Exploring a simple CNN/RNN for radar nowcasting in memory-constrained environments

State of the art nowcasting solutions are usually very resource intensive, owing to the fact that most of them produce all outputs at the same time. This is done to minimize accumulation of errors in the recurrent predictions. This talk will explore a smaller RNN/CNN architecture using design decisions inspired by larger architectures such as Deep Generative Modelling of Radar and NowcastNet including proper initialization of recurrent layers, inclusion of auxiliary data such as terrain and meteorological fields, and manipulation of loss functions that explicitly penalize dissipation of radar retrievals. We compare our results to an operational baseline which uses optical flow.