An NWP-Free, Observation-Driven Deep Learning Approach to Heavy-Rainfall Nowcasting Beyond the Three-Hour Limit

Abstract

Quasi-stationary convective bands over Kyushu, Japan, frequently trigger rainy-season disasters, and hours with ≥50 mm h −1 rainfall are increasing. However, skillful nowcasts beyond 3 h remain limited. This study presents FlowsNet, an observation-based multisensor fusion model that learns directly from radar/rain gauge-analyzed precipitation, surface variables from ground stations, geostationary satellite imagery, and satellite-derived precipitation context. The model targets category-4 (C4; ≥50 mm h −1 ) rainfall and incorporates two attention mechanisms: a channel-wise module that weights informative modalities and a spatial module that aligns features with banded structures at multi-hour leads. Training uses a tail-aware ordinal loss that couples focal reweighting with an Earth Mover’s Distance-based ordinal penalty to emphasize rare extremes. FlowsNet maintains a non-zero C4 Critical Success Index through 6 h. From 4 to 6 h, it matches or exceeds the Japan Meteorological Agency’s very-short-range forecast, and it outperforms a leading extrapolation-based method and state-of-the-art deep learning nowcasting models. Case studies show preserved band geometry and corridor placement at long lead over complex terrain. Ablation experiments identify satellite water vapor context and near-surface humidity as key for long-lead C4 prediction; combining satellite context with surface observations stabilizes placement and reduces false alarms. By avoiding numerical weather prediction model state and objective analyses/reanalyzes, the approach reduces latency and hardware demand, improves portability and resilience when model cycles degrade, and offers a practical route to earlier and more transferable warnings for extreme-rainfall events.