Dual-Polarisation Radar Nowcasting with ConvLSTM and Squeeze-Excitation Attention

Deep learning nowcasting framework for X-band dual-polarisation weather radar. A ConvLSTM encoder with Squeeze-Excitation (SE) channel attention predicts the next radar frame from a sequence of 4 input frames across 8 dual-polarisation channels (DBZH, ZDR, RHOHV, PHIDP, KDP, VRADH, WRADH, SQI). Developed for a Furuno WR-2100 radar near Debrecen, Hungary. Includes pretrained model weights, training/evaluation scripts, and HDF5 datasets for 17 precipitation events. The 8-channel model achieves RMSE=0.089 and SSIM=0.752, outperforming both persistence baseline and single-channel (DBZH-only) ablation. MIT licensed.