Abstract
Tornadoes are highly destructive weather events, and their timely detection is crucial for mitigating damage and saving lives. Doppler weather radar serves as the primary operational tool for tornado detection. However, existing detection methods mainly struggle with high false alarms. In addition, the potential mismatch between tornado reports and the corresponding radar signatures further degrade the detection accuracy. To tackle these issues, we propose TorDet, a novel two-stage deep learning approach designed for tornado detection using Doppler weather radar data. TorDet introduces a quantitative tornado labeling scheme that integrates the prominence of tornado-related radar features and tornado reports, explicitly categorizing samples (TOR, WRN, WEK, and NUL) to reduce ambiguity in both manual labeling and model training. TorDet’s two-stage design divides the detection process into distinct yet complementary tasks. In the Detection stage, it generates low-resolution feature maps to identify potential tornado-related regions by capturing large-scale information. The Positioning stage refines these detections, accurately locating tornado centers and classifying their categories. To better exploit limited data, we introduce deep supervision at intermediate layers in both stages, improving performance by enhancing the loss function and enriching gradient for backpropagation. We conduct experiments using radar data collected between 2017 and 2024. Experimental results demonstrate that TorDet, benefiting from its unique design and training strategy, reduces false alarms and maintains high correct detections compared with baseline deep learning models. On the test set, it improves the Probability of Detection (POD) from 0.511 to 0.647, reduces the False Alarm Ratio (FAR) from 0.795 to 0.697, and increases the Critical Success Index (CSI) from 0.171 to 0.260. Additional evaluation on an open-source dataset further supports its robustness and generalization capability across diverse radar data sources.
Showing the abstract — retrieve the full paper via the Exa API.