Super resolving spatial-temporal weather predictions using Deep Learning

Weather predictions typically result from observations on a sparse network of weather stations or models at coarse spatial resolution. However, the impact of weather events such as floods and hailstorms is hyper-local. We propose to use Generative Deep Learning Models to downscale weather predictions and improve the spatial resolution. Besides typical reconstruction loss at training time, additional constraints from simple laws of physics are introduced. We test the model on global weather data and compare the results for cases with and without constraints from the laws of physics. The models are found to be more robust when physical constraints are applied and provide more accurate results.