/DeepSolar

Re:Counting solar panels in the U.S. with machine learning a... by lawlorino on Friday February 01, 2019 @07:36AM (#19053589)
> I think your speculation about monetization is unfounded.

I think it isn't bad speculation at all - I used to work in the energy industry as a data scientist and I know there are several US energy companies working on this exact problem.

Re:Counting solar panels in the U.S. with machine learning a... by jsight on Friday February 01, 2019 @08:49AM (#19054020)
I'm not sure that sampling would work in this case. There's a real risk, IMO, of ending up with sampling bias.

For example, some communities are very anti-solar and the local "homeowners" association blocks them. Others are favorable and approve them everywhere having relatively high density.

Maybe there's a way to account for this, but I'm afraid that the clustering of installs might cause issues.

EDIT: And, to go to your final point, yes, I suspect one goal is to be able to provide really accurate data at a zip code level. I can imagine lots of useful reasons for that, some related to what I said above, too.