Advances in Statistical Downscaling of Meteorological Data: Development, Validation and Applications John Abatzoglou University of Idaho Department t of Geography EPSCoR Western Tri-State Consortium 7 April, 2010, Incline Village, NV
Desirable Qualities of Downscaling Resolution: Daily time resolution, high spatial resolution (<10km) Products: Multiple variables (not just T & P), physically based predictors Accuracy: Reproduction of high-resolution historical records Feasibility: Not too computationally burdensome (ability for ensembles) Synchronicity: Downscaled weather fields adhere to physical laws across variables and across spatial domains Theoretical: Retains core information from driving GCM; e.g., future synoptic patterns aren t the same as historical Practical: Tractable to observations (gridded/station) for immediate use in applied studies.
Downscaling for Applications Simple (Simple Change Factors) Region-wide changes in mean temperature Intermediate (Synthetic Statistical) Measures of Variability Advanced (Deterministic Statistical/Dynamical) Daily time-scales Multiple variables Ability to capture regional climate/wx patterns Complexity
Delta Method + False Realism?? Resolution Products Accuracy Feasibility Synchronicity y Theoretical Practical
Downscaling for Applications Simple (Simple Change Factors) Region-wide changes in mean temperature Intermediate (Synthetic Statistical) Measures of Variability Advanced (Deterministic Statistical/Dynamical) Daily time-scales Multiple variables Ability to capture regional climate/wx patterns Complexity
Bias Correction Spatial Downscaling (BCSD) (Wood et al, 2004; Maurer, 2007; Salathe et al, 2007) Bias Corrects GCM output Allows for GCM to determine time series Interpolation based method overlooks influence of complex terrain on weather and climate Resolution* Products* Limitations of monthly time scales Accuracy* and/or temporal disaggregation Feasibility Synchronicity y Theoretical Practical http://gdo-dcp.ucllnl.org/downscaled_cmip3_projections/dcpinterface.html
Downscaling for Applications Simple (Simple Change Factors) Region-wide changes in mean temperature Intermediate (Synthetic Statistical) Measures of Variability Advanced (Deterministic Statistical/Dynamical) Daily time-scales Multiple variables Ability to capture regional climate/wx patterns Complexity
Constructed Analogues (Hidalgo et al, 2008; Zorita and von Storch, 1999 ) Resolution Products Accuracy Feasibility Synchronicity Theoretical Practical Onus on GCM in simulating daily meteorology Concept Weather-typing approach based on spatial pattern matching of daily meteorological fields Relies on extensive library of fine and coarse scale observed patterns Identical : 10^30 years (Van den Dool, 1994) Single: best pattern match Constructed : many sampling opportunities (Hidalgo et al. 2008) Basis: model coarse scale fields from analogues and fine-scale will follow http://cascade.wr.usgs.gov/data/task1-climate/
Multivariate Adapted Constructed Analogues (MACA) Combines attributes of CA and BCSD (Maurer and Hidalgo, 2008, 2010) Input Data - Observations (fine + coarse) -GCM - Similar time period of obs (20C3M) - Future runs (SRES-A1B) Steps 1. Bias-correct GCM fields 2. Epoch adjustment: remove GCM difference (A1B/20C3M) + 3. Identify a subset of the 30 best patterns through spatial RMSE Constructed analog of patterns* 4. Adjusts end variables using quantile-mapping at fine-scale 5. Epoch adjustment: Introduce difference fields back +Epoch adjustment needed to account for disappearing analogs *Multivariate approach retains spatial covariance and covariance between variables
Common Gridspace GCMs interpolated to a common 2 degree grid space Fine-scale observations aggregated to common 2 degree grid space
Step I: Bias Correction of Daily Data Quantile-mapping of model to obs. Shift of hist. GCM applied to future GCM Processed on daily time-scales using a 45-day buffered window
Multivariate Adapted Constructed Analogues (MACA) Input Data - Observations (fine + coarse) -GCM - Similar time period of obs (20C3M) - Future runs (SRES-A1B) Steps 1. Bias-correct GCM fields 2. Epoch adjustment: remove GCM difference (A1B/20C3M) + 3. Identify a subset of the 30 best patterns through spatial RMSE Constructed analog of patterns* 4. Adjusts end variables using quantile-mapping at fine-scale 5. Epoch adjustment: Introduce difference fields back +Epoch adjustment needed to account for disappearing analogs *Multivariate approach retains spatial covariance and covariance between variables
Step II: Disappearing Analogues What about the day without an analog? e.g., July 4 th, 2080? (temperature extremes), Dec 30, 2080 (precip extremes) Solution: Epoch adjustment Removal of difference field prior to CA (scaling factors; even across CDF) Removal of difference field prior to CA (scaling factors; even across CDF) Reintroduction of difference fields after final step
Multivariate Adapted Constructed Analogues (MACA) Input Data - Observations (fine + coarse) -GCM - Similar time period of obs (20C3M) - Future runs (SRES-A1B) Steps 1. Bias-correct GCM fields 2. Epoch adjustment: remove GCM difference (A1B/20C3M) + 3. Identify a subset of the 30 best patterns through spatial RMSE Constructed analog of patterns* 4. Adjusts end variables using quantile-mapping at fine-scale 5. Epoch adjustment: Introduce difference fields back +Epoch adjustment needed to account for disappearing analogs *Multivariate approach retains spatial covariance and covariance between variables
Step III: Constructed Analogues Target (GCM) Library of Potential Analogs 30 Best Analogs From 45-day window of target date Constru ucted Anal log Coarse Res. n 30 n Z n og. Constru ucted Anal Downsc caled Res n 30 n Y n After Hidalgo et al. 2008 But applied on absolute values, not anomalies
Multivariate Analogs Joint analog search performed for grouped variables by physical descriptors Predictor Variables Downscaled Fields Maximum Temperature Minimum Temperature Specific Humidity TMAX TMIN RHMAX RHMIN MSLP PPT PPT U V
Multivariate Adapted Constructed Analogues (MACA) Input Data - Observations (fine + coarse) -GCM - Similar time period of obs (20C3M) - Future runs (SRES-A1B) Steps 1. Bias-correct GCM fields 2. Epoch adjustment: remove GCM difference (A1B/20C3M) + 3. Identify a subset of the 30 best patterns through spatial RMSE Constructed analog of patterns* 4. Adjusts end variables using quantile-mapping at fine-scale 5. Epoch adjustment: Introduce difference fields back +Epoch adjustment needed to account for disappearing analogs *Multivariate approach retains spatial covariance and covariance between variables
Validation of MACA Demonstration of skill needed for value-added downscaling NCEP/NCAR reanalysis as surrogate GCM Cross Validation 1979-2008 Compare skill to interpolation-only based approach Variables Maximum Temperature Minimum Temperature Maximum Relative Humidity Minimum Relative Humidity Daily Accumulated Precipitation Daily Average Wind Speed Time Period All Year Cool Season (Nov-Apr) Warm Season (May-Oct)
Validation: Daily Data MACA Interpolation r-value
Validation: Daily Data RMSE ( C)
Validation: Daily Data TMAX RHMIN PPT WS mm mm
Validation: Multivariate Data 22 Oct 2007 TMAX anom (shaded), Wind Vectors and area of RHMIN Anom below -15% (dashed line) Monthly Correlation of Accumulated Snowfall OBS DOWN
MACA Resolution: Daily, 8-km Products: TMAX, TMIN, RHMAX, RHMIN, 10-m U & V and PPT Accuracy: See validation figs Feasibility: Yes Synchronicity: Simulates integrated response across variables Theoretical: Driven by daily synoptic wx from GCMs Practical: Direct applicability to raster or station data Data Availability Time Slices: 1971-2000; 2046-2065; 2080-2099 Scenarios: SRES-A1B Models: 13 CMIP AR4 models
Th Additional Thanks to: Von Walden 1 Brandon Moore 1 Crystal A. Kolden 1,2 Timothy J. Brown 2 1 University of Idaho, Moscow, Idaho 2 Desert Research Institute, Reno, Nevada John Abatzoglou John Abatzoglou jabatzoglou@uidaho.edu
Applications Wildland Fire Danger and Proactive Fuel Treatments (T. Brown, DRI; C. Kolden, UI) Hydrologic sensitivity to climate change and forest management practices (E. Du, T. Link, UI) Salmonid habitat changes (S. Gillis, UI)
Validation: Monthly Data MACA Interpolation ti r-value
Statistical Downscaling Resolve sub-synoptic features using statistical relationships between observed coarse-scale (the predictors) and local (the predictand) climate. Advantages Easy to do Computationally cheap Ability for ensembles Ability to downscale to stations Results comparable to obs. Disadvantages Easy to do poorly Often violates physical processes Assumes stationary relationships Relies on obs. for guidance