A new deep learning technique increased the precision of short-term rainfall forecasts.
What's new: Suman Ravuri, Karel Lenc, Matthew Willson, and colleagues at DeepMind, UK Meteorological Office, University of Exeter, and University of Reading developed the Deep Generative Model of Radar (DGMR) to predict amounts of precipitation up to two hours in advance.
Key insight: State-of-the-art precipitation simulations struggle with short time scales and small distance scales. A generative adversarial network (GAN) can rapidly generate sequences of realistic images. Why not weather maps? A conditional GAN, which conditions its output on a specific input — say, previous weather history — could produce precipitation maps of future rainfall in short order.
How it works: Given a random input, a GAN learns to produce realistic output through competition between a discriminator that judges whether output is synthetic or real and a generator that aims to fool the discriminator. A conditional GAN works the same way but adds an input that conditions both the generator’s output and the discriminator’s judgment. The authors trained a conditional GAN, given radar images of cloud cover, to generate a series of precipitation maps that represent future rainfall.
- The generator took as input four consecutive radar observations recorded at five-minute intervals in the UK between 2016 and 2019. It used a series of convolutional layers to generate a representation of each and concatenated the representations. Given these observations and a random vector, a series of convGRU blocks (a type of convolutional recurrent neural network block) generated 18 grids that represented a 90-minute sequence of predicted precipitation per square kilometer.
- Two discriminators evaluated the generator’s output. A spatial discriminator made up of a convolutional neural network randomly selected eight of the 18 generated maps (for the sake of memory efficiency) and decided whether they were real. A temporal discriminator used 3D convolutions to process the 18 generated maps concatenated with the four input maps. Then it decided whether the generated sequence was real.
- In addition to the comparative loss terms, the discriminators used a loss term that encouraged the generator to minimize the difference, in each grid square, between real radar measurements and the average of six generated maps. This loss term increased the output resolution.
- At inference, the authors ran the generator multiple times and averaged the outputs. They used the variance to estimate uncertainty (for instance, a 20 percent chance of rain).
Results: The authors tested their approach at multiple time intervals and distance scales according to the continuous ranked probability score, a modified version of mean average error in which lower is better. Its output was on par with or slightly more accurate than that of the next-best competitor, Pysteps. Of 56 meteorologists who compared the generated and ground-truth precipitation maps, roughly 90 percent found that the authors’ predictions had higher “accuracy and value” than the Pysteps output with respect to medium and heavy rain events.
Why it matters: GANs can produce realistic images whether they’re cat photos or precipitation maps. A conditional GAN can turn that capability into a window on the future. Moreover, by averaging multiple attempts by the conditional GAN, it’s possible to compute the certainty of a given outcome.
We're thinking: Predicting the weather isn’t just hard, it’s variably hard — it’s far harder at certain times than at others. An ensemble approach like this can help to figure out whether the atmosphere is in a more- or less-predictable state.