A Heterogeneous Spatiotemporal Attention Fusion Prediction Network for Precipitation Nowcasting

Abstract

Precipitation nowcasting underlying various public services from rainstorm warning to flight safety is quite important and remains challenging due to the fast change in convective weather. Although some deep learning models have been proposed to make prediction automatically, most of them just deal with a single radar echo data source, making them hard to adapt to heterogeneous and diverse data in practice. In this article, a heterogeneous spatiotemporal attention fusion prediction (HST-AFP) network is proposed for radar echo extrapolation (deterministic output) and further precipitation nowcasting, which deals with mining and fusing knowledge from multiple heterogeneous spatiotemporal (ST) data sources, including history radar echo observations and numerical weather prediction data. With the help of the proposed attention-based ST diffusion module, the multiencoder is designed to extract information from both dense ST tensor and sparse ST tensor. On the other hand, the fusion decoder achieves very deep trainable residual fusion prediction by integrating scalewise attention fusion module and deep residual spatial and temporal attention mechanism. It can adaptively blend multisource ST features and rescale the multiscale temporalwise and spatialwise features for better prediction. Experiments in a real-world dataset of South China show that compared with the ingenious recurrent-neural-network-based methods and newly proposed UNet-based methods, our HST-AFP network can handle complex input with heterogeneity in both space and time domains, and performs better on the precipitation nowcasting metrics as well as requires remarkable shorter forecast time.

Keywords

• Deep residual spatiotemporal attention (DRSTA)
• heterogeneous spatiotemporal (ST) data
• precipitation nowcasting
• ST diffusion