A Lightweight Multimodal Spatiotemporal Learning Framework for Precipitation Nowcasting

Abstract

Accurate and timely precipitation nowcasting is essential for meteorological hazard mitigation, particularly at monitored sites with limited sensing inputs. Capturing the nonlinear, regime-dependent relationship between precipitable water vapor (PWV) and rainfall remains challenging. This study proposes a lightweight multimodal spatiotemporal learning framework with physics-aware inductive biases for station-centered, short-horizon nowcasting within an existing or deployable monitoring network. The framework combines global navigation satellite system-derived PWV with environmental-condition variables to generate probabilistic forecasts of threshold-based precipitation events. Its architecture integrates a spatially aware multimodal module, an efficient fusion stage, and a receptance weighted key-value-based temporal dynamical module to capture spatial correlation, multimodal complementarity, and long-history temporal dependencies under practical computational constraints. Evaluation involved more than 100 stations across three climatically distinct Chinese provinces: Zhejiang, Hubei, and Sichuan. Experiments focused on 1-hour nowcasting, with extensions to 3-hour forecasts. Ablation studies confirmed the contributions of multimodal integration, long-temporal modeling, and spatial correlation modeling. Compared with the iTransformer baseline under matched evaluation conditions, the framework consistently achieved better discrimination and lower false positive rates. Area under the receiver operating characteristic curve values exceeded 91% in the best-performing settings, while the false positive rate for heavy-rain events was approximately 21% at a true positive rate near 87%. Strong first-hour-of-rain accuracy further supports its potential for early warning. These results demonstrate an efficient approach to station-centered precipitation nowcasting from sparse observations.