Comparative study of cloud evolution for rainfall nowcasting using AI-based deep learning algorithms

It is a critical need to provide timely and valuable alerts of rainstorms and floods to the public. However, it still remains a world-class challenge to achieve serviceable nowcasting rainstorms with even a short lead time of one hour. Different deep learning algorithms have been adopted to improve nowcasting accuracy. Unfortunately, it is still a question which algorithm is more suitable and how to interpret the rainstorm nowcasting results from deep learning. To this end, this paper focuses on modelling the evolution of rainstorm clouds using deep learning algorithms that can be applied to nowcast rainstorms for the next few hours. Adopting three deep learning algorithms, the study provides a detailed analysis of the nowcasting results of three typical cases of different rainfall intensities from a radar echo mosaic image dataset. The dataset was collected in Guangdong, China, and the analysis interprets the performance differences. The analysis further discloses that an AI-based method can provide more skilful nowcasting for medium and strong rainfall cases than for weak ones. Moreover, a deep learning algorithm trained by the dataset for one region can be skilfully used to nowcast rainfall for another region with a similar weather system. This explains the nowcasting capability of deep learning algorithms as well as their robustness. Besides, experiments on the number of iterations reveal that more iterations do not achieve higher nowcasting accuracy. With improved interpretability of deep learning from the perspective of real-world application in the study, it is expected that the algorithms producing higher accuracy and longer lead time nowcasts will be made possible.