High resolution probabilistic regional climate projection ...€¦ · multi-model ensemble experiments. An appropriate combination of statistical methods and optimization of climate

High resolution probabilistic regional climate projection using a regression method with multi--model ensemble

Koji DAIRAKU1,

Noriko N. ISHIZAKI1 , Genta UENO2

1) National Research Institute for Earth Science and

Disaster Resilience, Japan

2) The Institute of Statistical Mathematics, Japan

18th May 2016, Stockholm, Sweden

Future scenario of regional precip.

※陰影：両者の差が信頼度95%で統計的に有意

MIROC3.2

a)  NHRCM  JJA b)  NRAMS  JJA c)  T-­‐WRF  JJA

NHRCM N-RAMS T-WRF

Ø  Future changes in JJA mean daily Precip.: ・ Precip. increases are noticeable in the west and south sides of the mountainous regions where precipitation amounts are large in the recent climate. ・ Relationship between precip. changes and topography is unclear in GCM simulations. Downscaling added values Tsunematsu, Dairaku, Hirano, JGR, 2013

u Climate change caused by human activities will continue for centuries. Need at least several decades until a mitigation take effect. Need to put adaptation together immediately. u Fundamental information for risk assessment on climate change, such as the frequency and magnitude of wind and water disaster, is of great concern. u Probabilistic climate scenario information for increasing ensemble experiments.

Climate information and services for Impacts, Adaptation and Vulnerability (IAV) Assessments are of great concern. In order to develop probabilistic regional climate information that represents the uncertainty in climate scenario experiments, we compared the physics ensemble experiments using the 60km global atmospheric model of the Meteorological Research Institute (MRI-AGCM) with multi-model ensemble experiments with global atmospheric-ocean coupled models (CMIP3 and CMIP5) . We also develop a prototype of probabilistic climate information using the multi-model ensemble experiments. Climate information and services for Impacts, Adaptation and Vulnerability (IAV) Assessments are of great concern. In order to develop probabilistic regional climate information that represents the uncertainty in climate scenario experiments in Japan, we compared the physics ensemble experiments using the 60km global atmospheric model of the Meteorological Research Institute (MRI-AGCM) with multi-model ensemble experiments with global atmospheric-ocean coupled models (CMIP3) of SRES A1b scenario experiments. The MRI-AGCM shows relatively good skills particularly in tropics for temperature and geopotential height. Variability in surface air temperature of physical ensemble experiments with MRI-AGCM was within the range of one standard deviation of the CMIP3 model in the Asia region. On the other hand, the variability of precipitation was relatively well represented compared with the variation of the CMIP3 models. Models which show the similar reproducibility in the present climate shows different future climate change. We couldn’t find clear relationships between present climate and future climate change in temperature and precipitation. We develop a prototype of probabilistic climate scenario using the multi-model ensemble experiments. An appropriate combination of statistical methods and optimization of climate ensemble experiments are discussed.

u Bottom-up Approach Assessment of potential vulnerabilities of resources (water, food, energy, health, ecosystem) . Climate change is one of many factors. Emphasis on risk assessment and disaster management. Focused on adaptation and mitigation measures to reduce vulnerability (Pielke Sr. et al., 2009). Robust adaptive strategies for "possible scenario" instead of “accurate prediction for optimization"

→ Complement mainstream top-down approach

Introduction

(UK Climate Projections Briefing report, 2009 )

General concept of probabilistic climate projection

（Collins et al., 2012）

・M: Model ・p：parameter space in a model ・R：radiative forcing Rh（past）, Rf（future） ・Ch：Simulated space in historical simulations ・Cf：Simulated space in future simulations 1. Identify more reasonable parameter space based on reproducibility.

2. Constrain parameter space in future climate.

Climate variable (C) simulated by a model (M)

Research Purposes Develop probabilistic climate scenario information by quantification of model’s structural and parametric uncertainties.

Performance of ensemble experiments

Probability distribution estimated by weighted ensemble Probability map of future

climate change scenario

Probability of △△ K (mm/day) increase (decrease) of temperature (precipitation) in future ○○ (30/100) years under a socio-economic scenario

・Perturbed physics ensemble（PPE） 　→ UKCP09（Sexton et al., 2012)

Sampling a possible combination of parameters Develop emulator based on PPE Estimate probability using a multivariate Gaussian mixture. 　　-needs a large number of ensemble experiments 　　-need to preset discrepancy term (model error) and

observation error

・Multi-model ensemble（MME） 　→ Tebaldi and Knutti (2007), Tebaldi et al. (2004), Greene et al. (2006), Furrer et al. (2007)

Assume statistically independence of multi-model ensemble. Multi-model ensemble may not sample all possible future

climate scenarios.

Two approaches

( )

( )0 1

2~

,

0,

N ny xt n t tn

Nt

w

w

β β

σ

= + ∑=

+

Regression Model

Variance does not depend on the time (t)

Selection of γ, α: Cross-validation

( )2

1 ( ) ( )ˆ ˆCV 01 1

T N nt ty xt n tT t nβ β

⎛ ⎞− −= −∑ ∑⎜ ⎟⎜ ⎟= ⎝ ⎠

−=

•  Suppress absolute values of regression coefficients (prevent overfitting） •  Avoid multi-collinearity between explanatory variables

•  Estimate β without yt. Then compared with yt

Estimate coefficient β: Regularized least squares method (Elas%c  net)

( ) ( )2

2101 1 0 0

N NT N nJ y xt n t n nt n n nβ β γ α β α β

⎡ ⎤⎛ ⎞= − −∑ ∑⎜ ⎟ ∑ ∑− + ⎢ ⎥⎜ ⎟ ⎢ ⎥= = ⎠ ⎣

+⎝ = = ⎦

•  Coefficient  of  redundant  explanatory  variables    are  set  to  0  (model  selec%on)  

ü  No need to preset model error and observation error ü  Development of emulator is unnecessary ü  Applicable to other regions and other variables

Estimate probability distribution

Estimated probability distribution (Gaussian mixture)

Estimated Gaussian distribution of each time (t)

yt

Construct Regression Model

σ̂

( )ˆ ˆ0 1

N nxn tnβ β+ ∑

=

Estimate coefficient β by Regularization (Elas%c  net) & Cross-validation

Estimate Gaussian distribution at a time (t)

Estimate Gaussian distribution at each time (t)

Observation： NIAES 1km AMeDAS mesh data（aggregated onto 20km grid） Climate models： CMIP3 SRES A1b 21 models（2.5°grid） Variable： Monthly 2m air temperature Present： 1979-1998 Future： 2079-2098 Area： Japan Regression Method： Elastic Net Comparison with a bias-correction: CDFDM (Cumulative distribution function-based downscaling method, Iizumi et al., JGR,2011)

Probability map in Japan

Validation

Monthly 2m air temperature range of mean value ±1 standard deviation in present climate in Tokyo (1979-1998).

•  Relatively larger warm biases of GCMs in winter

•  Biases of GCMs are corrected by probabilistic model (Regression model: REG) and the bias correction method (CDFDM).

•  Probabilistic model (REG) is better than the bias correction (CDFDM).

•  Relatively worse skill of CDFDM is partly due to data sorting and correction regardless of season and month.

Probability map in Japan

20-km grid probability map of 2m temperature increase exceeding 2K (a,e), 4K (b,f), 6K (c,g), and the increase of 90 %ile for January (upper panels) and July (lower panels).

The probabilistic model can produce various probabilistic information. This model is applicable to any regions where sufficient observation is available.

Jan

Jul

CMIP3 SRESA1b 21GCMs

Ishizaki et al., in preparation

Future(2079-2098) - Present（1979-1998)　

Summary

u Develop a method using a regression model to produce probabilistic climate scenario information by quantification of structural uncertainties of ensemble experiments. u Develop prototype of probabilistic climate scenarios in Japan (also CORDEX-EA: PA-028）by using CMIP3 21 GCMs. u In Japan, the probabilistic model shows better skill in mean and standard deviation) of 2m air temperature than simple ensemble mean of GCMs and also a bias-correction method (CDFDM). u The probabilistic model can produce climate information such as exceedance probability of temperature increase and quantile values (Addded values).

u Extend to low-frequent extreme events (high impact low-probability events) by employing big data, large number of ensemble experiments u Develop hierarchical probabilistic model (e.g., Kerkhoff et al., JC,

2015) for heterogeneous GCMs x RCMs multi-model ensemble downscaling in East Asia (Tsunematsu et al., JGR, 2013; Li et al., IJC in press; Tang et al., IJC in press; Nayak et al., JMSJ, in minor revision)

u Guidance information of experimental design for a rich diversity of future multi-model ensembles. u  Climate service (Science as a service) 　Establish CORDEX Pan-Asia statistical downscaling(SD) group

　New project SI-CAT (Social Implementation Program on Climate Change Adaptation Technology:2015-2020). Engagement of stakeholder(co-design).

　Develop advanced “interface” with stakeholders (DIAS: Data Integration & Analysis System).

Issues and Next Steps

Acknowledgments This study is supported by the SOUSEI Program Theme –C (PI: Izuru Takayabu in MRI), the SI-CAT (PI: Yoichi Ishikawa in JAMSTEC), and was conducted as part of the research area "Vulnerability and Adaptation to Climate Change in Water Hazard Assessed Using Regional Climate Scenarios in the Tokyo Region” (NIED; PI: Koji Dairaku) in the RECCA Program, funded by MEXT, Government of Japan.

Part of data collection and analysis is supported by the Data Integration and Analysis System (DIAS). In addition, a program developed by the National Institute for Agro-Environmental Sciences (NIAES) was used for AMeDAS mesh data.