System Architecture for Early Prediction of Tornadoes in U.S.
Open Access DepositedTornado prediction in the United States remains difficult because tornadoes form quickly,the atmosphere behaves in a very nonlinear way, and traditional forecasting methods often have trouble processing large radar datasets in real-time, which leads to false alarms and delayed warnings. This praxis fills in that gap by proposing a system architecture that is based on a Kalman-filtered Convolutional Bidirectional Long Short-Term Memory network with Multi- Head Attention and a Feedforward Network for predicting tornadoes early. The study employs six NOAA Level II radar-derived features—reflectivity (DBZ), radial velocity (VEL), specific differential phase (KDP), correlation coefficient (RHOHV), differential reflectivity (ZDR), and spectrum width (WIDTH)—to train and validate the model on TORNET data, while also testing it on NOAA Level II RADAR data. The experiment has about 5 million training records, 0.6 million validation records, and 0.5 million test records. It also compares the results to KNN, LightGBM, SVM, RNN, LSTM, BiLSTM, and Kalman-Conv BiLSTM with Attention baselines. Radar sequences were denoised using Kalman filtering, and focal loss, class weighting, and threshold tuning were used to address the significant class imbalance in tornado datasets. The proposed architecture increased the average recall by more than 27 percentage points, from 72.01% in previous studies to over 99% in the very-high-recall version. The balanced variant had an average F1-Score of 91.11, which is 9.37 points higher than the 81.74 average F1-Score in previous research. The study also showed that adding the Feedforward network improved recall by 27%, precision by 21%, and F1-Score by 16% compared to the old architecture. Feature- importance analysis further indicated that DBZ, VEL, KDP, RHOHV, ZDR, and WIDTH were the most informative indicators for tornado classification in this study. Overall, the results indicate that the proposed architecture is a promising machine learning approach for tornado prediction and contributes to ongoing research in radar-based tornado prediction models.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.