Performance assessment of drone-based photogrammetry coupled with machine-learning for the estimation of hail size distributions on the ground

Hail-producing thunderstorm are extreme weather phenomena that can have devastating impacts on the environment, agriculture and infrastructure, leading to large financial losses. Remote sensing techniques such as dual-polarimetric weather radar are the state-of-the-art for observing hail over large areas, but don’t necessarily represent conditions on the ground.

To verify weather radar-based products, current ground-based observations provide valuable information but exhibit various limitations: Crowd-sourced reports provide information over widespread areas, but only indicate the size of the largest hail stones and are biased towards populated areas, while automatic hail sensors and hailpads provide hail size distribution (HSD), based only on small observational areas of <1m², implying that only a partial HSD is retrieved. Drone-based hail photogrammetry can help to close this observational gap by sampling thousands of hailstones within a hail core across large areas of hundreds of square-meters at high resolution. Combining the approaches of previous studies, images captured during a drone flight are combined into a rectified image (orthophoto), from which hailstones are detected and their sizes are estimated using machine-learning models for object detection and image segmentation.

To assess the uncertainty of the hail size distribution retrievals from machine-learning models, we set up experiments on different grass surfaces using synthetic hail objects with a well-defined ground truth in terms of size and number. The results of the experiments are compared to drone-based HSD retrievals of a real hail event, which occurred in 2022 in Locarno, Switzerland. In the experimental setup and the real event, 98% of the synthetic hail objects and 81% of hailstones were correctly detected. Overall, the estimated size is in good agreement with the real size of the synthetic hail objects and only slight underestimations of the detected objects could be found. Across all different size classes, the underestimation is around -0.75mm for both synthetic hail and hailstones. The high performance in detection and size estimation, coupled with large sampling areas allow us to estimate representative HSDs on the ground under favorable conditions, but due to its time-intensive and challenging data collection process, it is best coupled with other measurement methods. In combination, a reliable ground dataset can be created to validate radar estimates and potentially improve forecasting of hail events. Additionally, we present past experiences of using drone-based hail photogrammetry and ongoing improvements of the method to extend its application in the field under challenging conditions.