Self-supervised learning for climate downscaling

Earth system models (ESM) are computer models that quantitatively simulate the Earth's climate system. These models are the basis of modern research on climate change and its effects on our planet. Advances in computational technologies and simulation methodologies have enabled ESM to produce simulation outputs at a finer level of detail, which is important for policy planning and research at the regional level. As ESM is a complex incorporation of different physical domains and environmental variables, computational costs for conducting simulations at a finer resolution are prohibitively expensive. In practice, the simulation at the coarser level is mapped onto the regional level by the process of ''downscaling''. In this presents a self-supervised deep-learning solution for climate downscaling that does not require high-resolution ground truth data during the model training process. We introduce a self-supervised convolutional neural network (CNN) super-resolution model that trains on a single data instance at a time and can adapt to its underlying data patterns at runtime. Experimental results demonstrate that the proposed model consistently improves the climate downscaling performance over the widely used baselines by a large margin.