Abrupt changes in the Earth’s climate system: opportunities and challenges for changepoint detection Claudie Beaulieu Example: The Nile problem Annual discharge of the Nile River at Aswan 4500 Discharge (m3/s) 4000 3500 3000 2500 2000 1500 1000 1860 1880 1900 1920 Year Cobb (1978) 1940 1960 1980 2 Outline • Abrupt change nomenclature • Specific examples of applica8ons – Ar8ficial climate shi9s -­‐ precipita8on – Terrestrial ecosystems -­‐ land uptake of carbon dioxide – Timing of paleoclimate transi8ons • Discussion of challenges and opportuni8es 3 Types of abrupt changes Abrupt change Ar8ficial shi9 Unforced Forced e.g. due to changes in the measurement procedures Red noise or bimodality e.g. North Atlantic Oscillation Driver shi9 -­‐ response shi9 e.g. climate shift leads to an ecosystem shift Threshold e.g. increasing anthropogenic pressure pushes the system to cross a threshold 4 Driver – response relationship The change in the response is much larger than the change in the forcing (IPCC, 2007). Adapted from Andersen et al., 2008 5 Example: regime shifts in the Pacific Ocean Scheffer et al., 2001 6 Example: regime shifts in the Pacific Ocean “Red noise and regime shifts” Rudnick and Davis, 2003 Scheffer et al., 2001 7 Example: Vegetation cover over the Sahara Scheffer et al., 2001 8 Potential abrupt change under anthropogenic forcing Lenton et al., 2008 9 Example 1: Artificial shifts in precipitation Opportunity: improve data underlying assessment of climate and ecosystem change Challenge: adapting statistical techniques to specific problems Observed trends Intergovernmental Panel on Climate Change, 2007 11 Observed trends Problem with instrumental 1me series: Changes in the measurement procedures Intergovernmental Panel on Climate Change, 2007 12 Problems with observations Inhomogeneous Closer to the ’true’ precipitation 1600 Ptot (mm) 1400 1200 1000 800 600 1920 1940 1960 Year 1980 Photo courtesy of Lucie Vincent, Environment Canada 2000 13 Homogenization techniques Change point methods extended to detect ar1ficial shi:s only Restric1ve techniques: • At most one shi9, • Normality, independence, … Large number of techniques available. Choice depends on: • Network density (possible to find neighbour series?) • Metadata (available?) • Variable (temperature or precipita8on) • Sta8s8cal proper8es (normality, autocorrela8on, …) Compara1ve studies: • Temperature: Ducré-­‐Robitaille et al. (2003), Menne & Williams (2005), DeGaetano (2006), Reeves et al. (2007) • Precipita8on Artificial shifts in precipitation series Comparison study to iden1fy the most promising methods. Total annual precipita1on series were generated (100-­‐year long) Types of series generated • No shi9 • With a single shi9 in the mean • With mul8ple shi9s in the mean 1800 • Three neighbour series 1600 Ptot (mm) 1400 1200 1000 800 600 400 0 20 40 60 Position (year) 80 100 *9HC@>IJ9 Results: series with a single shift 0()* 07+, 07.' 078, 67+3 415' '*3' 012* +,-/ +,-. '()* Well-­‐posi)oned shi.: posi)on detected ± 2 years far from the real posi)on !" #" $" %" 49::!;<=>?><@9AB=C>D?=BEFG Beaulieu et al., 2008; 2009; 2010 &"" Results: series with a single shift p=prior probability of no-­‐change p=0.01 p=0.5 50 0 3 2.5 2 1.5 1 Magnitude 0.5 85 6575 55 3545 1525 Position 50 0 3 2.5 2 1.5 1 Magnitude 0.5 0 Beaulieu et al., 2010 100 Shifts (%) 100 Shifts (%) Shifts (%) 100 p=0.99 85 6575 55 3545 1525 Position 50 100 50 0 3 2.5 2 1.5 1 Magnitude 0.5 85 6575 55 3545 1525 Position Example 2: Abrupt change in the terrestrial carbon cycle Opportunity: gain better understanding of the underlying processes Challenge: distinguish different types of change Carbon cycle 19 Carbon sources and sinks 1959-2010 Sources Sinks Land use 1.2 Pg C/yr Fossil fuel 5.6 Pg C/yr Atmospheric growth rate 3 Pg C/yr Ocean and land sink 3.7 Pg C/yr Between 1959 and 2010, 45% of each year’s CO2 emissions remained in the atmosphere on average; the rest was absorbed by carbon sinks on land and in the oceans. Le Quéré et al., 2009; Ballantyne et al. 2012 20 Net land uptake Net land uptake = fossil fuel emissions -­‐ CO2 atmospheric growth rate -­‐ ocean uptake ) ( 1 /+19:;-;1+<1,=>12'$$%8 /?@+AB:CDE1+<1,=>12'$$&8 F?B4+-C1+<1,=>12'$$&8 G+<H+=1+<1,=>12'$$I8 ! $ Fossil fuel emissions 4 2 CO2 atmospheric growth rate 1970 1980 1990 Year 2000 3 !! !' !(1 !"#$ 6 0 1960 !"%$ !"&$ !""$ *+,- '$$$ Ocean uptake (Pg C/yr) ./012341567-8 ' agr (Pg C/yr) 8 2 1 0 1960 Le Quéré et al. (2007) Lovenduski et al. (2008) Rodgers et al. (2008) Wetzel et al. (2005) 1970 1980 1990 Year 2000 21 Land sink 1959-2010 Sarmiento et al. (2010) analysis suggests an increase of approximately 0.9 Pg C/yr around 1988-­‐1989 Sarmiento et al., 2010 22 Models considered Shift in the mean 60 40 40 40 20 20 20 0 0 0 −20 −20 −20 −40 20 30 40 50 Time Shift in the mean and variance 60 10 10 20 30 40 Time Linear trend −40 50 60 40 40 40 20 20 20 Y 60 Y Y Y 60 −40 0 0 0 −20 −20 −20 −40 20 30 40 50 Time Shift in the intercept and trend 60 10 40 Y Shift in the variance 60 Y Y Constant mean 20 0 −20 −40 10 20 30 Time 40 50 −40 10 20 30 Time 40 50 −40 10 20 30 40 Time Shift in the intercept 50 10 50 20 30 Time 40 We use an informa)on criterion to iden)fy the )ming of the change and to discriminate between the models with or without changes. We can also integrate covariate effects and/or autocorrela)on in the model (Beaulieu et al., 2012a). 23 Best models Mean net land uptake Atmospheric growth rate Beaulieu et al. (2012b) 24 Estimate of the shift Mean net land uptake Shi. in 1988 1.2 Pg C/yr Shi. in 1988 0.9 Pg C/yr Beaulieu et al. (2012b) 25 Regions where the productivity increased NPP differences between 1990-­‐2008 and 1982-­‐1989 a c b Only significant differences are shown (t-test, 5% critical level). 26 Example 3: Timing of major transitions in paleoclimate time series Opportunity: better understand the past Challenge: additional layer of uncertainty using proxy data Change-points in paleo-proxy records Beaulieu et al. (2012a; Rodgers et al., 2011) 28 Discussion on opportunities and challenges Opportunities • Remove ar8ficial signals from data underlying assessment of global environmental change • Detect, characterize and monitor shi9s in climate and ecosystem 8me series • Contribute to beher understanding of the mechanisms leading to environmental shi9s 30 Potential for monitoring climate For monitoring, we need to detect the shi: soon a:er it occurred. Simula1on study to inves1gate how long it takes to detect a shi: a:er it occurred. Synthe1c series • start in 1960 (lots of instrumental series of observa8ons that started around this 8me) • shi9 in 1990 (evidence for large scale-­‐shi9 around 1990 in the carbon cycle, in cloud and aerosols, intensifica8on of the hydrological cycle over land) Search for a shi: and decide whether a model with a shi: is beRer using two different decision rules: SIC( p) < SIC (n) SIC( p) = min {SIC(k), k = 2,..., n − 2} SIC( p) + c α < SIC (n) Beaulieu et al. (2012a) Challenges • Extending exis8ng techniques to be less restric8ve (e.g. autocorrela8on, distribu8on) • Tailor new methods to environmental data and global change problems • Account for addi8onal layer of uncertainty of the data sets and provide uncertainty for the 8ming of the change 33 Distinguish between different types of changes Evidence of piecewise linear and step-­‐like changes in global temperature • need sta8s8cs to iden8fy, quan8fy and understand temporal evolu8on Seidel & Lanzante (2004) 34 Detecting shifts in extremes Extremes • Heavy rainfall • Floods • Droughts • Heat waves Take shi:s into account when assessing the risk of future extremes (e.g. Villarini et al., 2009) Zhao & Chu (2010) 35 Detecting a threshold Toms and Lesperance, 2003 36 Multivariate change-point detection Beaugrand et al., 2008 37 Uncertainty IPCC, 2007 38 Communication • Choosing an appropriate sta8s8cal methodology and comparing the op8ons is challenging • Gaining a good understanding of the phenomena to do a meaningful analysis 39 Acknowledgements • Jorge L. Sarmiento et al., Atmospheric and Oceanic Sciences, Princeton University • Wolfgang Buermann, University of Leeds • Jie Chen, University of Missouri • Taha B.M.J. Ouarda (University of Quebec and Masdar MIT) and Ousmane Seidou (University of Ohawa) 40