Different from traditional radar exploration methods, we used a method named DeepRNN which is originally proposed by Pang et al. 2019, and incorporate optical flow during the model training. This study converts the nowcasting task to an images series prediction task.
Unlike straight input an images series to the model, the DeepRNN method splits the images prediction to spatial learning and temporal predictions. The spatial information is extracted by a convolutional neural network (CNN) to learn the static information of the cloud in single images such as structure, intensity and location.
The motion information is transformed from a recurrent neural network (RNN) model to learn the variation of the extracted high-level information between frames.
Eventually, the Representation Bridge Model (RBM) module in the DeepRNN framework further splits the spatial and temporal information flow, and connects them by bridges.
To ensure the motion information is preserved, we incorporate optical flow information during the prediction of the first several frames.
Traditional statistic scores are calculated to evaluate the performance of the proposed method. The results show that this method improves around 10% on each target (CSI, POD, FAR) than other state-of-the-art and conventional radar exploration methods on 2 hours forecasting of radar echo. In the future , this new method will be improved to incorporate the information from the NWP.
Key words: radar reflectivity, nowcasting, recurrent neural network (RNN), convolutional neural network (CNN) , DeepRNN
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner