Barras, Hélène Christine Louise (2021). On hail in Switzerland – crowdsourcing, nowcasting and multi-day hail clusters. (Thesis). Universität Bern, Bern

Preview

Text 21barras_hcl.pdf - Thesis Available under License Creative Commons: Attribution (CC-BY 4.0). Download (27MB) | Preview

Abstract

Hail is one of the costliest atmospheric hazards in Switzerland, causing substantial damage to crops, cars, buildings, and infrastructure every year. Currently, the Swiss population is warned operationally about thunderstorms, but no information is given on specific hazards such as hail, severe wind gusts and lightning. One reason is the gap in ground-based hail observations, without which predictions could not be verified. To address these gaps, this dissertation presents a multi-approach advancement to hail prediction. Three projects explore crowdsourced hail reports, create hail nowcasting models and characterize large- and local scale atmospheric conditions of multi-day hail clusters. To close the gap in available ground-based hail observations, the first part of this dissertation uses crowdsourced hail size reports submitted via mobile application of the Swiss Federal Office for Meterorology and Climatology (MeteoSwiss). The reporting function was added in May 2015 and has collected more than 100'000 reports since. These reports are explored, filtered using an automatic plausibility filtering method based mainly on three criteria, and compared to two operational radar-based hail algorithms. The most important criterion guarantees a minimum proximity of reports to thunderstorms. Other criteria remove duplicate reports and artificial patterns and limit the time difference between the event time and the report submission time. If "no hail" reports are excluded, 53 % of reports collected until September 2020 remain after filtering. The comparison of crowdsourced hail reports with the algorithms probability of hail (POH) and maximum expected severe hail size (MESHS) indicates that some hail events were missed by the algorithms. While there is significant variability between size categories, the matched reports and radar-based algorithms correlate positively. MESHS values are typically 1.5 cm larger than the reported sizes. This study shows that the crowdsourced reports are invaluable for hail research and suggest that crowdsourcing could be applied to other atmospheric hazards. In the second part of this dissertation, extreme gradient boosted tree (XGBoost) models are developed to nowcast the occurrence and size of hail for individual thunderstorms. Statistics of environmental variables from radar, satellite, lightning, topography and numerical weather models serve as features (also called predictors) to predict the maximum POH and MESHS, in 5-minute time steps, up to 45 minutes in advance. For each lead-time, binary XGBoost models predict the occurrence of hail (POH ≥ 10 %, MESHS ≥ 2 cm) and, subsequently, linear XGBoost models predict the non-zero maximum POH and MESHS values. Additional models with a reduced number of input features assess how many features are needed to reach the same nowcast quality as models using all features. The binary XGBoost models predict the occurrence of hail better than the Lagrangian persistence for all lead-times ≥ 10 minutes. For a lead-time of 5 minutes, both predictions skills are equal. About 500{1000 features are necessary to reach a similar skill to models that used all features. Although all data sources are present in the top 100 features, radar-based features are the most important. Features indicating an intense thunderstorm activity at the most recent time step increase the probability of POH ≥ 10 % and MESHS ≥ 2cm. The Lagrangian persistence predict the POH values with a smaller standardized centered root-mean squared error than linear XGBoost models, up to a lead-time of 25 minutes. A likely reason is the smaller sample size used to train and test linear XGBoost models. This chapter demonstrated the effectiveness of machine learning in nowcasting and will serve as a base for future projects. Multi-day hail clusters cause significant damage in a short time. To increase their predictability, the third part of this dissertation explores the large- and local-scale atmospheric conditions during and up to three days before multi-day hail clusters and isolated hail days. Hail days between 2002-2019 are defined for two regions, north and south of the Alps, within 140 km of the Swiss radar network. The conditions are described using a weather type classification, reanalysis data, objectively identified fronts and atmospheric blocks. For both regions, composite atmospheric variables indicated a more stationary and meridionally amplified atmospheric flow during multi-day hail clusters. North of the Alps, blocks are more frequent over the North Sea and surface fronts are located farther from Switzerland on clustered hail days than on isolated hail days. Furthermore, clustered hail days are characterized by significantly higher convective available potential energy (CAPE) values, warmer daily maximum surface temperatures, and a higher atmospheric moisture content than isolated hail days. South of the Alps, these differences in CAPE, temperature and moisture are not as significant. However, the mean sea level pressure is significantly deeper on isolated hail days. For both regions, the Rossby waves are already more amplified three days before multi-day hail clusters, than before isolated hail days. Furthermore, prior to more than 10 % of clustered hail days, atmospheric blocks occur over Scandinavia, which is not the case for isolated hail days. This chapter shows that the temporal clustering of hail days is coupled to specific large- and local-scale flow conditions, providing an added value for short- to medium-range forecasts of hail in Switzerland. Furthermore, the conditions during multi-day hail clusters north of the Alps raise the question, whether multi-day hail clusters may occur more frequently with global warming. Altogether, this dissertation explores a way of closing the hail observation gap, creates nowcasting models for hail, which could lead to an operational hail warning system, and characterizes the atmospheric conditions during and before multi-day hail clusters and isolated hail days. The latter provides an added value for hail forecasts. This dissertation makes a further step towards warning the Swiss population of hail and preventing its damage.

Item Type:	Thesis
Dissertation Type:	Cumulative
Date of Defense:	29 June 2021
Subjects:	500 Science > 550 Earth sciences & geology 900 History > 910 Geography & travel
Institute / Center:	08 Faculty of Science > Institute of Geography
Depositing User:	Hammer Igor
Date Deposited:	03 Jun 2024 14:17
Last Modified:	03 Jun 2024 22:25
URI:	https://boristheses.unibe.ch/id/eprint/5109

Actions (login required)

View Item