Sebastian Scher, Stefanie Peßenteiner

Creating spatially coherent rainfall patterns with high temporal resolution
from data with lower temporal resolution is an important topic in
many geoscientific applications. From a statistical perspective, this
presents a high-dimensional and highly under-determined
problem. However, recent advances in unsupervised machine learning
provide methods for learning such high-dimensional probability distributions.
We show that it is possible to use Generative Adversarial Networks
(GANs) for estimating the full probability distribution of spatial
rainfall patterns with high temporal resolution, conditioned on a
spatial field of lower temporal resolution, requiring no knowledge
of the underlying processes. The GAN is trained on rainfall radar
data. Given a new field of daily precipitation sums, it can be used
to sample scenarios of spatiotemporal patterns with sub-daily resolution,
at very low computational cost. While the generated patterns do not
perfectly reproduce the statistics of the observations, they are visually
hardly distinguishable from the real patterns.