Deep Learning U-Net Improves Hyperlocal Precipitation Nowcasting.

Accurate, short-term precipitation forecasts, known as nowcasting, are critical for managing floods and making real-time decisions in urban areas. Traditional numerical weather prediction models are computationally intensive and suffer from latency, making them less suitable for rapid, hyperlocal predictions. This research introduces a compact, radar-only nowcasting framework built around a multi-variable U-Net. It processes recent radar volume scans, incorporating multi-elevation reflectivity, Doppler radial velocity, and derived kinematic features like divergence and vorticity. This allows the model to learn storm evolution directly from high-frequency observations. The framework also includes a high-reflectivity attention module to improve sensitivity to convective cores and uses physics-guided attribution for interpretability. Tested on Mumbai radar data, the model demonstrates superior performance over persistence forecasts, providing rapid, high-resolution precipitation predictions up to 90 minutes ahead.