Thread by @lak_gcp

DeepMind has just published a paper in Nature on doing nowcasting using neural networks. This paper uses generative models.

How does this compare to previous precipitation nowcasting models from Google?

nature.com/articles/s4158…

The first attempt was a U-Net implementation consisting of (1) spatial downsampler +CNN (2) spatial upsampler + CNN with skip connections to maintain knowledge at different scales.

This bettered the performance of optical flow methods, but not by much.
arxiv.org/abs/1912.12132

#2 MetNet (arxiv.org/abs/2003.12140) predicts probability of rainfall amounts > some value. It has: (1) spatial downsampler and CNN (2) temporal LSTM (3) spatial aggregator with self-attention to learn which part of input to focus.

MetNet gets much better accuracy scores. In fact, where the U-Net was barely able to beat HRRR after an hour, MetNet was able to successfully predict rainfall 8 hours ahead! This is far beyond what most people think nowcasting methods are capable of.

Exciting ... but ...

MetNet learns that creating a blurred output is an easier way to get high accuracy scores than to actually learn storm structures.

Variational generative models are probabilistic by nature (the encoder predicts probabilities; the decoder generates images from these)

So generative models provide a different way to represent uncertainty. Different than the traditional ensemble methods used in meteorology.

That said, the Deep Mind paper was done in the UK, and the UK doesn't really have much convective weather. So jury's out.

#wxtwitter