Fusing observations and model output for radar nowcasting

This presentation will describe the fusion of radar data, model data, and optical flow in a conditional Generative Adversarial Network (cGAN) to produce future radar. We focus on combining persistent radar imagery with variables known to affect convective initiation. This talk will outline the steps necessary to prepare such a dataset as well as the architecture of the model and implementation details. We document challenges with making this operational, including mitigation of grid imprinting artifacts, the choice of loss functions, the difficulties of resource allocation, and hyperparameter tuning. Preliminary verification results will also be presented.