Multi‐task framework of precipitation nowcasting

Abstract

Abstract Precipitation forecasting plays an important role in disaster warning, agricultural production, and other fields. To solve this issue, some deep learning methods are proposed to forecast future radar echo images and convert them into rainfall distributions. Prevailing spatiotemporal sequence prediction methods are usually based on a ConvRNN structure that combines a Convolutional Neural Network and Recurrent Neural Network. However, these existing methods ignore the image change prediction, which causes the coherence of the predicted image has deteriorated. Moreover, these approaches mainly focus on complicating model structure to exploit more historical spatiotemporal representations. Nevertheless, they ignore introducing other valuable information to improve predictions. To tackle these two issues, we propose GCMT‐ConvRNN, a multi‐ask framework of ConvRNN. Except for precipitation nowcasting as the main task, it combines the motion field estimation and sub‐regression as auxiliary tasks. In this framework, the motion field estimation task can provide motion information, and the sub‐regression task offers future information. Besides, to reduce the negative transfer between the auxiliary tasks and the main task, we propose a new loss function based on the correlation of gradients in different tasks. The experiments show that all models applied in our framework achieve stable and effective improvement.

Keywords

• deep neural networks
• motion estimation