Transcript 20130814 QED2013 JWinkler Tart Cherries and Bamboo

Using Downscaled Data in the
Real World: Sharing Experiences
Julie Winkler
Michigan State University
Introduction
• Describe the selection and application of climate
projections for four climate change assessments that
vary in terms of:
–
–
–
–
assessment objectives
climate variables of interest
relevant spatial and temporal scales
nature of the “downstream” impact models that employ
the downscaled climate projections
– geographic focus
• The assessments share in common an “end-to-end”
approach (also referred to as “top-down” approach)
Assessment Projects
• Potential impacts of climate change on perennial crop
production and tourism/recreation in Michigan
• Vulnerability of understory bamboo habitat and panda
distribution in China’s Quinling Mountains to climate
change
• Climate change impacts on corn and soybean
production in the Upper Great Lakes Region of the
United States
• Development of an integrated framework for climate
change impact assessments for international market
systems with long-term investments
SOUR CHERRY PRODUCTION IN
MICHIGAN
Background/Considerations
• Production occurs in the lake-modified regions along
Lake Michigan
• Damaging spring temperatures are the major limiting
factor on perennial crop yield in Michigan, with
precipitation during pollination a secondary factor
• Downstream models were phenology and yield models
– These models were developed at the “point” (i.e., local)
level using climate observations from COOP stations with
relatively long-term records
– Considerable stakeholder involvement
Climate Projections
• “Custom” projections (i.e., developed specifically for the
project)
– Local scale (i.e., individual sites)
– Daily time step
– Maximum temperature, minimum temperature, wet/dry days,
precipitation amount
• Indices (e.g., frequency of freezing temperatures after reaching heat
accumulation thresholds)
– Statistical downscaling of GCM simulations
– Perfect Prog approach
• Multiple regression using surface and upper-level “circulation”
variables as predictors
– Ensemble
• four GCMs, two greenhouse gas emissions scenarios (SRES A2 and
B2), several “variants” of downscaling transfer functions
– Web-based user tools for stakeholders
NCAR
ECHAM
Canadian
Hadley
A2, B2
multiple
downscaling
methodologies
• 15 Locations
• 4 climate parameters
– Tmax
– Tmin
– Wet/dry days
– Precipitation amount
• 4 GCMs
– CCC CGCM2
– HadCM3
– MPI ECHAM4
– NCAR CSM1.2
• 2 Emission scenarios
– A2, B2
• Multiple empirical
downscaling
methodologies
Advantages/Limitations
• Advantages
– Projections had the temporal and spatial resolution necessary for
impact models
– Local site factors implicitly included
– Spatial autocorrelation retained
– Research team intimately familiar with nature of the projections and
their limitations
– Large ensemble to characterize uncertainty
• Limitations
–
–
–
–
Considerable time, effort, and expense to develop the projections
Projections available for only a small number of locations
Projections for climate variables developed separately
Local site factors (e.g., Lake Michigan) included implicitly rather than
explicitly
– Explained variance for precipitation transfer functions was small
– Working with a large scenario ensemble caused some angst among
team members
– Stationarity?
References
• Web site: www.pileus.msu.edu
• Winkler, J.A., J.A. Andresen, J. Bisanz, G.S. Guentchev, J. Nugent,
K. Piromsopa, N. Rothwell, C. Zavalloni, J. Clark, H.K. Min, A.
Pollyea, and H. Prawiranta, 2012: Michigan’s Tart Cherry
Industry: Vulnerability to Climate Variability and Change In S.C.
Pryor [Ed] Climate Change in the Midwest: Impacts, Risks,
Vulnerability and Adaptation, Indiana University Press, 104-116.
ISBN: 978-0-253-00682-0
• Winkler, J.A., J.P. Palutikof, J.A. Andresen, and C.M. Goodess, 1997:
The simulation of daily time series from GCM output. Part 2: A
sensitivity analysis of empirical transfer functions for downscaling
GCM simulations. Journal of Climate, 10, 2514-2532.
UNDERSTORY BAMBOO HABITAT AND
PANDA DISTRIBUTION IN CHINA’S
QUINLING MOUNTAINS
Background/Considerations
• Elevation a key consideration
• Bioclimatic models developed by the research
team
– Dependent variable:
• bamboo presence data from field plots covering the
elevational range of the distributions of three dominant
bamboo species within the Qinling Mountains
– Independent variables (from WorldClim database):
•
•
•
•
gridded values of 19 bioclimatic variables
Long-term (1950-2000) averages
30×30 arc second resolution (ca. 1 km2)
thin plate spline interpolation
Climate Projections
• “Off the shelf” projections:
– WORLDCLIM
• Three time slices 2010 – 2039, 2040 – 2069 and 2070 – 2099) [IPCC
TAR]
• Four GCMs (CCSR/NIES, CGCM2, CSIRO-Mk2 and HadCM3) [IPCC TAR]
• SRES A2 and B2 greenhouse gas emissions scenarios
• ca. 1 km2 resolution
– International Center for Tropical Agriculture (CIAT)
•
•
•
•
One future time slice (2040 – 2069)
15 GCMs (IPCC AR5)
SRES A2 greenhouse gas emissions scenario
ca. 1 km2 resolution
• Projections in the form of “deltas” from a reference period
Advantages/Limitations
• Advantages
–
–
–
–
Readily available, ease of use
Fine resolution
Includes widely-used bioclimate variables
Explicit consideration of topographic variations
• Disadvantages
– “black box”
– sensitivity of projections to different interpolation
algorithms is unknown
– difficult for users to evaluate
– needed to “piece together” projections from two sources
to cover time period of interest
Reference
•
Tuanmu, M-N, A. Viña, J.A. Winkler, Y. Li, W. Xu, Z. Ouyang, and J. Liu, 2013:
Climate change impacts on understory bamboo species and giant pandas in China’s
Qinling Mountains. Nature Climate Change, 3: 249–253 doi:10.1038/nclimate1727.
Projected future distributions of climatically suitable areas (CSAs) in
2070 – 2099 for the three bamboo species studied under the climate
projections from four IPCC TAR GCMs
GCM-related uncertainty of projected changes in
giant panda habitat area for the time slice of 2040 –
2069 under the SRES A2 greenhouse gas emissions
scenario.
CORN AND SOYBEAN PRODUCTION IN
THE UPPER GREAT LAKES REGION OF
THE UNITED STATES
Background/Considerations
• Goal was to evaluate potential latitudinal
shifts/expansion of corn and soybean production
in the Upper Great Lakes region
• Impacts models employed were:
– CERES-Maize
– CROPGRO-Soybean
• Interested in county-level yield
• Required climate variables:
– Daily maximum temperature, minimum temperature,
precipitation, solar radiation
Climate Projections
• Developed from NARCCAP simulations
– Used 8 RCM/GCM simulations (CRCM/ccsm, CRCM/cgcm3,
HRM3/hadcm3, HRM3/gfdl, RCM3/cgcm3, RCM3/gfdl, WRFG/ccsm,
and WRFG/cgcm3)
– SRES A2 greenhouse gas emissions scenario
– 50 km2 resolution
– Time slices:
• future period (2041-2070)
• historical periods (1971-2000)
• Used “delta” method (calculated by month) to adjust for biases and
downscale to local level
• Adjusted daily time series of precipitation and maximum and
minimum temperature for 34 USHCN stations across the study
region
– The stations were selected for their representativeness of the regional
climate variations and the quality (i.e., percent missing values) of their
time series
– Employed a climate regionalization (PCA/cluster analysis)
– Counties were assigned to stations
CRCM/cgcm3 CRCM/ccsm
RCM3/cgcm3 HRM3/hadcm3 HRM3/gfdl
RCM3/gfdl
WRFG/cgcm3 WRFG/ccsm
Projected changes in maximum temperature (left),
minimum temperature (middle), and precipitation
(right) between the future (2041-2070) and historical
(1971-2000) period for the eight NARCCAP models.
Temp. Change
<=1.0
1.0 - 1.5
1.5 - 2.0
% Precip Change
2.0 - 2.5
2.5 - 3.0
3.0 - 3.5
3.5 - 4.0
>4.0
<=-10
-10 - -5
-5 - 0
0-5
5 - 10
10 - 15
15 - 20
20 - 25
25 - 30
>30
Climate Stations
Representative Stations
Climate Region
1
4
6
2
5
7
3
ARGL
MADA
GRFR
PRMN
PRDM
THBR
CLQT
IRWD
STBG
SPNR
MWES
MILA
CBGN
MSFD
PSTN
NWUM
MNSG NWBR
BWLR OCNT
ZMRT
ETWS
FDLC
GMDW
HART
BRWT
PRDC
BRHD
STHV
ALGN
MDLN
OWSO
MCLN
ANBR
CDWR
ADRN
Selection of
representative climate
stations for the regional
climate change impact
assessment on corn and
soybean production in
the Upper Midwest
based on county-climate
memberships (above)
and their assignment to
counties (below). Four
characters are the
abbreviation to
distinguish the
representative climate
stations with colors
indicated counties
assigned with the same
representative climate
station.
Advantages/Limitations
• Advantages
– Realistic location/outline of Great Lakes
– Delta method is simple to apply
• Limitations
– Small number of USHCN stations and non-uniform
distribution
– Lost some of the spatial detail available from
NARCCAP simulations
– Did not consider changes in variability or frequency of
wet/dry days
– Stationarity assumption
ET
CRCM-ccsm
CRCM-cgcm3
HRM3-gfdl
HRM3-hadcm3
RCM3-cgcm3
RCM3-gfdl
WRFG-ccsm
WRFG-cgcm3
Average
Change in the median of
cumulative seasonal
evapotranspiration (ET,
above) and crop yields
(Yield, below) between
the historical (1971-2000)
and the future (20412070) period for corn
production in the Upper
Midwest at the reference
level of CO2 (370 ppm)
concentration
Yield
CRCM-ccsm
CRCM-cgcm3
HRM3-gfdl
HRM3-hadcm3
RCM3-cgcm3
RCM3-gfdl
WRFG-ccsm
WRFG-cgcm3
Average
ET Changes (%)
<=-15
-15 - -10
-10 - -5
-5 - 0
Yield Changes (%)
0-5
5 - 10
10 - 15
>15
<=-50
-50 - -25
-25 - 0
0 - 25
25 - 50
50 - 100
100 - 150
>150
References
• Perdinan, 2013. Crop production and future
climate change in a high latitude region: a
case study for the Upper Great Lakes region of
the United States. PhD Dissertation. Michigan
State University. Completed May, 2013.
DEVELOPMENT OF AN INTEGRATED
FRAMEWORK FOR CLIMATE CHANGE
IMPACT ASSESSMENTS FOR
INTERNATIONAL MARKET SYSTEMS
WITH LONG-TERM INVESTMENTS
(CLIMARK)
Background/Considerations
• Impetus came from stakeholders of the tart cherry industry
• Traditional local/regional climate impact and adaptation
assessments do not consider important spatial and
temporal interactions for international market systems
• Other important production regions include central and
eastern Europe
• Assume that supply and demand are linked through
international trade
• Local, daily climate projections needed for several locations
within each of the major production regions
Expanded Assessment Framework
Climate
Projections
Climate
projections
for
production
regions for
Time Slice
#1
Climate
projections
for
production
regions for
Time Slice
#2
Base Situation
(industry structure, economic factors, and regional
constraints)
Major System Components for Time
Slice #1:
•Regional climate scenarios
•Phenology and yield models
•Trade model
Major System Components for Time
Slice #2:
•Regional climate scenarios
•Phenology and yield models
•Trade model
Between Time Slice Projected Changes
Adaptation
(e.g.,
cultivars,
growing
regions)
Economic
factors (e.g.,
consumer
preferences,
income)
Regional
constraints
(e.g.,
infrastructure,
institutions)
Between Time Slice Projected Changes
Climate
projections
for
production
regions for
Time Slice
#3
Major System Components for Time
Slice #3:
•Regional climate scenarios
•Phenology and yield models
•Trade model
Adaptation
(e.g.,
cultivars,
growing
regions)
Economic
factors (e.g.,
consumer
preferences,
income)
Regional
constraints
(e.g.,
infrastructure,
institutions)
Climate Projections
• Hybrid downscaling
– Start with dynamically-downscaled projections:
• NARCCAP (mid-century time slice)
• ENSEMBLES (21st century)
– Apply bias correction and empirical downscaling
• Hybrid projections supplemented with statisticallydownscaled projections using simple “delta”
approach applied to CMIP5 model output for 21st
century
Need for Bias Correction
Observed and simulated values of minimum
temperature for winter (December, January,
February).
Types of Bias Correction and
Empirical Downscaling Techniques
Accuracy-driven:
Goal is to minimize overall prediction error
N
min

 L  y , f ( x ;  )  ( )
a
i 1
i
i
Examples: MLR and its variants (ridge and
lasso regression), analog methods, nonlinear
models (neural networks, HMM)
Distribution-driven:
Goal is to minimize error of fitted distribution
N
min

 L z , f ( x ;  )  ( )
i 1
d
i
i
such that Fobs z i   Fw  f ( xi ;  ) 
Examples: quantile mapping (QM), histogram
equalization (Piani et al)
MLRCDF: Multi-Objective Regression
• Current techniques are designed to optimize either
accuracy or fit to observed distribution, but not
both
• MLRCDF: a multi-objective regression method that
combines both objective functions
N
min

 L  y , f ( x ;  )  L z , f ( x ;  )  ( )
i 1
a
i
i
d
i
i
–  controls the trade-off between accuracy and fitting
the distribution
*adjusted time stamp for time of observation
QQ plot for daily precipitation over the test period for unadjusted and adjusted
output from WRFG driven by the NCEP reanalysis The blue line corresponds to
the QQ line while the dotted black line is the diagonal. Top row: Eau Claire;
Middle row: Maple City; Bottom row: Hart
Advantages/Limitations
• Advantages
– Capture some of the benefits of both dynamic and
statistical downscaling
• Limitations
– Only one future time slice for NARCCAP
– NARCCAP and ENSEMBLES do not use the same
emissions scenario
– Time consuming
– More ensemble members?
References
• Winkler, J.A., S. Thornsbury, M. Artavio, F.-M. Chmielewski, D.
Kirschke, S. Lee, M. Liszewska, S. Loveridge, P.-N. Tan, S. Zhong, J.A.
Andresen, J.R. Black, R. Kurlus, D. Nizalov, N. Olynk, Z. Ustrnul, C.
Zavalloni, J.M. Bisanz, G. Bujdosó, L. Fusina, Y. Henniges, P.
Hilsendegen, K. Lar, L. Malarzewski, T. Moeller, R. Murmylo, T.
Niedzwiedz, O. Nizalova, H. Prawiranata, N. Rothwell, J. van
Ravensway, H. von Witzke, and M. Woods, 2010: Multi-regional
climate change assessments for international market systems with
long-term investments: A conceptual framework. Climatic Change,
103, 445-470. DOI 10.1007/s10584-009-9781-1.
• Abraham, Z., P.-N.Tan, Perdinan, J. A. Winkler, S. Zhong, and M.
Liszewskak, 2013: Distribution Regularized Regression Framework
for Climate Modeling. Proceedings of SIAM International
Conference on Data Mining (SDM-2013), Austin, Texas. Available at:
http://knowledgecenter.siam.org/333SDM/333SDM/1
Closing Remarks
• Choice of climate projections is influenced by:
– Assessment goals
– Demands of impact models
• Hybrid downscaling is likely to become more
common