Super-Resolution and Deep Learning for Climate Downscaling

Statistical downscaling is crucial to local climate change adaptation and is a long-standing problem in the climate community with a vast literature of techniques, including a variety of approaches using the method of analogues. However, machine learning and computer vision advances over the past few years have not been leveraged for statistical downscaling. In this work, we show that dictionary learning applied to image super-resolution is a generalization of constructed analogues. We then discuss the connection between dictionary learning and super-resolution deep learning frameworks and present two deep learning architectures, Super-resolution Convolutional Neural Networks (SRCNN) and Deep Residual Learning (Resnet). SRCNN and Resnet are compared to Locally Constructed Analogues (LOCA) when downscaling daily precipitation at 1/8° over the contiguous United States. Furthermore, we discuss open research problems including the use of Generative Adversarial Networks to project realistic weather patterns and hypothesized generalizability of spatio-temporal non-stationarity.