New physics-guided U-Net model forecasts hyperlocal rain 90 mins ahead

A team of researchers (Akshay Sunil, Muhammed Rashid, and colleagues) has introduced a compact, physics-based deep learning framework for hyperlocal precipitation nowcasting. The model employs a multi-variable U-Net encoder-decoder architecture that ingests radar volume scans—including multi-elevation reflectivity, Doppler radial velocity, and a radial-velocity-gradient proxy. Instead of full wind field retrieval, it computes derived channels representing velocity magnitude, divergence, directional shear, and vorticity, capturing kinematic signatures of convergence and boundary interactions. A high-reflectivity attention module sharpens sensitivity to convective cores, and physics-guided attribution ensures the learned patterns are meteorologically meaningful.

Trained on Mumbai Doppler radar observations from May to August 2023, the model generates 12 future composite reflectivity fields at 7.5-minute intervals covering nowcasts from 10 to 90 minutes lead time. Performance metrics at 90 minutes show a Critical Success Index of 0.437 for ≥10 dBZ, 0.332 for ≥20 dBZ, and 0.193 for ≥30 dBZ thresholds. The model outperforms persistence forecasting in terms of lower RMSE and higher spatial correlation at longer lead times. A key practical advantage: after training, the entire system runs on a standard computer, delivering predictions in seconds—making it viable for real-time deployment in flood-prone urban regions like Mumbai.