Discussion

Reed, D.W. 2002a. Discussion on "Tall tales about tails of hydrological distributions". J. Hydrologic Engg, Am. Soc. Civil Engrs., 7, 192-193. [Papers 21176 and 21177 by V. Klemeš, J. Hydrologic Engg, Am. Soc. Civil Engrs., July 2000, 5, 227–239.]

It is inevitable that opinions divide most strongly on crucial and difficult issues. Flood estimation is important because property, health and lives can be lost if defence schemes fail to perform to their intended standard, or if communities lose sight of the residual risk of an overtopping event. Flood estimation is difficult because the key variable - e.g. the 100-year flood peak - is never measured. Exceptional floods occur but their rarities are only inferred, never measured. In this respect, flood estimation differs fundamentally from large branches of applied science in which the primary objective is to understand the relationship of measurable inputs to measurable outputs.

While there are important opportunities to exploit hydrological understanding and develop scientific method along the way, flood estimation is at root a statistical problem. Engineering practice has much to gain from systematic procedures that increase the effective use of relevant data. The author is right to highlight the importance of experience, not least in keeping sight of the uncertainties inherent in flood risk estimation. But are the frequency analysts really suggesting otherwise? As the title of their book makes plain, Hosking and Wallis (1997) recommend regional L-moment methods not single-site methods. It is therefore curious that the author targets Hosking and Wallis for criticism when so little of his material (just p. 236) is concerned with regional frequency analysis.

Reed et al. (1999) review the principles and building blocks of regional frequency analysis. They highlight the fundamental tension between the desire to base estimates on large quantities of data (to avoid undue extrapolation of a fitted frequency curve) and the wish to avoid contamination by inappropriate pooling. The concept of pooling data from hydrologically similar basins (e.g. Acreman and Wiltshire, 1989) reduces the scope for inappropriate couplings, whether of Rock Creek and the Mississippi or Turkey Brook and the Thames. The Flood Estimation Handbook (IH, 1999) implements such a system for use in the UK.

The balances to be struck between pooled and single-site analyses, and between statistical and rainfall-runoff methods of flood estimation, remain difficult to judge. The discusser agrees with the author's conclusion that illusion of knowledge can do more harm than awareness of ignorance. However, the author is wrong to lay the blame for unsound flood estimates at the door of applied researchers who develop, validate and disseminate powerful analytical and diagnostic tools. It is human nature that advisory warnings are often ignored. While researchers have a duty to validate and disseminate their methods clearly, the responsibility for ensuring good practice lies with those appointing, regulating, supervising and executing studies.

APPENDIX. REFERENCES