Tremendous-resolved rainfall prediction with physics-aware deep studying
Authors: S. Moran, B. Demir, F. Serva, B. Le Saux
Summary: Rainfall prediction on the kilometre-scale up to a couple hours sooner or later is vital for planning and security. However it’s difficult given the complicated affect of local weather change on cloud processes and the restricted ability of climate fashions at this scale. Following the set-up proposed by the emph{weather4cast} problem of NeurIPS, we construct a two-step deep-learning resolution for predicting rainfall incidence at floor radar excessive spatial decision ranging from coarser decision climate satellite tv for pc photos. Our method is designed to foretell future satellite tv for pc photos with a physics-aware ConvLSTM community, which is then transformed into precipitation maps by a U-Web. We discover that our two-step pipeline outperforms the baseline mannequin and we quantify the advantages of together with bodily data. We discover that local-scale rainfall predictions with good accuracy ranging from satellite tv for pc radiances might be obtained for as much as 4 hours sooner or later.