ISSN 1000-0526
CN 11-2282/P
CN 11-2282/P
TFM+CNN Based Tornado Identification Algorithm Driven by Radar Feature Parameters of Storm
- Article
- Figures
- Metrics
- Preview PDF
- Reference
- Related
- Cited by
- Materials
Abstract:
In response to the issue of insufficient accuracy in identifying tornadoes using conventional threshold methods,this study proposes a machine learning-based tornado identification method driven by multiple radar-derived storm features. The method first calculates and constructs a storm feature dataset based on radar base data from tornado cases in China between 2003 and 2023. The dataset includes three categories: (1) tornadic storms, (2) non-tornadic storms with significant three-dimensional vortex signature(3DVS), and (3) non-tornadic storms without significant 3DVS. Feature importance ranking is then performed using XGBoost, and the top 14 features are selected as model inputs after combining with PCA. A TFM+CNN deep learning model for tornado recognition is constructed by using the encoder part of the TFM architecture and utilizing CNN to increase the input vector dimension and adding multiple fully connected layers at the end of the model. Three binary classification experiments are conducted using this model: (1) tornado storms vs. non-tornadic storms without significant 3DVS, (2) tornado storms vs. non-tornadic storms with significant 3DVS, and (3) tornado storms vs. non-tornadic storms. Comparative experiments with the XGBoost model are also performed. The results show that the TFM+CNN model achieves CSI values of 84%, 71%, and 73%; POD values of 94%, 83%, and 83%; and FAR values of 11%, 16%, and 14% in the three binary classification experiments, respectively. Compared to the XGBoost model, the TFM+CNN model performs better in most metrics, except for a 5% higher FAR in experiments 1 and 3. Both models exhibit the worst performance in experiment 2. Both models performed worst in experiment 2, indicating that tornadic storms and non-tornadic storms with significant 3DVS are the most challenging to distinguish. Furthermore, the TFM+CNN model has a larger AUC under ROC, indicating stronger generalization ability. It can be seen that the TFM+CNN model has strong tornado recognition ability.
Keywords:
Project Supported:
Clc Number:
Mobile website
