TITAN‐LSTM: A Weather Radar Nowcasting Tool

Abstract Precipitation nowcasting is essential for a wide range of meteorological and hydrological applications, and weather radar plays a central role in this task. Extrapolation algorithms such as TITAN (Thunderstorm Identification, Tracking, Analysis and Nowcasting) provide efficient storm tracking but do not directly use rainfall rate fields, while deep learning models capture storm dynamics using gridded radar images, resulting in extensive preprocessing and high computational cost. To bridge this gap, we have introduced TITAN‐LSTM, a hybrid radar nowcasting algorithm that combines TITAN with a Long Short‐Term Memory (LSTM) network. TITAN identifies and tracks rainstorm life histories and predicts their displacements, while LSTM forecasts rain area and precipitation intensity based on TITAN tracking. The algorithm was trained and validated using data from the DAEE‐SPOL weather radar from 2016 to 2019, generating 5–30 min forecasts. Compared to the TITAN standalone version, TITAN‐LSTM improved rainfall prediction skill, achieving POD > 0.70 and FAR < 0.20 at 15 min and substantially reducing rain‐area errors by decreasing RMSE between 60% and 80% when longer storm histories were used as input. Rainfall‐rate errors increased slightly with lead time, from about 3 mm/hr at 5 min to 4 mm/hr at 30 min. TITAN‐LSTM performed comparably to more complex deep learning architectures reported in the literature, which reported POD values between 66% and 90% at short lead times but with the advantage of directly using rainfall rate fields and rain area time series instead of 2D images.