Multitask Learning for Tornado Identification Using Doppler Radar Data

Tornadoes, as highly destructive small-scale weather events, demand accurate detection and identification for effective weather decision-making. While weather radar serves as a primary tool for tornado identification, traditional radar-based tornado identification algorithms, such as Tornado detect algorithm (TDA) and Tornado Vortex Signature algorithm (TVS), are susceptible to radar noise, with limited tornado feature extraction capability leading to high rates of false alarms and low probability of detection. In response to these challenges, this study introduces an innovative multi-task learning network based on spatial-temporal information (TS-MINet) to improve tornado identification. Leveraging continuous three-frame radar Level-II data as inputs, including reflectivity, radial velocity, and spectral width, TS-MINet adopts a multi-task learning structure, simultaneously performing tornado detection and number estimation tasks to comprehensively extract tornado-related information. TS-MINet integrates channel recalibration blocks, spatial construction module, and temporal construction module, constructing a robust tornado identification model that overcomes the limitations of traditional algorithms with single-frame radar data. The introduction of channel recalibration blocks refines local representations, capturing micro-scale features crucial for accurate tornado identification. Inspired by the transformer architecture, the spatial construction module enriches global spatial dependencies by assimilating information from different spatial regions. Simultaneously, the temporal construction module captures the time-relatedness of consecutive radar frames, providing a nuanced understanding of tornado evolution. Given the limited number of tornado samples, data augmentation techniques like random rotation and cropping are implemented during model training to enhance robustness. Compared with the traditional TDA method with a Critical Success Index (CSI) of 0.15, the proposed method successfully improves the CSI to 0.54, which highlights the potentially advantages of deep learning methods in identification tasks. Even compared with the classical deep learning model UNet, which has a Probability of Detection (POD) of 0.62 and a False Alarm Rate (FAR) of 0.50, the proposed method achieves 0.75 and 0.32, respectively, and possesses more superior accuracy and robustness. The innovative TS-MINet model provides new insights and solutions for tornado detection, providing strong support for accurate prediction and timely response to future weather events.