Deep Temporal Interpolation of Radar-based Precipitation

Abstract

When providing the boundary conditions for hydrological flooding models and estimating the associated risk, interpolating precipitation at very high resolutions in time, say, for horizons of 10 minutes, is desired not to miss the critical cause of flooding in local regions. In this paper, we study optical flow-based interpolation of weather radar images from satellites, which are globally available nowadays. The proposed approaches are based on a latest deep neural network for multiple video frames interpolation, while terrain information is combined to use with temporarily coarse-grained precipitation radar observation as inputs for self-supervised training. The experiment with the Meteonet radar precipitation dataset for Aude 2018 flooding demonstrated the advantage of the proposed method over a linear interpolation baseline, up to 20% error reduction.