Downscaling Numerical Weather Models with GANs

Alok Singh, Brian L. White, Adrian Albert

Abstract:

Abstract-Numerical simulation of weather is
resolution-constrained due to the high computational cost
of integrating the coupled PDEs that govern atmospheric
motion. For example, the most highly-resolved numerical
weather prediction models are limited to approximately
3 km. However many weather and climate impacts
occur over much finer scales, especially in urban areas
and regions with high topographic complexity like
mountains or coastal regions. Thus several statistical
methods have been developed in the climate community
to downscale numerical model output to finer resolutions.
This is conceptually similar to image super-resolution
(SR) [1] and in this work we report the results of
applying SR methods to the downscaling problem. In
particular we test the extent to which a SR method
based on a Generative Adversarial Network (GAN)
can recover a grid of wind speed from an artificially
downsampled version, compared against a standard
bicubic upsampling approach and another machine
learning based approach, SR-CNN [1]. We use ESRGAN
([2]) to learn to downscale wind speeds by a factor of 4
from a coarse grid. We find that we can recover spatial
details with higher fidelity than bicubic upsampling or
SR-CNN. The bicubic and SR-CNN methods perform
better than ESRGAN on coarse metrics such as MSE.
However, the high frequency power spectrum is captured
remarkably well by the ESRGAN, virtually identical
to the real data, while bicubic and SR-CNN fidelity
drops significantly at high frequency. This indicates that
SR is considerably better at matching the higher-order
statistics of the dataset, consistent with the observation
that the generated images are of superior visual quality
compared with SR-CNN.