A Radar-Based AI Framework for Nowcasting Severe Hailstorms

Short-term prediction of severe hailstorms remains a major challenge, especially when forecasts must preserve the location, intensity, and internal structure of rapidly evolving high-reflectivity cores. Although Numerical Weather Prediction (NWP) models have improved synoptic-scale forecasting, physics-based models are limited at short lead times by model spin-up. This makes it harder to accurately capture rapidly evolving hail-producing storms. Useful hail nowcasting also requires guidance that can represent storm scale evolution with finer spatial and temporal fidelity. In this work, we present a radar-based AI nowcasting framework for short time range forecasting of intense convection over the Colorado-Wyoming region. The model is trained using composite radar reflectivity fields and is designed to generate high-resolution forecasts out to 1 to 3 hours.  

To assess performance, the framework was evaluated on 15 severe hail events over the Colorado-Wyoming region using a combination of verification metrics. These include Probability of Detection (POD), False Alarm Ratio (FAR), Critical Success Index (CSI), Fractions Skill Score (FSS), and Structural Similarity Index Measure (SSIM), with a focus on the 40 dBZ reflectivity threshold as a proxy for hail-relevant storm intensity. The results show that the model provides high skill at shorter lead times and retains meaningful forecasts through 180 minutes. Comparisons against traditional extrapolation based nowcasting approaches and physics-based forecast models, including the High-Resolution Rapid Refresh (HRRR), indicate that the AI based framework is particularly effective at preserving storm structure, intensity, and spatial placement. Hail-producing storms are usually characterized by localized high-reflectivity maxima embedded within rapidly evolving convective morphology, so maintaining both intensity and spatial realism is essential for forecast usefulness. These findings suggest that radar-based AI driven nowcasting is a good tool for severe hail prediction.