Climate impacts and future trends of hailstorms in China based on millennial records

Abstract

Understanding how hailstorm trends have changed in the context of climate change is a persistent challenge, mainly because of the lack of long-term consistent observations of hailstorms. Here, we leverage hail damage records from Chinese historical books and extend hailstorm records to approximately 2890 years ago, exploring variations in the number of hailstorm days between 1500 and 1949 based on reliable and consistent data. We show that the number of hailstorm days was constant before 1850, but has increased significantly afterwards. This increase in hailstorm days seems to be associated with the increase in surface temperature after the population effect is removed. In addition to the trend, hailstorm activity is found to display both quasicentennial and multidecadal variability, with the former (later) dominating before (after) the 1850s, driven by the Pacific Decadal Oscillation (PDO). These results suggest that long-term changes in hailstorm days in China are modulated by climate warming and natural variability, via the PDO. Future projections based on different climate change scenarios and a convolutional neural network model show a further increase in the number of hailstorm days in the 21^st century.

Fig. 1. Long-term records of hailstorm days in China.

a Ancient hail damage records from 886 BC to 1100 (grey bars), with key historical markers at 886 BC and the Han dynasty indicated by pink triangle and blue star, respectively. (b) Ancient damage records from 1100−1948 (grey bars); the black solid line shows the 20 year smoothed time series. The beginning of the Ming dynasty is marked by a yellow diamond. c Hail damage data from 1949−2000 (grey shading); colored lines represent CMA (China Meteorological Administration) station records for hailstorms with maximum hailstone sizes ≥2 mm (purple), ≥5 mm (yellow), ≥7 mm (orange), and ≥10 mm (blue). The grey solid line shows the 20 year smoothed hail damage data.

Fig. 2. Long-term variations in hailstorm damage days, population, and temperature from 1500 to 1948.

a Observed hailstorm damage days (light grey greysolid line) and population-detrended hailstorm damage days (dark grey solid line). The brown line represents the nonlinear trend component of population-detrended hailstorm damage days decomposed via the CEEMDAN method. The orange line shows the nonlinear trend, quasicentennial variability (QCV), and multi-decadal variability (MDV) components. b Time series of the population (blue dotted line), temperature anomaly (light green line), and the nonlinear temperature trend component (dark green line), which is also derived from the CEEMDAN method.

Fig. 3. Variability in quasicentennial and multidecadal components of hailstorm day and climate indices.

Solid lines represent the quasicentennial variability (QCV) components, and dashed lines represent the multidecadal variability (MDV) components. a Hailstorm day variability, b Pacific Decadal Oscillation (PDO) variability, c North Atlantic Oscillation (NAO) variability.

Fig. 4. Historical and projected trends of hailstorm days in relation to temperature anomalies and the PDO.

Annual hailstorm days are shown as light grey lines, with 50 year smoothed trends in dark grey for historical (pre-1950, left Y-axis) and current (1950−2000, right Y-axis) periods. Colored solid lines from 1960−2000 represent projections (post-2000, right Y-axis) based on observed temperature anomalies and the Pacific Decadal Oscillation (PDO) index. Colored solid lines from 2000−2100 show future projections under RCP scenarios: RCP-8.5 (red), RCP-6.0 (orange), RCP-4.5 (light blue), and RCP-2.6 (dark blue). Shading around projections indicates ±1 standard deviation from 200 ensemble training runs. PDO index is shown as shading, with data from 1870 to 2013 based on HadISST1.1, and from 2014 onward derived from CMIP5 models. Temperature anomalies are shown as dashed green lines (10 year low-pass filtered), with different shades representing the four RCP scenarios. Data from 1880 and 2000 are from NASA/GISS, and from 2000 to 2100 are from the CMIP5 multimodel ensemble mean.