Weather radar nowcasting of rainfall using optical flow with Recurrent All-Pairs Field Transforms (RAFT)

Weather radar nowcasting is a crucial technique in real-time urban hydrological applications. Uncertainties related to weather radar observations, motion field estimation, and rainfall evolution prognostication are inevitable. In this research, the deep learning model Recurrent All-Pairs Field Transforms (RAFT), which was developed for optical flow predictions in image processing and trained on synthetic data, is implemented to estimate motion fields in weather radar nowcasting. This research aims to determine whether this model has the potential to enhance the accuracy of rainfall prediction. The study is conducted using C-band weather radar data and 51 rainfall events, encompassing both linear and non-linear rainfall motion patterns, to systematically evaluate the prediction accuracy. The accuracy evaluations are performed by comparing RAFT-derived motion fields with those derived with the global vector, COntinuity of TREC (COTREC), Variational Echo Tracking (VET), Lucas-Kanade, and Dynamic and Adaptive Radar Tracking of Storms (DARTS) approaches. The RAFT model has been shown to perform statistically as well as the well-established, high-performing VET and Lucas-Kanade approaches. It is furthermore shown that RAFT, VET, and Lucas-Kanade outperform the global vector, COTREC, and DARTS methods. The presented results demonstrate that RAFT has significant potential for weather radar rainfall nowcasting, and that applying optical flow with RAFT for motion field estimation yields rainfall predictability comparable to that achieved by state-of-the-art methodologies.