Thursday, 31 August 2023
Boundary Waters (Hyatt Regency Minneapolis)
We present a new radar-derived hail product that uses a large dataset of insurance claims and radar data coupled with meteorological variables to accurately estimate hail damage using a deep neural network approach. A gridded observed damage dataset was created by normalising the reported losses from the insurance dataset to minimise the effect differences among the property types had on the sensitivity of individual properties towards hail damage. This dataset was then matched to radar observations in order to correct for horizontal advection of hail during its fall. Using Shapley Additive Explanations analysis we identified 5 key meteorological variables that coupled with radar observations achieved a critical success index of 0.92 and a coefficient of determination of 0.82 against observed damage. Comparing this hail damage estimate to a popular hail size product (MESH), we identified meteorological conditions associated with biases on MESH. Our results highlight the potential of our approach to quantify hail impact and provide insights into hail size distributions and/or mixed phase volumes in different environments.

