Algorithmic Hallucinations of Near-Surface Winds: Statistical Downscaling with Generative Adversarial Networks to Convection-Permitting Scales

Providing small-scale information about weather and climate is challenging, especially for variables strongly controlled by processes that are unresolved by low-resolution (LR) models. This paper explores emerging machine learning methods from the fields of image super-resolution (SR) and deep learning for statistical downscaling of near-surface winds to convection-permitting scales. Specifically, Generative Adversarial Networks (GANs) are conditioned on LR inputs from a global reanalysis to generate high-resolution (HR) surface winds that emulate those simulated over North America by the Weather Research and Forecasting (WRF) model. Unlike traditional SR models, where LR inputs are idealized coarsened versions of the HR images, WRF emulation involves non-idealized LR inputs from a coarse-resolution reanalysis. In addition to matching the statistical properties of WRF simulations, GANs quickly generate HR fields with impressive realism. However, objectively assessing the realism of the SR models requires careful selection of evaluation metrics. In particular, performance measures based on spatial power spectra reveal the way that GAN configurations change spatial structures in the generated fields, where biases in spatial variability originate, and how models depend on different LR covariates. Inspired by recent computer vision research, a novel methodology that separates spatial frequencies in HR fields is used in an attempt to optimize the SR GANs further. This method, called frequency separation, resulted in deterioration in realism of the generated HR fields. However, frequency separation did show how spatial structures are influenced by the metrics used to optimize the SR models, which led to the development of a more effective partial frequency separation approach.