Transcript here - LOCA Statistical Downscaling

Statistical downscaling
using Localized
Constructed Analogs
(LOCA)
David Pierce and Dan Cayan
Scripps Institution of Oceanography
Bridget Thrasher, Edwin Maurer, John
Abatzoglou, Katherine Hegewisch
Development sponsored by
The California Energy Commission
Department of Interior/US Geological Survey via the
Southwest Climate Science Center
NOAA RISA Program through the California Nevada
Applications Program
Production runs sponsored by
U.S. Army Core of Engineers/USBR
NASA via computing resources
Downscaling system
Global
Models
Regridding
Bias
Correction
Spatial
Downscaling
Quantile Mapping
(QM)
Constructed Analogs
(CA; Hidalgo et al. 2008)
Bias Correction and
Constructed Analogs
(BCCA; Maurer et al. 2010)
Multivariate Adapted
Constructed analogs
(MACA; Abatzoglou & Brown 2012)
Bias Correction with
Spatial Disaggregation
(BCSD; Wood et al. 2004)
Issues with bias correction
1. QM does not preserve model-predicted changes
(Maurer and Pierce, HESS, 2014)
• Tmax
• Difference
between original
model-predicted
change and
change after bias
correction
• 2070-2100 minus
1976-2005
• Ensemble
averaged across
21 GCMs
deg-C
EDCDFm reference:
Li, H., J. Sheffield, and E. F. Wood, 2010: Bias correction of monthly precipitation and temperature
fields from Intergovernmental Panel on Climate Change AR4 models using equidistant quantile
matching. J. Geophys. Res. Atmos., 115 (D10101), doi:10.1029/2009JD012882.
What about precipitation?
• Evaluate temperature changes as a difference (degrees C)
• Evaluate precipitation changes as a ratio (percent)
– Positive definite
– Wide dynamic range
– Rain shadow regions
• Precipitation
• Difference
between original
model-predicted
change and
change after bias
correction in
percentage points
• 2070-2100 minus
1976-2005
• Ensemble
averaged across
21 GCMs
“PresRat” scheme
• Like EDCDFm (Li et al. 2010) except:
1. Preserves the ratio of model-predicted changes (not the difference)
2. Zero-precipitation threshold (preserve observed number of dry days in
historical period)
3. Final correction factor to preserve mean change
“PresRat” scheme
• Like EDCDFm (Li et al. 2010) except:
1. Preserves the ratio of model-predicted changes (not the difference)
2. Zero-precipitation threshold (preserve observed number of dry days in
historical period)
3. Final correction factor to preserve mean change
Correction factor
• Correction factors
necessary to
preserve modelpredicted
changes (20702099 vs. 19762005) in mean
precipitation
• Averaged across
21 GCMs
• Precipitation
• Difference
between original
model-predicted
change and
change after bias
correction in
percentage points
• 2070-2100 minus
1976-2005
• Ensemble
averaged across
21 GCMs
2. Model Errors can be a
Function of Frequency
If log-RMSE is f, then
models are off by factor
of (1 + f), on average
Log-RMSE metrics
How much does frequency-dependent bias correction change values?
3. Standard QM not multivariate
•
•
Temperature on precipitating days affects snow cover (Abatzoglou et al.)
Bias correct temperature conditional on precipitation > 0 or not
Issues with spatial downscaling
Spatial Downscaling with constructed analogs
Hidalgo, H.G., Dettinger, M.D., and
Cayan, D.R., 2008
Slide from Mike Dettinger, USGS (tenaya.ucsd.edu)
Spatial Downscaling with constructed analogs
Hidalgo, H.G., Dettinger, M.D., and
Cayan, D.R., 2008
Slide from Mike Dettinger, USGS (tenaya.ucsd.edu)
Spatial Downscaling with constructed analogs
Hidalgo, H.G., Dettinger, M.D., and
Cayan, D.R., 2008
Slide from Mike Dettinger, USGS (tenaya.ucsd.edu)
Spatial Downscaling with constructed analogs
Hidalgo, H.G., Dettinger, M.D., and
Cayan, D.R., 2008
Slide from Mike Dettinger, USGS (tenaya.ucsd.edu)
Issues with current downscaling (BCCA)
•
•
“BCCA” = “Bias correction with Constructed Analogs”
Averaging step reduces temporal variance (i.e., mute extremes)
2. Frequency of occurrence -> percent of amount
• Take an extreme example for illustration:
60% of the time
40% of the time
Contributes to reduction
in extremes
Slide 22
3. Drizzle problem from downscaling
Slide 23
New downscaling (LOCA) (Step 1 of 2)
•
BCCA uses 30 best matching analog
days over entire domain
•
LOCA starts with 30 best matching
analog days over the region around
the point
•
Region: everywhere correlation with
point being downscaled is > 0 (in
obs)
•
Regions are calculated by season
(DJF, MAM, JJA, SON) and variable
(pr, tasmax, tasmin, etc.)
•
Gives a natural domain
independence to LOCA (extending
domain past region does not affect
results at the point)
Example shown for precipitation
New downscaling (LOCA) (Step 2 of 2)
•
Once 30 regional analog days are
selected:
•
Find best one (of the 30) matching
days in a small localized region (~1
degree) around each point
This two step process means each
point:
– Is consistent with what’s
happening regionally
– Is the best match locally
•
• Points whose selected analog day is
different from a neighbor’s (“edge
points”) use a weighted average of
the relevant analog days
• ~30% of points are edge points
• Greatly reduced averaging means:
– Better extremes
– Better spatial coherence
– Far less “drizzle” problem
Example for precip,
1 Jan 1940
(P=0)
4. Run out of analogs for extreme days?
Existing methods:
1950-99……………………………………………………………………………….2070-99
Anomaly w.r.t. historical period (Tmin, Tmax)
LOCA:
1950-99……2000-2009….…2010-2039…..2040-2069……….2070-99
Anomaly w.r.t. 30-year climatology
Use LOCA to downscale changes in climatology
5. Averaging increases spatial coherence
precip (red = more coherent)
Project in association with Keith Dixon, GFDL
Evaluation:
Seasonal mean of daily
precipitation (mm/day) in
CCSM4
Error in %
Evaluation:
Seasonal mean of daily Tmax
(degC) in CCSM4
Error in degC
Evaluation:
Standard deviation of daily
precip (mm/day), averaged by
season
CCSM4
Error in %
Evaluation:
Standard deviation of daily
Tmax (degC), averaged by
season
CCSM4
Error in degC
Goal: Realistic daily extremes (Precip)
Winter,
mm/day
Summer,
mm/day
Goal: Realistic daily extremes (Temp)
Winter, C
Summer, C
Goal: Preserve model-predicted changes
(Precipitation, CCSM4, rcp 8.5, 2070-2100 minus 1950-1999)
Winter, %
Summer, %
Goal: Preserve model-predicted changes
(Tmax, CCSM4, rcp 8.5, 2070-2100 minus 1950-1999)
Winter, C
Summer, C
Example VIC
output
(water yr avgs)
Available variables:
Evapotranspiration
Snowpack
Humidity
Total runoff
Soil moisture
Black = with obs forcing
Green = 10 models
Red = model average
Slide 36
Summary of Production Runs
• 32 CMIP5 models
• Historical: 1950-2005. RCP 4.5 and
RCP 8.5: 2006-2100 (2099 some models)
• Climatological period: 1950-99
• Interpolated model calendars to
standard calendar w/leap days
• North America 24.5 N to 52.8 N at
1/16th degree resolution
• Daily Tmin, Tmax, Precip (specific
humidity? 23 models).
ACCESS1-0
ACCESS1-3
CCSM4
CESM1-BGC
CESM1-CAM5
CMCC-CM
CMCC-CMS
CNRM-CM5
CSIRO-Mk3-6-0
CanESM2
EC-EARTH
FGOALS-g2
GFDL-CM3
GFDL-ESM2G
GFDL-ESM2M
GISS-E2-H
GISS-E2-R
HadGEM2-AO
HadGEM2-CC
HadGEM2-ES
IPSL-CM5A-LR
IPSL-CM5A-MR
MIROC-ESM
MIROC-ESMCHEM
MIROC5
MPI-ESM-LR
MPI-ESM-MR
MRI-CGCM3
NorESM1-M
bcc-csm1-1
bcc-csm1-1-m
inmcm4
Summary
•
•
Many bias correction & downscaling schemes…
Quantile mapping, BCCA:
–
–
–
–
–
•
Muted extremes
Different biases at different frequencies
Too much spatial coherence
Drizzle problems
Wrong temperature of precipitation
New bias correction and LOCA downscaling
–
–
–
–
–
–
Extremes preserved pretty well, along with seasonal means and std deviations
Reasonable preservation of original model-predicted changes
Frequency dependent bias correction
Spatial coherence not degraded as much
Greatly reduces drizzle problem
Bias correct temperature conditional on precipitation
Pierce, D. W., D. R. Cayan, and B. L. Thrasher, 2014: Statistical downscaling using
Localized Constructed Analogs (LOCA). Journal of Hydrometeorology, v. 15, page 25582585
Analysis plots: loca.ucsd.edu
39