Developing an Object‐Based Climatology of Tornadic Storms Over the United States Using Machine Learning and Radar Reanalysis

Abstract

Tornado climatologies over the contiguous United States have traditionally relied on report‐based data sets, which are subject to biases related to population density, time of day, and evolving observational practices. This study presents an object‐based machine learning framework for constructing a tornado climatology using storm cluster data derived from a radar reanalysis data set within the Multi‐Radar Multi‐Sensor framework spanning 2005–2011. A two‐stage pipeline is developed, where Stage 1 performs binary detection of tornadic storm clusters, and Stage 2 estimates severity across 2‐class (EF0–1 vs. EF2+), 3‐class (binned by pairs of EF‐scales), and 6‐class (EF0–EF5) configurations using over 500 radar‐derived and near‐storm environmental predictors. CatBoost and Random Forest algorithms were trained and evaluated, with CatBoost demonstrating superior performance and subsequently being selected for climatological applications. The Stage 1 detector achieves a tornado‐class Critical Success Index (CSI) of 0.49 and a macro‐averaged CSI of 0.67 at a calibrated threshold that prioritizes probability of detection at the cost of an elevated false‐alarm rate. Error analysis shows that many false alarms occur near verified tornadoes and in similar radar and environmental contexts, suggesting that the model captures a broader envelope of tornado‐favorable storm characteristics. The 2‐class Stage 2 model provides the most robust severity discrimination, while finer‐grained models show reduced skill due to the rarity of EF4–EF5 events. Applying the selected models to approximately 10.1 million storm clusters yields a severity‐stratified climatology that reproduces the observed seasonal cycle, primary spatial tornado corridor, and late‐afternoon diurnal peak, with elevated nighttime predictions partially compensating for known nocturnal underreporting.