WRF data for downscaling, used in Learned multi-resolution dynamical downscaling for precipitation

This study uses regional climate model (RCM) simulated precipitation at low and high spatial resolution, to develop convolution neural network (CNN) based approaches, that can emulate high resolution modeled data using low resolution modeled data with cheaper computational resource than running dynamical downscaling at the high spatial resolution. Specifically, we  define two types of CNNs, one that stacks variables directly and one that encodes each variable before stacking, and train each CNN type both with a conventional loss function, such as  Mean Square Error (MSE), and with a conditional generative adversarial network (CGAN), for a total of four CNN variants. We compare the four new CNN-derived high resolution precipitation with precipitation generated from a bi-linear interpolater and the state-of-the-art CNN-based super-resolution (SR) technique, using the original high resolution precipitation from the RCM as ground truth. We find that SR technique produces similar results to the interpolator with smoother spatial and temporal distributions and smaller data variabilities and extremes than ground truth shows. While the new CNNs trained by MSE generate better results over some regions than the interpolator and SR technique, their predictions are still not as close as ground truth. The CNNs trained by CGAN generate more realistic and physically reasonable results. This advanced technique improves not only the data variability in time and space, and but also the extremes, such as intense and long-lasting events, based on event-feature tracking algorithm.