HELCOM workshop, Uppsala 18 May 2015 An uncertainty framework for marine indicators based on monitoring data - Variations in time, space and methodology Jacob Carstensen Department of Bioscience Aarhus University www.waters.gu.se
Decision-making is inherently uncertain! www.waters.gu.se
Why bother about uncertainty? WFD classification and uncertainty • Benefit-of-doubt (polluters option) • Face-value (sharing option) • Fail-safe (environmental option) CIS guideline #7 www.waters.gu.se
How can we determine the confidence in status classification? www.waters.gu.se
How can we determine the confidence in status classification? www.waters.gu.se
The simple and ” convenient ” approach 10 Station 3 m depth 8 Station 7 m depth Chlorophyll a (µg L -1 ) 6 4 2 0 2007 2008 2009 2010 2011 2012 2013 Spreadsheet solution Mean = 1.61 µg L -1 Mean = 1.52 µg L -1 St.dev. = 1.56 µg L -1 St.dev. = 1.35 µg L -1 www.waters.gu.se
The simple and ” convenient ” approach 10 Station 3 m depth 8 Station 7 m depth Chlorophyll a (µg L -1 ) 6 4 2 0 2007 2008 2009 2010 2011 2012 2013 Spreadsheet solution Mean = 1.61 µg L -1 Mean = 1.52 µg L -1 St.dev. = 1.56 µg L -1 St.dev. = 1.35 µg L -1 Mean = 1.56 µg L -1 St.dev. = 1.46 µg L -1 www.waters.gu.se
The simple and ” convenient ” approach 10 Station 3 m depth 8 Station 7 m depth Chlorophyll a (µg L -1 ) 6 4 2 0 2007 2008 2009 2010 2011 2012 2013 Spreadsheet solution Mean = 1.61 µg L -1 Mean = 1.52 µg L -1 St.dev. = 1.56 µg L -1 St.dev. = 1.35 µg L -1 Mean = 1.56 µg L -1 St.dev. = 1.46 µg L -1 www.waters.gu.se
www.waters.gu.se
The uncertainties to be considered Fixed Random www.waters.gu.se
Including fixed factors to reduce random variation 70% 20% 10% www.waters.gu.se
Including fixed factors to reduce random variation 95% 70% 20% 10% 4.5% 0.5% www.waters.gu.se
Reducing uncertainty by including seasonal variation 10 Chlorophyll a (µg L -1 ) 1 Rel. Uncertainty: 108% and 113% Station 3 m depth Station 7 m depth 0.1 2007 2008 2009 2010 2011 2012 2013 www.waters.gu.se
Reducing uncertainty by including seasonal variation 10 Chlorophyll a (µg L -1 ) 1 Rel. Uncertainty: 108% and 113% Station 3 m depth Station 7 m depth 0.1 2007 2008 2009 2010 2011 2012 2013 2.5 Rel. Uncertainty (season): 2 Chlorophyll a (µg L -1 ) 101% and 99% 1.5 1 Station 3 m depth 0.5 Station 7 m depth 0 www.waters.gu.se Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
Reducing uncertainty by including interannual variation 10 Chlorophyll a (µg L -1 ) 1 Rel. Uncertainty: 108% and 113% Station 3 m depth Station 7 m depth 0.1 2007 2008 2009 2010 2011 2012 2013 2.5 2 Rel. Uncertainty (season) 2 Chlorophyll a (µg L -1 ) Chlorophyll a (µg L -1 ) 1.5 101% and 99% 1.5 1 1 Rel. Uncertainty (season & year): 0.5 Station 3 m depth Station 3 m depth 0.5 93% and 95% Station 7 m depth Station 7 m depth 0 0 www.waters.gu.se Jan 2007 Feb Mar 2008 Apr May 2009 Jun Jul 2010 Aug Sep 2011 Oct Nov 2012 Dec
Splitting the residual variation between random components 10 Chlorophyll a (µg L -1 ) 1 Rel. Uncertainty (season & year): Station 3 m depth 93% and 95% Station 7 m depth 0.1 2007 2008 2009 2010 2011 2012 2013 Component VAR SE Rel. Uncertainty Station 0 0 0% Station × month 0 0 0% 0 0 0% Station × year year × month 0.1472 0.3837 47% Residual 0.2730 0.5225 69% www.waters.gu.se
Chlorophyll a uncertainty components Component VAR SE Rel. Unc. Station 0.1212 0.3481 42% Station × month 0.0249 0.1578 17% Station × year 0.0321 0.1791 20% year 0.0160 0.1265 13% year × month 0.1255 0.3543 43% Residual 0.4113 0.6414 90% Total variation 0.7310 0.8550 135% To reduce uncertainty it is important to focus on the largest sources www.waters.gu.se
Another example: Eelgrass shoot density www.waters.gu.se
Another example: Eelgrass shoot density Area Site # # # divers # observations stations years Vitsandsbrygga 1 2 2 12 North Sound Höganäs 2 12 2 144 Landskrona 2 17 4 288 Central Sound Bjärred 4 16 3 288 Lomma 3 13 3 168 Limhamn 2 4 2 120 Bunkeflo 3 11 3 204 South Sound Klagshamn 5 17 4 450 Bredgrund 3 11 2 167 Ö. Haken 1 1 1 6 Lilla Hammar 1 1 1 6 Segelskär 1 1 1 6 South coast Fredshög 1 11 3 60 www.waters.gu.se
Model 1: Analysis of individual sites Eelgrass shoot density V[G × Y] Area Site V[GRADIENT] V[YEAR] V[PERSON V[PATCHINESS] ] Vitsandsbrygga - 0.1102 - - 0.0511 North Höganäs 0.2119 0.0164 - 0.0027 0.1761 Sound Landskrona 0.0262 0.0082 0.0697 0.0082 0.2275 Central Bjärred 0.0520 0.0292 0.0718 0.0111 0.1370 Sound Lomma 0.0777 0.0067 0.0123 - 0.1432 Limhamn - 0.0177 0.0167 0.0121 0.4577 Bunkeflo 0.5013 - 0.0987 0.0508 0.1132 South Klagshamn 0.1142 0.0368 0.0137 - 0.3550 Sound Bredgrund 0.0734 0.0388 0.1986 0.0117 0.0391 Ö. Haken - - - - 0.0646 Lilla Hammar - - - - 0.0089 Segelskär - - - - 0.0568 South coast Fredshög - 0.0648 - 0.0012 0.0917 www.waters.gu.se
Model 2: all sites combined 2000 A) Shoot density (m -2 ) 1500 1000 Component VAR SE Rel. Unc. 500 YEAR 0.0098 0.0990 10% STATION 0.1273 0.3570 43% 0 YEAR × STATION 0.0133 0.1153 12% PERSON 0.0599 0.2447 28% 1200 B) PATCHINESS 0.2151 0.4638 59% 1000 Shoot density (m -2 ) 800 Total variation 0.4254 0.6522 92% 600 400 200 0 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 www.waters.gu.se
Model 3: Explaining the gradient by depth 2000 A) Shoot density (m -2 ) 1500 1000 Component VAR SE Rel. Unc. 500 YEAR 0.0063 0.0794 8% STATION 0.0181 0.1345 14% 0 YEAR × STATION 0.0222 0.1490 16% PERSON 0.0768 0.2771 32% 1200 B) PATCHINESS 0.0944 0.3072 36% 1000 Shoot density (m -2 ) 800 Total variation 0.2178 0.4667 59% 600 400 200 0 www.waters.gu.se 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
Estimating indicator variance – crossed design Estimation of indicator variance – crossed design Number of years Number of replicates Number of stations www.waters.gu.se
Uncertainties at the indicator level propagate through the integrated assessment www.waters.gu.se
Uncertainties at the indicator level propagate through the integrated assessment www.waters.gu.se
Uncertainties at the indicator level propagate through the integrated assessment www.waters.gu.se
Conclusions • BQE indicators are influenced by many different sources of uncertainty that must be considered • Estimation of variance components with a reasonable precision requires a large dataset with a structure that allows for identification of these components • The uncertainty associated with the various variance components can be reduced by including fixed, explanatory factors in the model • Uncertainties from indicators propagate through the integrated assessment and this allows for estimating the confidence in classification and the risk of misclassification www.waters.gu.se
So, how do I calculate uncertainties in my spreadsheet? www.waters.gu.se
So, how do I calculate uncertainties in my spreadsheet? YOU DON ’ T www.waters.gu.se
Uncertainty www.waters.gu.se
Uncertainty Uncertainty www.waters.gu.se
Uncertainty Uncertainty www.waters.gu.se
Recommend
More recommend
