Downscaling Numerical Weather Models With GANs (My CI 2019 Paper)


I saw some people on Twitter recently talking about work that’s similar to stuff I’d done last year (~April 2019). For a moment, I felt annoyance that I wasn’t cited, but then that quote about standing on each other’s shoulders and not each others’ toes came to mind, so I decided to blog about the paper since my marketing clearly wasn’t good enough.

Paper link

Background

I did this work while at Terrafuse, and the basic idea is to do super-resolution, but for the weather.

My first pass at it was to do the simplest possible thing and just pretend it’s an image problem.

Setup

Get some pics of weather and super-res them.

The interesting bit: Confusion about Results

Here’s a sample of the results: It goes ground truth vs GAN vs SR-CNN vs bicubic upsampling. There’s also a zoom-in of the red region.

So the GAN looks pretty good.

But it does worse on metrics like MSE and SSIM.

The GAN images have a sharpness to them that the other predictions do not have, so I looked at their power spectral density.

The fit is so close that plotting it basically overlaid 2 curves on each other and I wasted 2 days trying to debug what I thought was an issue in matplotlib.

Later experiments showed that GANs match spectral density and other statistics of the data very quickly, within 2 epochs. But to also get good image quality takes almost a 100. I had a few ideas about how to use this, that made it into another paper.

Related Posts

Why do things break in a 2 element set?

Braindead Way to Derive Taylor Series of Exponential Function

What is the prime factorization of zero?

Use of emphasis in speech

Generating a lot of language data with a theorem prover

"Litany Against Fear" in Present Tense

When it's time to party we will party hard

these are people who died

divine carrot

the frog