An Efficient Spatiotemporal Representation Learning Model for Quantitative Precipitation Estimation

Real-time and accurate rainfall estimation, also called quantitative precipitation estimation (QPE), is critical for environmental protection and water resources management. Weather radars are the main instruments for QPE tasks in many nations. However, how to represent the complex nonlinear relationship between radar images and real precipitation intensities, as well as to explore the spatiotemporal dependencies of precipitation for more accurate precipitation estimations remains a challenging problem. This study develops an efficient spatiotemporal representation (ESTR) learning network that uses multi-head attention to capture global information and a Gated Recurrent Unit (GRU) to extract temporal features after residual blocks. In particular, the dual-polarization radar images are utilized as inputs, and point-wise rain gauge measurements as ground truths. The performance of the ESTR network is evaluated using radar images collected from U.S. Weather Surveillance Radar - 1988 Doppler (WSR-88D) during 2016-2019 in Melbourne, Florida. The experimental results show that the ESTR models exhibit superior performance than conventional QPE methods, which achieves the best evaluation metrics, with a mean absolute error (MAE) of 0.89mm, root mean squared error (RMSE) of 1.83mm, normalized standard error (NSE) of 29% and correlation coefficient (CC) of 0.93. These results suggest that the ESTR learning network has great potential for regional rainfall estimation and mapping.