An extreme value method to study decadal hurricane wind trends

TL;DR

This paper introduces a novel extreme value theory-based method for estimating smooth wind-speed percentiles, enabling robust analysis of decadal hurricane wind trends with minimal distribution assumptions.

Contribution

It develops a percentile-smoothing technique for wind extremes using satellite and model data, addressing data scarcity and dependency issues for trend analysis.

Findings

Robust estimates of wind extremes down to 99.9999th percentile.

Consistency across different distribution fits and data sources.

Applicability to various scatterometer and reanalysis datasets.

Abstract

This paper presents a method developed using techniques from extreme value theory to estimate smooth wind-speed percentiles, allowing us to consider more extreme wind speeds while being less sensitive to the noise that stems from the scarcity of extreme data. A reliable characterisation of wind extremes is the first required step for studying decadal trends in tropical-cyclone and extra-tropical-cyclone winds. We develop a percentile-smoothing method using ASCAT-A Level-3 products, focusing on a number of tropical basins (Caribbean and Atlantic), estimate the uncertainty with the block-bootstrap technique to address the issue of dependency, and apply our method to both scatterometer winds (ASCAT-A at two different resolutions) and collocated ERA5 model data. The results obtained are very robust at basin level, without having to rely on a strong assumption for the distribution tail: they are very consistent whether we use exponential fits, generalised-Pareto fits, or even no fit at all, down to at least truly extreme wind percentiles such as 99.999th (main result), and remain quite consistent within uncertainties down to 99.9999th. As ensuring scientifically sound decadal-trend conclusions would require going back sufficiently in time, spanning the lifetimes of different instruments with different characteristics and extreme-wind statistics, a natural follow-on study would be to apply this method not only to ASCAT, but also to its predecessors on QuikSCAT and ERS - comparing each scatterometer individually against ERA5, and also to each other as partial overlaps exist between instruments.