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




Annau, Nicolaas J.

Journal Title

Journal ISSN

Volume Title



Providing timely and accurate small-scale information about weather and climate is challenging – especially for variables strongly controlled by processes that are unresolved by low-resolution (LR) models. Motivated by this challenge, this thesis employed emerging artificial intelligence methods from the fields of computer vision and deep learning for the statistical downscaling of surface wind variables to convection-permitting scales. Specifically, Generative Adversarial Networks (GANs) were conditioned on LR inputs to deterministically generate possible arrangements of high-resolution (HR) surface wind patterns using super-resolution (SR). SR models were trained over several subregions in the contiguous United States and southern Canada, allowing for a systematic analysis of the methods. In addition to matching the statistical properties of the target dataset, GANs generate fields with impressive realism and computational efficiency, making them attractive for operational statistical downscaling. However, objectively assessing the realism of the SR models required a careful selection of evaluation metrics. Using power spectra successfully revealed how altered GAN configurations changed spatial structures in the generated fields, where biases in variability originate, and the role of including additional LR covariates in the SR methods. Inspired by recent work in the computer vision field, a novel methodology that separates spatial frequencies in HR fields was used in an attempt to optimize the SR GANs further. This method, called frequency separation, ultimately deteriorated the realism of the generated HR fields. However, frequency separation revealed how spatial structures are influenced by the metrics used to optimize the SR methods. A covariate-sensitivity analysis was also performed, illustrating how physical relationships between the large and small scales are mirrored in the GANs. Furthermore, a methodology was used to evaluate serial dependence in the SR methods. Even without explicitly including temporal information while training, serial dependence in the fine-scale structures of the generated fields was present. Although this thesis did not use fully-stochastic SR models (as has been done in existing work), the results provide valuable insights into variability, biases, covariates, and the optimization problem of SR.



climate science, artificial intelligence, deep learning, downscaling, super-resolution, computer vision, generative modelling