Delivering local-scale climate scenarios for impact assessments. Mikhail A. Semenov Rothamsted Research BBSRC, UK

Transcription

1 Delivering local-scale climate scenarios for impact assessments Mikhail A. Semenov Rothamsted Research BBSRC, UK

2 Rothamsted Research Sir Ronald Fisher Founded in 1843 by John Lawes. Fisher ( ) developed statistical methods such as design of experiments, analysis of variance, maximum likelihood, Fisher's linear discriminator and Fisher information. Rothamsted considered to be the most important birthplace of modern statistical theory and practice.

3 The Millionaire" Calculating Machine

4 Outline Nonlinearity of process-based impact models Downscaling of global climate models Weather generators ELPIS: a dataset of local-scale climate scenarios Ensembles and uncertainty Impact on wheat

5 How to build a process-based wheat model Photosynthesis Biomass Biomass = RUE(T,W,CO2,...)*R R intercepted radiation Light intercepted Light Int tersepted Beer's Law P = 1-exp(-k LAI) P proportion of light intercepted Leaf Area Index LAI Leaf Area Index anthesis max LAI Thermal time Emergence Anthesis Maturity

6 Nonlinear responses environmental variations Rate of photosynthesis as a function of temperature Light intensity

7 High-impact short-duration extreme events Relative reduction of grain numbers β Tcr Heat shock at flowering can decrease the grain number (Mitchell et al, 1995; Wheeler et al 1996) Heat shock shortly after flowering can decrease the potential grain size (Wardlaw et al 1989) Temperature at anthesis, C

8 Requirements for climate scenarios Should be daily site-specific and be able to reproduce climatic variability and extremes Should include the full set of climate variables required by impacts models Should contain an adequate number of years to permit risk analysis Should include predicted changes in mean climate and climate variability

9 Effect of changes in climatic variability on yield Yield CV ~ ~0.1 0 base no Var Var base no variability with variability (Semenov & Barrow, 1997)

10 Predicting the evolution of climate Coarse spatial resolution of GCM means that sub-grid scale processes, such as precipitation, are not adequately resolved

11 GCM output is biased Absolute bias in average surface temperature from ERA- 40 (averaged across all IPCC AR4 models) (Knutti et all, 2010)

12 Downscaling techniques Statistical downscaling Derives statistical relationships between observed small-scale variables and larger (comparable to GCM) scale variables, using various techniques (circulation typing, regression analysis, or neural network) Dynamic downscaling A regional climate model is nested within GCM and run over a region using boundary conditions from GCM Weather generator

13 Biases of RCMs from the PRUDENCE ensemble Temperature Precipitation (Jacobs et al, 2007)

14 Stochastic weather generator Weather generator is a stochastic numerical model to generate daily weather series statistically similar to observed. It can be used: to generate long weather time-series suitable for risk assessment including extreme events to provide the means of extending the simulation of weather to unobserved locations to serve as a computationally inexpensive downscaling tool to produce local-scale climate scenarios

15 Single-site weather generators 0.12 Parametric (WGEN, Richardson & Write, 1984) small set of parameters, prior assumptions on distributions (mean, sd) Semi-parametric (LARS-WG, Semenov et al 1998) small set of parameters, no prior assumptions on distributions Bootstrap resampling Non-parametric(Resampling, Lall& Sharma,1996 ) Original dataset used to resample time-series

16 Skills of WGs in diverse climates Statistical tests Kolmogorov-Smirnov test for probability distributions T-test for means and F-test for variances Significant test results Site Series Rain Month mean Month var Athens Bismarck Boise Bologna Caribou Debrecen Indianapolis Jokioinen Mobile Montpellier Moscow Munich Rothamsted Seville Tashkent Tucson Verhojansk Wageningen (Semenov et al, 1998)

17 Interannual variability & WGs WGs often fail to capture interannual variability Monthly temperature Daily temperature Generated 4 2 Generated Observed -3 Observed

18 Extreme events Prob bability nor m Extreme value theory The Three Types Theorem states the distribution for maxima converges to one of the three types: Gumbel, or Fréchet, or Weibull Generalized Extreme Value (GEV) distribution Precipitation, mm ( ; µ, σ, ξ ) G x 1/ ξ ( x µ ) exp 1 + ξ, ξ 0 σ + = ( x µ ) exp exp, ξ = 0 σ

19 Simulation of extreme events by LARS-WG Precipitation Maximum temperature Heat waves Precipitation Maximum temperature Heat-waves WG 20 years return values, mm Obs 20 years return values, mm WG 20 years return values, C Obs 20 years return values, C values, day WG 20 years return Obs 20 years return values, day

20 Spatial interpolation Interpolation LARS-WG for UKCIP02 at 5-km grid Local interpolation Global correction Interpolated baseline parameters 138 sites with daily weather 138sites with daily weather monthly means of 2376interpolated with thin plate smoothing splines (Semenov & Brooks 1999; Hutchinson 1995; Semenov 2007)

21 Multi-site weather generators A parametric hidden Markov model for precipitation (HMM, Hughes & Guttorp1994) The underlying hypothesis of an unobserved discrete weather state that corresponds to distinguished spatial rainfall distribution patterns over the ground. A multi-site stochastic precipitation model (Wilks 1998; Wilks 2009) Extended a single site WG of rainfall to multisite by driving a collection of single-site models with serially independent, but spatially correlated random numbers.

22 Multi-site weather generators A resampling K-nearest neighbour (KNN) model (Lall& Sharma 1996; Buishand) Sampling of rainfall at multiple locations, from the historical rainfall record. This involves identifying nearest neighbours in the historical record that have similar characteristics as the previous day, and using observations for the current day as the basis for resampling.

23 WGs as a downscaling tool for local-scale climate scenarios Baseline: estimate WG parameters of the distributions for climatic variables for a site using observed daily weather or pre-calculated dataset Derived from GCM/RCM -changes for climatic means and variability between baseline and future climate predictions Adjust WG distributions for a site with -changes of climate predictions Using adjusted WG parameters for a site generate localscale climate scenarios

24 ELPIS: a dataset of local-scale climate scenarios for Europe Underlying observed daily weather from the European Crop Growth Monitoring System (CGMS) meteorological dataset from the EC Joint Research Centre Climate predictions from multi-model ensemble of GCMs from the IPCC AR4 ensemble LARS-WG stochastic weather generator (Semenov et al, 2010)

25 CGMS meteorological dataset Daily data for precipitation, temperature and radiation from ~3000 stations for have been interpolated to a regular 25-km grid over Europe Interpolated daily data represents daily weather at a typical site from the grid used for agricultural production In conjunction with agricultural models CGMS is used by JRC for various agricultural assessments for the EC.

26 Multi-model ensemble of GCMs from IPCC AR4 Projections from 15 GCMs were used including HADCM3, ECHAM5, CNRM and GFDL Most of GSMs were run for A1B and B1 emission scenarios (some for A2) Predictions were available for , and GCM resolution vary from to 4 5. Monthly mean data were available for the whole globe

27 LARS-WG stochastic weather generator Generates precipitation, min and max temperature, radiation and potential evapotraspiration Modelling of precipitation event is based on wet/dry series Semi-empirical distributions are used for distribution of climatic variables Temperature and radiation are conditioned on the wet/dry status of a day and auto- and cross-correlated LARS-WG is used in more than 65 countries for research and in several Universities as an educational tool LARS-WG is available for academic, governmental and non-profit organizations from

28 Performance of LARS-WG for CGMS dataset Proportion of CGMS grids with the number of significant test results at the significance level α = 0.01

29 Test results for mean temperature

30 Test results for precipitation distributions and means

31 ELPIS: open architecture Ensembles of climate predictions from GCM/RCMs Datasets of baseline parameters for regions LARS-WG Local-scale climate scenarios

32 Ensembles and uncertainty Multi-model ensembles of GCMs or RCMs AR4 or ENSEMBLES, Centres around the world, ~20 Structural model uncertainty Perturbed physics ensembles HADAM3/SM3, UK Met. Office, ~1,000 Uncertainty to model parameterization Climateprediction.net HADSM3, ~100,000 a distributed project for climate predictions up to 2100

33 Use of ensemble predictions For small ensembles all members should be used for impact assessment models are equally probable, or weighted according to performance metrics (Gleckler et al, 2008) For large ensembles Sample a sub-set of projections Use of a statistical emulator for probabilistic sampling (Murphy et al, 2007; UKCP09)

34 Mean model outperforms individual GCMs (Knutti et al, 2010)

35 Impact on wheat: identifying future threats Maximum temperature and precipitation at SL, Spain and RR, UK for 2050 as predicted by IPCC AR4 MME and for Maximum te emperature, C RR SL Monthly precipitation, mm Month Month

36 Yield losses from drought stress decreased Soil water deficit at anthesis and 95-perccentiles of DSI (drought stress index) for baseline and for 2050 (A1B) 100 r deficit, mm Soil water percentile of DSI TR WS WA RR MA DC CF MO SL Cultivar Avalon Maturity 8 Aug 18 July (Semenov & Shewry, submit)

37 Probability of heat stress at flowering increased 1.0 Probability Probability consecutively Probability of maximum temperature exceeding 27 C within 3 days of anthesis or at and after anthesis for baseline and for 2050 (A1B) TR WS WA RR MA DC CF MO SL Cultivar Mercia Flowering 19 June 5 June Tmax at flowering, C

38 Publications Gleckler PJ, Taylor KE, Doutriaux C(2008) Performance metrics for climate models. J. Geophysical Research-Atmospheres 113 Hutchinson MF(1995) Interpolating mean rainfall using thin plate smoothing splines. Int. J. Geog. Infor. Systems 9: Knutti R, Furrer R, Tebaldi C, Cermak J, Meehl GA(2010) Challenges in Combining Projections from Multiple Climate Models. J. Climate 23: Pierce DW, Barnett TP, Santer BD, Gleckler PJ(2009) Selecting global climate models for regional climate change studies. PNAS 106: Semenov, M.A. & Barrow, E.M., Use of a stochastic weather generator in the development of climate change scenarios. Climatic Change, 35: Semenov M.A., Brooks R.J., Barrow E.M. & Richardson C.W (1998) Comparison of the WGEN and LARS-WG stochastic weather generators in diverse climates. Climate Research 10: Semenov M.A.& Brooks R.J.(1999) Spatial interpolation of the LARS-WG stochastic weather generator in Great Britain. Climate Research 11: Semenov MA 2007 Development of high resolution UKCIP02-based climate change scenarios in the UK, Agric. Forest Meteorology, 144: Semenov MA(2008) Simulation of weather extreme events by a stochastic weather generator, Climate Research 35: Semenov MA& Stratonovitch P(2010) The use of multi-model ensembles from global climate models for impact assessments of climate change. Climate Research 41:1-14 Semenov MA, Donatelli M, Stratonovitch P, Chatzidaki E, and Baruth B. (2010) ELPIS: a dataset of local-scale daily climate scenarios for Europe. Climate Research(DOI: /cr00865) Semenov MA & Shewry PR Modelling predicts that heat stress, not drought, will increase vulnerability of wheat in Europe Proceedings of the Royal Society B,(submitted) Wilks DS(2009) A gridded multisite weather generator and synchronization to observed weather data. Water Resources Research 45 WWW: