A two-stage trained hybrid Unet-ConvLSTM2D for enhanced precipitation nowcasting

Current precipitation nowcasting models such as ConvLSTM, ConvGRU, PredRNN, and PFST-LSTM face challenges in capturing complex spatiotemporal patterns and preserving fine spatial details. These models often struggle with high-intensity precipitation events and lose accuracy over longer lead time. Although PFST-LSTM addresses spatial alignment and feature preservation and SAC-LSTM enhances long-range spatial dependency modeling through self-attention, they are constrained by increased computational complexity. In this study, we propose a hybrid Unet-ConvLSTM2D model that combines Unet’s superior spatial feature extraction capabilities with ConvLSTM2D’s temporal sequence modeling strengths. By introducing time-distributed layers in each Unet block, the model effectively handles temporal sequences while retaining high-resolution spatial features. The model is trained using a two-stage approach: pre-training on the moving-MNIST++ dataset to learn basic temporal dynamics, followed by fine-tuning on the CIKM AnalytiCup 2017 dataset to adapt to real-world meteorological data. Experimental results demonstrate that the proposed model significantly outperforms existing methods, with average improvements of 3.73% in critical success index (CSI) and 3.63% in Heidke skill score (HSS) across all dBZ thresholds, along with a 4.72% reduction in mean squared error (MSE).