Reproduce “A Benchmark Dataset for Tornado Detection and Prediction using Full-Resolution Polarimetric Weather Radar Data” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mark S. Veillette, James M. Kurdzo, Phillip M. Stepanian, John Y. N. Cho, Siddharth Samsi, Joseph McDonald
A new benchmark dataset called TorNet is introduced, providing full-resolution polarimetric weather radar data to advance machine learning (ML) algorithms for tornado detection and prediction. The dataset enables deep learning models, particularly Convolutional Neural Networks (CNNs), to significantly outperform the operational Tornado Vortex Signature (TVS) algorithm in identifying tornadic events.
Key insights
- Directly processing full-resolution polarimetric radar imagery with deep learning, especially with CoordConv layers to account for polar coordinate characteristics, provides superior performance over traditional feature engineering.
- Including 'difficult' non-tornadic cases (tornado warnings without confirmed tornadoes) in the dataset is crucial for training models that can effectively reduce false alarms.
- While CNNs offer the highest potential, their training process is more sensitive to initialization and hyperparameters compared to traditional ML models, indicating a need for robust training strategies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.