Verification of the linguistic uncertainty of warning uncertainty Martin Göber 1,3, Tobias Pardowitz 2,3, Thomas Kox 2,3 1 Deutscher Wetterdienst, Offenbach, Germany 2 Institut für Meteorologie Freie Universität, Berlin, Germany, 3 Hans-Ertel-Centre for Weather Research, Germany Göber, Pardowitz, Kox: Linguistic uncertainy of warning uncertainty 1
Operational praxis of probabilistic warning information at DWD 1. 7-day forecast of weather hazards compulsary use of possible, likely, very likely, non-quantitativ 2. Report on the regional warning situation (next 36 hours) a variety of terms used ( can not be excluded, might occur, are expected, ) 3. County based warnings (now, up to next hours/day) No uncertainty used apart fron spatial and temporal restrictions ( exposed areas, temporal ) Göber, Pardowitz, Kox: Linguistic uncertainy of warning uncertainty 2
Question: Imagine, the DWD states the advent of a storm for your area for the next day as possible, likely or very likely. Which of the following probability terms would you associate with this prediction? In: Kox, T., Gerhold, L. and Ulbrich, U. (2014) : Perception and use of uncertainty in severe weather warnings by emergency responders in Germany. In Atmospheric Research. Göber, Pardowitz, Kox: Linguistic uncertainy of warning uncertainty 3
19 counties within Berlin+Brandenburg: 30.000 km^2 Berlin: 900 km^2 Source: http://de.wikipedia.org/wiki/brandenburg#mediaviewer/datei:brandenburg,_administrative_divisions_-_de_-_colored.svg Göber, Pardowitz, Kox: Linguistic uncertainy of warning uncertainty 4
1. Can human forecasters estimate warning uncertainty? 2. How well is this done in verbal terms? probabilistic short range (T+36h, in 6h-intervals) forecast for warning events (gusts, thunderstorms) for Berlin (900 km2) or larger area of Berlin+Brandenburg (30.000 km2) since February 2013 3 different forecasts: human forecaster: from regional office in Potsdam 1) numerical estimate (deliberately produced for our project) 2) textual (operational text issued 4 times a day) 3) WarnMOS: Warning Model Output Statistics based on the global models GME and IFS includes latest observations Göber, Pardowitz, Kox: Linguistic uncertainy of warning uncertainty 5
Wind gusts thunderstorms N=1924 Göber, Pardowitz, Kox: Linguistic uncertainy of warning uncertainty 6
wind gusts man vs machine: reliability similar Brier Skill Score = 16 % Göber, Pardowitz, Kox: Linguistic uncertainy of warning uncertainty 7
thunderstorms man vs machine: reliability similar Brier Skill Score =6 % Göber, Pardowitz, Kox: Linguistic uncertainy of warning uncertainty 8
P(usage) P(obs given terms) Göber, Pardowitz, Kox: Linguistic uncertainy of warning uncertainty 9
P(usage) P(obs given terms) Göber, Pardowitz, Kox: Linguistic uncertainy of warning uncertainty 10
thunderstorms gusts Göber, Pardowitz, Kox: Linguistic uncertainy of warning uncertainty 11
Summary and Conclusions 1. human forecasters can estimate warning uncertainty reliably (and tend to have higher resolution than statistical estimates) 2. minimise overlap and vagueness by restricting oneself to only a few terms Göber, Pardowitz, Kox: Linguistic uncertainy of warning uncertainty 12
India warnings http://amssdelhi.gov.in/4dham/pages/gangotri.html Göber, Pardowitz, Kox: Linguistic uncertainy of warning uncertainty 13
Summary and Conclusions 1. human forecasters can estimate warning uncertainty reliably (and tend to have higher resolution than statistical estimates) 2. minimise overlap and vagueness by restricting oneself to only a few terms 3. reduce underspecificity by sharply delineate categories and specification of relationship between words and numbers Göber, Pardowitz, Kox: Linguistic uncertainy of warning uncertainty 14
USA NWS http://www.nws.noaa.gov/wsom/manual/archives/nc118411.html#z8-34 Göber, Pardowitz, Kox: Linguistic uncertainy of warning uncertainty 15
Summary and Conclusions 1. human forecasters can estimate warning uncertainty reliably (and tend to have higher resolution than statistical estimates) 2. minimise overlap and vagueness by restricting oneself to only a few terms 3. reduce underspecificity by sharply delineate categories and specification of relationship between words and numbers 4. stronger words for thunderstorms points to forecasting of risk rather than uncertainty alone Göber, Pardowitz, Kox: Linguistic uncertainy of warning uncertainty 16
UK weather warnings http://www.metoffice.gov.uk/guide/weather/warnings Göber, Pardowitz, Kox: Linguistic uncertainy of warning uncertainty 17
Summary and Conclusions 1. human forecasters can estimate warning uncertainty reliably (and tend to have higher resolution than statistical estimates) 2. minimise overlap and vagueness by restricting oneself to only a few terms 3. reduce underspecificity by sharply delineate categories and specification of relationship between words and numbers 4. stronger words for thunderstorms points to forecasting or risk rather than uncertainty alone The possibility of a potentially not too bad conference might possibly has to be accounted for! Göber, Pardowitz, Kox: Linguistic uncertainy of warning uncertainty 18