Linking the climate change scenarios and weather generators with agroclimatological models

Transcription

1 Linking the climate change scenarios and weather generators with agroclimatological models Martin Dubrovský (IAP Prague) & Miroslav Trnka, Zdenek Zalud, Daniela Semeradova, Petr Hlavinka, Eva Kocmankova, Lenka Bartosova (MUAF Brno) Sassari, May 25 June 5, 2009 acknowledgements: - Sassari University & CNR - Institute of Atmospheric Physics (IAP) - Mendel University of Agriculture and Forestry (MUAF) - grant agencies: GAAV (project IAA ), GACR (project 205/05/2265, NAZV (project QG60051)

2 part 3 weather generators

3 what are weather generators, and what they may be used for? What the WGs are: software, which produces weather series which are similar (in terms of statistics AVGs, VARs, extremes, correlations, ) to the observed series What they may be used for generation of arbitrarily long weather series generation of weather series for sites where no observations exist (because it provides better results than interpolation of weather series) generation of weather series which continue the observed weather (e.g. for probabilistic crop yield forecasting) generation of weather series representing changed climates (climate sensitivity studies, climate change impact studies various characteristics /not only means!/ may be changed) [in doing this, WG parameters are modified accordning to climate change scenarios]

4 weather generators - introduction spatial resolution: single-site (OK for crop growth model) multi-site or spatially continuous (required in hydrological modelling; but may be also used in wild-fire modelling) temporal resolution (~time step) hourly daily monthly number of variables single-variate multi-variate (CERES: 4 vars; WOFOST: 6 vars) stand-alone surface weather generator vs conditioned on circulation parametric vs. non-parametric parametric: WGEN, SIMMETEO, Met&Roll, M&Rfi /= Met&Roll flexible and improved) semi-parametric (Semenov: LARS-WG) non-parametric (nearest neighbours resampling)

5 Nearest neighbours resampling learning SRAD TMAX TMIN RAIN... xx xx xx xx xx xx xx xx xx xx xx xx xx xx xx synthetic SRAD TMAX TMIN RAIN

6 Met&Roll / M&Rfi - history * 1995: first version (based on WGEN [Richardson, 1981]) to be used with CERES crop models since 1995: improvements of the model Markov(1) > Markov(3) conditioning on monthly WG previous applications: - crop growth modelling (together with MUAF; in climate change impact studies, probabilistic seasonal crop yield forecasting -> PERUN system /2001/) - hydrological modelling : interpolation 2007: M&Rfi developed based on Met&Roll

7 Met&Roll / M&Rfi : model Met&Roll = 4-variate stochastic daily weather generator: step 1: PREC occurrence ~ Markov chain (order: 1-3; parameters: trans.prob.) Prob(PREC(t)>0) = P01 if PREC(t-1) = 0 P11 if PREC(t=1) = 1 step 1b (only if PREC(t)>0) : PERC. amount ~ Gamma distribution (parameters: α, β /~ shape, scale/) step 2: SRAD, TMAX, TMIN: AR(1) model (parameters: A, B, avg(x i ), std(x i ),) X*(t) = AX*(t) + Be where X = [SRAD, TMAX, TMIN] X* i = [X i avg(x i )] / std(xi) e = white noise A, B = [3x3] matrices - all parameters are assumed to vary during the year -daily WGis linked to AR(1)-based monthly WG (to improve low-frequency variability)

8 4-variate 6-variate (nearest neighbours resampling) 4-variate SRAD TMAX TMIN RAIN variate SRAD TMAX TMIN RAIN VAPO WIND learning SRAD TMAX TMIN RAIN VAPO WIND... xx xx xx xx xx xx xx xx xx xx xx xx xx xx xx

9 testing quality of the weather generator motivation: WG cannot perfectly fit the structure of the real-climate weather series direct validation comparison of observed vs. synthetic weather series in terms of derived climatic characteristics (synthetic wea.series should resemble observed series) indirect validation comparison of characteristics derived from model output (e.g. crop growth model) fed by OBS and SYNT weather series (outputs from impact model fed by OBS and SYNT wea.series should resemble each other)

10 scheme of WG validation obs.wea.series (~15-30 years ) WG (present climate) synt.weather series (present climate) info about - plant genetics - soil properties -growing site - management calculation of selected climatic characteristics climatic chars. (obs.wea.) B climatic chars. (synt.wea.) CROP GROWTH MODEL model yields (obs.wea.) C model yields (synt.wea: presence ) B: direct validation of WG C: indirect validation of WG

11 Met&Roll - direct validation normality of SRAD, TMAX, TMIN: variability of monthly means: length of dry periods:

12 Met&Roll indirect validation Motivation: How the WG imperfections (to fit the structure of real-world weather series) affect output from impact models fed by synthetic series?

13 indirect validation of Met&Roll using crop model AVGs and STDs of wheat yields (17 stations x 3 versions of WG) crop model: CERES-Wheat; 30-y simulations for 17 Czech stations weather generator: - WG-BAS: basic WG: no annual cycle of AR matrices; 1 st order Markov chain - WG-A3: improved WG: annual cycle of AR matrices; 3rd order Markov chain - WG-A3M: best WG: WG-A3 + conditioned on monthly WG (Dubrovsky et al, 2004, Climatic Change)

14 indirect validation of Met&Roll using rainfall-runoff model monthly maxima of daily model (SAC-SMA) streamflows (39-y simulations) [DBZ, 2004]

15 Probability distribution of 5-day average model streamflow simulated with observed (CB) and synthetic weather series [DBZ, 2004]

16 interpolation of Met&Roll weather generator (calimaro project) Czech institutes (9 people) participated Main aim: interpolation of Met&Roll parameters motivation: applicability of Met&Roll for sites without observations sub-aims: 1. choice of the interpolation methods 2. validation in terms of the climatic characteristics 3. validation in terms of outputs from models fed by synthetic series produced by the interpolated generator crop model hydrological rainfall-runoff models

17 Methodology - stochastic daily weather generator Met&Roll [only basic version used in these experiments ~ Richardson s WGEN (1981)] 4 daily variables: PREC: - occurrence ~ Markov chain (order = 1) - amount ~ Gamma distribution (parameters: α, β /~ shape, scale/) SRAD, TMAX, TMIN: standardised deviations from their mean annual cycle are modelled using AR(1) model (parameters: A, B, avg(x i ), std(x i ),) - all parameters are assumed to vary during the year

18 Data - Station weather data: 125 stations from Czechia with data coverage greater or equal to 75% during (= at least 75% of terms have observed values of TMAX, TMIN, PREC, and CLOUD or SUNSHINE used to estimate SRAD) - Topography of Czechia is derived from the global digital elevation model GTOPO30 (horizontal grid spacing = 30 arc seconds; i.e. approximately m resolution within the territory of Czechia) [

19 Region (= Czechia) stations available: a) circles: learning set, b) squares: validation set - altitude varies from 115 to 1602 m a.s.l.

20 Scheme: Testing interpolated weather generator WG parameters: Met&Roll Parameters of site-calibrated WG Met&Roll interpolation A(int) Parameters of interpolated WG Met&Roll Daily weather series (125 st.): observed series synthetic series synthetic series Climatic characteristics: climatic characteristics B(wg) climatic characteristics B(int) climatic characteristics impact model (crop-growth model, hydrological model,...) Output from impact models: Crop yields, river streamflows C(wg) Crop yields, river streamflows C(int) Crop yields, river streamflows A(int) B(wg) B(int) C(wg) C(int) accuracy of interpolation ability of WG to reproduce climatic characteristics effect of interpolation of WG on climatic characteristics in synt. series effects of WG inaccuracies on impact models output effect of interpolation of WG on impact models output

21 1) choice of the interpolation technique 1) co-kriging (used via ArcGIS) 2) neural networks [Multilayer Perceptron network type = 3-5-1, 29 degrees of freedom, Back Error Propagation and Conjugate Gradient Descent training algorithms used] 3) weighted nearest neighbours y(x,y,z) = weighted average from the surrounding stations (d<100km; bellshaped weight function) corrected for the zonal + meridional + altitudinal trends + WG parameters are mapped using GTOPO30 digital elevation map

22 validation in terms of Γsc Γsh (Jul) (highest RV) co-kriging: RV = 59% nearest neighbours: RV = 70% neural networks: RV = 63% topography Color of the stations symbols relate to station-specific interpolation error

23 validation in terms of Pwet (Jan.) co-kriging: RV = 56% nearest neighbours: RV = 69% neural networks: RV = 59% topography Color of the stations symbols relate to station-specific interpolation error

24 validation in terms of P wet dry (Jan) co-kriging: RV = 17% nearest neighbours: RV = 41% neural networks: RV = 28% topography Color of the stations symbols relate to station-specific interpolation error

25 validation in terms of Γsh (July) (lowest RV) co-kriging: RV = 8% nearest neighbours: RV = 10% neural networks: RV = 6% topography Color of the stations symbols relate to station-specific interpolation error

26 interpolated WG: validation in terms of precipitation characteristics Nearest Neighbors interpolator was found the best and selected for the additional experiments (Box 2 and Box 3)

27 comparison of WG parameters: site-calibrated vs interpolated WG parameters Interpolation methods Neural Net. Nearest Neighb. co-kriging r(λ,φ,z) RV [%] RV [%] RV [%] a) Precipitation: Γsh (shape parameter of the Γ distribution) JAN JULY b) Precipitation: Γsh*Γsc (product of shape and scale parameters of the Γ distribution JAN JULY c) Precipitation: probability of wet day occurrence JAN JULY d) Precipitation: transitional prob. of wet day occurrence (given the previous day was JAN JULY e) Daily temperature maximum AVG (JAN.) AVG (JULY) STD (JAN.) STD (JULY) f) Daily temperature minimum AVG (JAN.) AVG (JULY) STD (JAN.) STD (JULY)

28 2) validation in terms of extreme precipitation characteristics - only NearestNeighb. and NeuralNetwork -

29 observed Annual max. length of dry spell B(wg) WG-site (site-calibr.) WG-int. (Neighb.) B(int) WG-int. (Neur.Netw.) B(int) - B(int):... WG underestimates max.length of dry spell (the fit is better if MC3 model is used) - B(int):. interpolated WG performs similarly as the site-calibrated WG - B(int):. both interpolation techniques perform similarly - B(wg) vs B(int):. differences due to interpolation are lower than those due to WG imperfections

30 observed Annual max. length of wet spell B(wg) WG-site (site-calibr.) WG-int. (Neighb.) WG-int. (Neur.Netw.) B(int) B(int) - WG simulates dry spell better than the dry spells

31 observed Annual max. 1day precipitation B(wg) WG-site (site-calibr.) WG-int. (Neighb.) B(int) WG-int. (Neur.Netw.) B(int) - B(wg): WG underestimates annual extreme precipitation - B(wg) vs B(int): differences due to interpolation are lower than those due to WG imperfections

32 observed Annual max. 5-day precipitation B(wg) WG-site (site-calibr.) WG-int. (Neighb.) B(int) WG-int. (Neur.Netw.) B(int) Similar as 1-day PREC but even more pronounced: - B(wg): WG underestimates annual extreme precipitation - B(wg) vs B(int): differences due to interpolation are lower than those due to WG imperfections

33 Annual extreme precipitation characteristics - summary L(dry) L(wet) 1d-PREC 5d-PREC OBS site-cal. WG WG-int. (Neighb.) WG-int. (Netw.)

34 30-year extreme precipitation characteristics - summary L(dry) L(wet) 1d-PREC 5d-PREC OBS site-cal. WG WG-int. (Neighb.) WG-int. (Netw.) Results are similar as in the case of the annual extremes

35 Czechia vs. Nebraska 12.7 times higher density of stations (in this experiment) in Czechia

36 L(Dry) (mean annual maximum) A B A B - the differences are due to WG imperfections - the differences are due to imperfections in interpolation

37 PREC-1d (mean annual maximum) A B A B - the differences are due to WG imperfections - the differences are due to imperfections in interpolation

38 TMAX (mean annual maximum) A B A B - the differences are due to WG imperfections - the differences are due to imperfections in interpolation

39 Czechia vs Nebraska : annual statistics shows inaccuracy of WG shows inaccuracy of interpolation

40 Czechia vs Nebraska : 30y statistics

41 3) indirect validation of interpolated Met&Roll Motivation: We have found imperfections in reproducing climatic characteristics by interpolated WG. Q: How these imperfections affect output from crop model (or any other model) fed by weather series produced by the interpolated WG?

42 Scheme: Testing interpolated weather generator WG parameters: Met&Roll Parameters of site-calibrated WG Met&Roll interpolation A(int) Parameters of interpolated WG Met&Roll Daily weather series (125 st.): observed series synthetic series synthetic series Climatic characteristics: climatic characteristics B(wg) climatic characteristics B(int) climatic characteristics impact model (crop-growth model, hydrological model,...) Output from impact models: Crop yields, river streamflows C(wg) Crop yields, river streamflows C(int) Crop yields, river streamflows A(int) B(wg) B(int) C(wg) C(int) accuracy of interpolation ability of WG to reproduce climatic characteristics effect of interpolation of WG on climatic characteristics in synt. series effects of WG inaccuracies on impact models output effect of interpolation of WG on impact models output

43 interpolated WG: Indirect validation via STICS model AVG(model wheat yields) [soil = Chernozem (CZ_01)] yields simulated with observed weather interpolated yields yields simulated with site-calibrated WG yields simulated with interpolated WG

44 validation of Met&Roll - summary Direct validation: imperfections exist in reproducing some climatic characteristics Indirect validation: - good applicability for crop models - hydrological models are more affected by WG imperfections (problems with fitting extremes) Interpolated WG: 1. nearest neighbours interpolator is the best 2. interpolated WG provides reasonable climatic characteristics errors due to interpolation are lower than errors due to WG imperfections CZ ~ NE 3. estimating crop yields in sites with no weather observations: running crop model with weather series generated by interpolated WG provides better results than interpolation of crop yields

45 M&Rfi - M&Rfi was originally developed for FAO, Rome - available on web: is a follower of Met&Roll ( = Met&Roll Flexible and Improved) the same core model: PREC is main variable other variables are (optional) conditioned on PREC daily WG (DWG) may be linked to monthly WG (MWG)!!! freely available from web: demo batch files + user s guide!!!

46 M&Rfi main features optional number of variables (<=8) [typically 3 or 4: (PREC, SRAD, (TMAX + TMIN) or (TAVG + DTR) or TAVG) optional time step (1d, 3d, 5d, 1w, 10d*, 2w, ½m, 1m) 1 variable (PREC) is optionally the conditioning variable transformation of variables may better treat non-normal variables (allows parametric & non-parametric transformations) VAPO and WIND are first candidates for inclusion) estimation of solar radiation from cloudiness or sunshine estimation of evapotranspiration using Penman-Monteith equation more user-friendly (guide available) run via command line [~M&R] all WG parameters stored in a single file, more stations may be stored in a single file the synthetic weather series may be forced to fit [~M&R] weather forecast for a forthcoming period (following days, month or whole season) climate change scenario (including changes in both high-frequency and lowfrequency variability) through modifying WG parameters through direct modification of input weather series

47 M&Rfi: effect of time step [Dubrovsky and Grieser 2007, EGU] Experiment: 8 European + 11 US stations 3 versions of M&Rfi: dt = 1 day, 10 days, 1 month each station: 1 observed series 30 synthetic series for each of 3 WGs Experiment: Result: the best performance in reproducing monthly variability is obtained by monthly WG