Enhancing Orographic Precipitation Nowcasting over the Santa Clara Valley Using Deep Learning

Accurate nowcasting (0–6 hours) of precipitation, particularly during high-impact weather events such as landfalling hurricanes, is vital for disaster preparedness and mitigation. Traditional methods such as radar reflectivity extrapolation techniques often struggle with predicting rapidly evolving precipitation patterns during severe weather events. The predictability of these methods diminishes quickly within the first hour, due to the complex, non-linear dynamics of atmospheric processes, making them less reliable for accurate warning operations. To address these limitations, recent advances in deep learning have shown promise in improving nowcasting. Leveraging models such as convolutional recurrent neural networks (ConvRNNs) offers a more sophisticated approach by better modeling the spatiotemporal variability of precipitation. This study introduces a novel deep learning architecture for precipitation nowcasting and extracting the global spatiotemporal dependencies inherent in severe weather systems.

The study focuses on the Western United States, with particular emphasis over the Santa Clara Valley — a region known for its vulnerability to atmospheric rivers (ARs). Santa Clara's unique geographical features, such as its proximity to the Pacific Ocean and complex topography, often pose challenges to orographic precipitation nowcasting. To address these challenges, we utilized a comprehensive dataset of precipitation measurements spanning from 2017 to 2023. This dataset encompasses a wide range of meteorological conditions, including severe storms, prolonged rainfall periods, and instances of both localized and widespread flooding. By incorporating such a diverse array of high-impact weather events, the dataset provides a robust foundation for training the deep learning model. This diversity allows the model to generalize effectively across different precipitation scenarios, ensuring its applicability not only in Santa Clara but also in other regions with similar meteorological challenges. The nowcasting model performance was evaluated using several skill scores, including the Heidke Skill Score (HSS), Critical Success Index (CSI), Probability of Detection (POD), and False Alarm Rate (FAR). The results demonstrate that the deep learning model significantly outperforms traditional nowcasting methods, offering more accurate and reliable forecasts that are crucial for disaster response and emergency management.