Scales of Tornado Damage Detectability using Machine Learning

The increasing availability of high-resolution satellite imagery and Uncrewed Aerial System (UAS) imagery has opened new avenues for storm damage assessment. This has been shown to be especially useful as a complement to ground surveys after high impact severe thunderstorms, including with tornadoes. However, the availability and use of satellite imagery to document and characterize damage has grown in recent years and also has proven useful as a supplement to ground surveys. To better understand the value added by satellite and UAS imagery, this study aims to compare the detectability of damage between satellite imagery with roughly 50 cm resolution and UAS imagery with 3 cm resolution using machine learning techniques.

By leveraging deep learning models, we evaluate the performance of damage detection algorithms on both datasets, focusing on the level of detectability across different land categories, such as crop fields, forests, and urban areas. The comparison highlights the advantages and limitations of each imagery type, particularly in detecting fine-scale damage features that are often critical in time-sensitive assessments. This presentation will discuss the implications of these results, providing insights into the optimal use of satellite and UAS imagery for damage detection and assessment.