Generative Adversarial Models for Extreme Super-Resolution of Climate Datasets

Keywords: Super-resolution, Machine learning, Climate data, Uncertainty modeling
Abstract Type: Paper Abstract

Abstract

Accurate and high-resolution climate datasets are fundamental in climatic research and decision-making. Many existing datasets, however, such as the outputs of Global Climate Models (GCMs), are only available at very coarse spatial scales. Many deep learning-based methods have been developed for super-resolution of climate datasets, in particular generative adversarial nets (GAN) models and the variants, and have shown great successes. In this paper, we describe a GAN-based sampling framework for the super-resolution of climate datasets. Compared to most existing methods that work best for super-resolution with low scaling factor (4X or 8X), the framework was able to generate high-resolution realistic images from very low-resolution inputs. More importantly, given an input the framework can produce a set of plausible high-resolution outputs instead of one single deterministic result. These plausible samples allow us to empirically explore the uncertainty spaces of the manifold implied by the input images. With a case study of climate datasets (wind and solar), we demonstrate the performance of the framework in super-resolution with very high scaling factors (up to 64X), and further highlight the advantages of the framework with a comprehensive comparison with commonly used downscaling methods, including area-to-point (ATP) kriging, deep image prior (DIP), enhanced deep super-resolution network (EDSR), enhanced super-resolution generative adversarial networks (ESRGAN), and physics-informed resolution-enhancing GANs (PhIRE GANs).