An Improvement Multitask Transformer Network for Dual-Polarization Radar Extrapolation

Severe convective weather, a meteorological phenomenon, is distinguished by its abrupt initiation, swift propagation, extreme atmospheric conditions, and formidable capacity for destruction, all of which have a significant impact on human productivity and livelihoods. In this study, we introduce an improved multitask transformer-based algorithm, MT-Transformer, for predicting parameters related to dual-polarization radar. These parameters encompass radar reflectivity Zh, differential reflectivity Zdr, and specific differential phase Kdp so that more information about the dynamic structure of convective storms can be obtained. MT-Transformer has the following main improvements. First, to overcome the insensitivity of the transformer to high-frequency information, before the data are fed into the transformer, 2-D convolution operation is used for the downscaling and image feature extraction. Second, for the input side of the Transformer decoder, we develop the future feature extraction (FFE) module for multiscale feature prediction, which is a structure for capturing the global multiscale contextual information of multivariate. Third, the fully connected structure in the feedforward network is replaced by a 3-D convolution layer, which can effectively extract the spatiotemporal information of precipitation. The quantitative results demonstrate a reduction in the root-mean-square error (RMSE) values for Zh, Zdr, and Kdp by 14.48%, 3.47%, and 6.91%, respectively, in comparison to those of the second-best MIM model for a 1-h forecast duration.