Informal seminar: A Machine Learning Approach to Stochastic Downscaling of Precipitation Forecasts
by ,
Despite continuous improvements, precipitation forecasts are still not as accurate and reliable as those of other meteorological variables as key processes occur below the grid-scale of global weather models. Generative adversarial networks (GANs) have been demonstrated by the computer vision community to be successful at super-resolution problems, i.e. learning to add structure to coarse images. Leinonen et al. (2020) previously applied a GAN to the atmospheric sciences problem of downscaling: improving the spatial resolution of low-resolution images by constructing an ensemble of possible high-resolution images.
We build on this work and employ two methods -- GANs but also Conditional Variational Autoencoders (CVAEs) -- to demonstrate that downscaling can also be learned when mapping ECMWF's weather predictions with the Integrated Forecasting System (IFS) at ~9 km resolution to precipitation observations over the UK at much higher resolution. Here, the machine learning task must learn to add resolution while correcting for model error. We show that machine learning tools can match the statistical properties of point-wise methods while creating spatially coherent precipitation maps.