Exploring deep learning-based approaches for precipitation downscaling

Despite the remarkable performance of Numerical Weather Prediction and Global Climate Models across a wide range of applications, the accurate simulation of local-scale precipitation and the reliable reproduction of its distribution is still a significant challenge. The limited spatial resolution of global models is among the primary factors undermining their forecast capabilities in this context. Indeed, the physical mechanisms underlying the onset and development of precipitation, particularly in extreme events, operate at spatio-temporal scales smaller than those explicitly resolved, thus struggling to be accurately captured. Downscaling aims at bridging the gap between the spatial resolution of models output, and the resolution needed for local-scale applications. Generative Adversarial Networks (GANs) have been successfully applied to super-resolution tasks, which involve enhancing the spatial resolution of images. Reasons of similarity provide a compelling motivation for applying this technique to the task of downscaling. This thesis explores the application of a conditional deep convolutional GAN to precipitation downscaling. The capabilities of the GAN are demonstrated in a perfect-model setup. Specifically, the spatial resolution of a precipitation dataset is artificially degraded, and the GAN is employed to restore it. The developed model exhibits superior performance when compared to one of the leading precipitation downscaling methods found in the literature. The GAN not only generates a synthetic dataset featuring plausible high-resolution spatial patterns and intensities, but also produces a precipitation distribution with statistics closely mirroring those of the ground-truth dataset.