A deep learning based physically-consistent super-resolution approach to climate downscaling

Physics-based climate models that incorporate process understanding, such as global climate models (GCMs), are usually run at coarse resolutions (for example, 100km) due to high computational costs. However, climate adaptation planning and policy- and decision-making often requires assessing the impact of climate change on regional and local levels, for which the availability of reliable fine-scale data is crucial. Many methods exist for downscaling climate data --- retrieving high resolution data from low resolution inputs --- and broadly fall into the categories of dynamical and statistical downscaling. The former is often preferred for its accuracy and physical consistency, but it is still computationally very demanding.

In recent years, deep convolutional neural networks (CNNs) have emerged as a promising alternative to statistical downscaling where a deep learning model is trained to learn the relationship between low resolution climate variables and their high resolution projections. This problem closely resembles single-image super-resolution (SR), a classic problem in image processing where the goal is to enhance a low-resolution image to approximate its true high-resolution image counterpart. In this study, we improve upon a deep generative super-resolution method by integrating established physical knowledge to downscale several atmospheric variables, including humidity and vertical winds (omega at 500 mB). A key motivation for the choice of variables is to predict better extreme precipitation events on fine scales. We downscale by a factor of four in each dimension and compare results against standard off-the-shelf methods, such as bicubic interpolation. Our approach not only captures the distribution of extremes more accurately, but also the power spectra of winds and humidity, a key requirement for maintaining physical consistency of complex weather and climate phenomena at fine scales.

This work is affiliated with the Jupyter Meets the Earth project, which is supported by the NSF EarthCube program (awards 1928406 & 1928374).