Multimodal Deep Learning in Precipitation Nowcasting: A Review from a Geometric Evolution

Deep learning has significantly advanced precipitation nowcasting by improving the representation of complex spatiotemporal atmospheric processes. However, multimodal fusion remains challenging due to geometric inconsistencies across heterogeneous meteorological observations, including radar grids, satellite imagery, station networks, and numerical weather prediction fields. This review provides a geometric evolution perspective for multimodal precipitation nowcasting by categorizing existing methods into four representation paradigms: Euclidean grid-based backbone models, non-Euclidean graph neural networks, global latent-space attention models, and probabilistic generative models. This taxonomy highlights how different architectures address geometric mismatch across modalities while enabling progressively richer spatiotemporal representations. We further summarize representative datasets, evaluation metrics, and key reproducibility challenges. Finally, we discuss emerging directions including meteorological foundation models, physics-informed learning, probabilistic forecasting, and interpretable multimodal fusion, which together point toward more robust and generalizable nowcasting systems.