Incorporating Prior Scientific Knowledge Into Deep Learning for Precipitation Nowcasting on Radar Images

Pattarapong Danpoonkij, Nutnaree Kleawsirikul, Patamawadee Leepaisomboon, Natnapat Gaviphatt, Hidetomo Sakaino, Peerapon Vateekul

Abstract: Precipitation nowcasting aims to precisely predict the rainfall intensity in the near future that can be applied in various applications. A common approach is to simulate the complex physical processes or extrapolate the rainfall from the current stage. The existing deep learning model for this task uses an end-to-end network to forecast, but this approach has often met with limited success due to the complexities of the problem. Therefore, this paper proposes a novel hybrid model that combines the scientific method from meteorology and the deep learning method from computer science. We experimented with the model on both simulated data and radar images. Also, we have created the simulated data to imitate important features from radar images. The results show that our hybrid modeling approach outperforms all baselines on almost all datasets (both simulated and the radar data).