TyrainNow: A Deep Learning-Based Model for Typhoon Rainfall Nowcast With Radar Products

Abstract

Tropical cyclone (TC)-induced rainfall is a drastic threat to human life and property, and thus rational rainstorm nowcasts within even a lead time of few hours play a key role in disaster mitigation. While recent deep learning-based algorithms have shown promise, predictions commonly suffer from the troubles of blur, dissipation, and location errors of TC rainbands, particularly for a lead time beyond 1 hr. Here, we develop a new nowcasting model, named TyrainNow, and show a significant improvement for nowcasting rainbands with a lead time up to 2 hr. Concretely, TyrainNow employs a refined multi-task loss function integrating geographical consistency, temporal coherence and radar image structural similarity. This tailored enhancement is architecture-agnostic and involves subtle adjustments. Secondly, TyrainNow adopts the quantile mapping technique to correct systematic attenuation biases inherent in the neural network outputs. The new model is verified on the basis of typhoon radar composite reflectivity products in South China, with a focus on the Greater Bay Area. Specifically, the new model achieves a critical success index of 0.099 at the 40 dBZ threshold, marking a substantial improvement from 27% to 330% compared to three other benchmark models, DGMR (0.070), PredRNN-v2 (0.023), and optical flow model (0.078), averaged over the lead times between 1 and 2 hr. We further verify the model's explainability and generalizability, and recommend it as a scalable and reliable model.