Climate Change impacts on hail

6^th June 2024

Summary

We understand the main atmosphere processes that cause damaging hail, and have built realistic models of storm cell development, yet our confidence in how a warming planet affects this peril remains low. This is largely due to the mixed quality of the observed hail climate, which creates uncertainty in past trends as well as limiting the effectiveness of model validation.

Here, I attempt to improve this situation by developing 33-year records (1990-2022) of hail days in larger US cities from Storm Prediction Center data. The use of cities ensures small reporting biases, and daily resolution fits the event timescale. City hail days provides a more homogeneous record of ground-truth hail.

A logistic regression model was built to predict hail days over these cities in 1990-2022. A quite standard set of predictors drawn from ERA5 weather variables was calibrated to the occurrence of city hail days, and the model reproduces the main variations across the U.S. and over the annual cycle.

The model was then applied to ERA5 daily weather from 1960 to 2023 over the whole of the contiguous U.S. and revealed the number of large hail days rising by around 0.5 to 1.0% per year over the 1960-2023 period for most parts east of the Rockies. Higher values of CAPE in a warming planet is the main cause of this increase; the lifting of the melting level has the opposite effect but is smaller magnitude.

The validity of the model across the diverse U.S. regional climates and their highly variable seasons suggests it can simulate hail in Europe too. It was found that the trend in hail over 1960-2023 was generally larger in Europe, at about 1% per year, with CAPE the main driving force again.

The rising trend in Europe fits with an emerging view here, but the widespread growing risk in the U.S. over the past six decades sits alongside varied signals in published work. The collection of studies on past hail trends in the U.S. use a very diverse mix of data and methods, and what we know for sure is that the simple average of their signals suggests U.S. hail risk has been trending up over the past 60 years.  

Climate models project increasing hail risk over both U.S. and Europe in the future, which fits with the past trends reported here.

The Stormwise CC Tool implements the rising hail risk into insurance workflows by perturbing climate baseline YLTs.

Introduction

Hail is the dominant cause of property damage from convective storms, and strongly related to the occurrence of other subperils such as tornadoes and straight-line winds. The changes to hail in a warming climate are a good guide to what happens to the full thunderstorm peril. However, there is a high level of uncertainty in how the hail subperil responds to global warming. For example, the IPCC sixth assessment report (Arias et al., 2021) states “In nearly all regions, there is low confidence in changes in hail…” and they go on to explain “… observations are often short-term or lack homogeneity, and models often do not have sufficient resolution or accurate parametrizations to adequately simulate them over climate change time scales“. Doubts about models would be smaller if they were shown to simulate reasonable hail climates, but the ground-truth is too poor to build much confidence in the models. 

This note describes the development of a new dataset of hail climate consisting of higher quality data around the bigger U.S. cities. This is then used to calibrate a model of city hail occurrence using large-scale environment variables. In turn, this model is applied to weather conditions over the wider U.S. mainland from 1960 to 2023, and trends in hail probability are examined. The model is also applied to European weather over the same period, and its trends are analysed.

 

A homogeneous record of past hail occurrence

The Storm Prediction Center (SPC) database of hail reports in the U.S. from 1955 onward is well known to contain strong trends due to increased reporting over the decades, and completely unrelated to climate. However, it is possible to use a fairly substantial subset of the SPC database to form a much more homogeneous hail climate across the U.S. over recent decades, using a quantity called city hail days.  I described city hail days when assessing the severe hail hitting the U.S. last year: Figure 3 of Allen and Tippett (2015) shows hail days over a specific area have smaller trends than the total number of reports, while larger populations ensure almost all hail days are reported in cities over recent decades. I post-process the SPC data to form historical records of city hail days from 1990 to 2022. This is done for (a) those 84 U.S. cities with populations above 200,000 and more than 150 km distant from any neighbouring larger cities (see Figure 1), and (b) for hail greater than or equal to three different threshold sizes, namely 1.25, 2.0 and 2.75 inches (about 3, 5 and 7 cm), reported within 0.25° (about 25 km) of the city centres. The northern Plains is not well sampled, though its low population densities ensure smaller losses too.

Figure 1: map of the cities, and their population, used in this study.

 

Model of hail prediction

The records of city hail days over the past three decades are used to calibrate a model of hail occurrences in the United States. Briefly, weather quantities at ERA5 (Bell et al., 2021) grid cells closest to cities are used as predictors in a logistic regression with city hail day occurrence, over the full period 1990-2022, for all 84 cities. The chosen weather variables were based on published research and model testing, leading to a model with five predictors, namely CAPE, bulk shear, the height of the melting level, the lapse rate from 800 to 400 hPa, and the vertical velocity at 500 hPa (w_500). Storm relative helicity (SRH) performed slightly better than bulk shear, but the cost of storing and processing so many extra levels of winds for SRH outweighed benefits in terms of model validation. The inclusion of w_500 is quite novel: basically, the well-designed convective mass flux scheme used in ERA5 works in concert with the large-scale dynamical wind field to remove instabilities, and w_500 proved to be a better model predictor than related diagnostics such as convective precipitation.

Figure 2 shows the number of city days with hail of 1.25 inches or bigger in the 1990-2022 calibration period for both observed and modelled, and their difference. The model captures the large-scale hail features in the U.S.: the peak in the southern Plains, and its progressive decline as we move further away from this region. The main model weakness concerns orography: it tends to under-estimate hail risk in the cities immediately downwind of steep mountains such as Denver, and a few cities in the lee of the Appalachians. Numerical weather prediction models are known to poorly resolve sharp changes for numerical stability considerations, and this effect is more pronounced nearer mountains, where they use a smoothed representation of surface elevation.

Figure 2: number of days with hail of 1.25” diameter or larger in the 1990-2022 period,

for the 84 validation cities: observed (top), modelled (middle) and their difference.

Hail trends over the U.S. in 1960-2023

The logistic regression models were applied to the ERA5 variables for the contiguous U.S. domain, to obtain the probability of a hail day at each grid cell from 1960 to 2023, for three different hail sizes. The upper plot in Figure 3 shows the annual mean number of days with hail diameter of 3 cm or larger over the 1960-2023 period. Peak activity occurs around the High Plains, though this might be shifted slightly eastwards by the blurring of orography in numerical models. For comparison, the hail climates from two published studies are included in the lower row of Figure 3. The study by Allen et al. (2015) simulates peak activity to the south and east of that shown in the new model, whereas the newer study by Battaglioli et al. (2023) is more similar, with peaks in the High Plains.

Figure 3: upper plot shows the modelled annual number of days with hail diameter of 3 cm or larger over the 1960-2023 period from this study. Lower left plot shows the annual mean number of 3-hourly periods with hail diameter of 2.5 cm or larger from the Allen et al. (2015) study while the lower right plot shows the annual number of hourly periods with hail diameter of 2 cm or larger from the Battaglioli et al. (2023) study. Besides different temporal binning, the three studies all have different spatial resolution – for example, Allen et al. simulate hail occurring in 1° grid cells – leading to the significantly different annual occurrence rates.

 

As a first step towards estimating long-term trends, a damage index was calculated at each grid-cell and every day. Using an assumption that hail damage varies with the fourth power of diameter (e.g. chapter 8 of Changnon et al., 2009), then the index was defined as the probability-weighted sum of the three different hail size diameters raised to the fourth power. The following results are not sensitive to the assumed relation between damage and hail diameter. Annual sums of damage were computed at each grid cell, from 1960 to 2023, then a simple linear regression was performed on the annual damage timeseries using year as the predictor. Finally, the gradient of the linear fit was re-expressed as the per cent change in hail damage per year. Figure 4 shows the linear trend of damage in all U.S. grid cells over the 1960-2023 period. Almost all higher-risk areas between the Rockies and the Appalachians have experienced rising hail damage over the past 60 years, mostly in the range from 0.5 to 1.25% per year.

A few studies have focused on past trends in the frequency of large hail, and a brief overview is given here.

• Allen et al. (2015) found no trend in the frequency of hail of 2.5 cm or larger in the U.S. from 1979 to 2012 (their Figure 12)
• Battaglioli et al. (2023) found weak and spatially mixed trends over the U.S. from 1950 to 2021 for 2 and 5 cm size thresholds (upper row of plots in their Figure 7)
• Tang et al. (2019) found increasing trends of around 5 to 10% per decade in the 1979-2017 period, for most areas east of the Rockies (their Figure 1) for hail of 5 cm or larger

There are significant differences in modelling methods and data used by these studies. For example (i) different hailstone size thresholds, and time periods studied, and (ii) Allen et al. use monthly mean weather predictors whereas Battaglioli et al. use hourly, and both Tang et al. and this study use daily maxima, and (iii) Allen et al. count numbers of 3-hourly reports, Battaglioli et al. count hourly values, while both Tang et al. and this study count hail days. Measuring the impacts of these different choices is beyond the scope of this short note. Instead, it is observed how two studies detect little change and two report quite significant upward trends, hence the balance suggests U.S. hail risk has been rising over recent decades. This is consistent with many climate model projections for the future, of increased hail risk (e.g. Trapp et al., 2019 and references therein).

 

Figure 4: linear trend in hail damage, in per cent per year, from 1960 to 2023.

The relative contributions of the five predictors to the grid-cell trends in Figure 4 were estimated, and as expected, CAPE was the main driver of the rising hail risk. Figure 5 shows the trend in hail risk caused by CAPE, and it is highly consistent with the full hail trends in Figure 4. The increases due to CAPE were ameliorated by reductions in risk due to the rising melting level (not shown), but CAPE is clearly the dominant change over the past 60 years. 

 

Figure 5: the per cent change in hail risk per year due to CAPE, in the 1960-2023 study period.

Hail trends over Europe in 1960-2023

The model is valid for most of the U.S. with its wide variety of climatic regions. Further, it is equally valid for all four seasons of the year (not shown). Therefore, its simulation of the European hail climate should be reasonable. Figure 6 shows the annual trend in hail risk (per cent per year) over the same 1960-2023 period for Europe. The model predicts hail risk is rising faster than over the U.S. (about 1% versus 0.7% per year) and once again, increases to CAPE are dominant. This is entirely consistent with the analysis of past trends in Battaglioli et al. (2023), and future trends analysed in Radler et al. (2019). Indeed, there is an emerging consensus on more severe hail damage in Europe as a result of a warming climate, as described by a group of international experts in Raupach et al. (2021), and the result here merely adds to the growing body of evidence.

Figure 6: as Figure 4, for Europe.

In summary, the new hail occurrence model strongly suggests hail risk has been on the rise over the past few decades in both the U.S. (about 0.7% per year) and Europe (around 1% per year). The upward direction in Europe is consistent with other analyses of past trends here, while the significant uplift of around 0.7% per year detected across the U.S. over the past 60 years is consistent with one recent study, but different from two others finding little net change. Climate models project increasing hail risk over both U.S. and Europe in the future, and the past trends reported here, based on a more homogeneous observed record, support this view.