1 Brendan Rogers. 2 SizePlacement of important currents.

1

Brendan Rogers

2

Size Placement of important currents

3

1)Change factor, perturbation, or delta method (my data)

1)Regression methods (Gachon & Dibike, 2007)

1)Regional Climate Models (Pan et al., 2001)

1) Intercomparison/summary (Fowler et al., 2007)

4

1) Calculate GCM-simulated historical average

2) Calculate anomalies between future GCM and (1)

3) Apply anomalies to observed (e.g. PRISM) average

Pros-rapidly applied to several GCMs/regions-relatively simple to implement

Cons-assumes constant bias through time-assumes spatial pattern of climate will remain the same-assumes anomalies are the same regardless of elevation, continental position, etc.-generally ignores the change in variability(?) or extreme events, just scaling mean, max, and min of variables-can’t do local feedback effects

“…sophisticated downscaling methods, despite the physically based concepts, are not yet completely satisfactory and they do not appear to be a good practical alternative to the perturbation method.” (Prudhomme et al., 2002)

03/09/11

What is Climate Mapping?What is Climate Mapping?The process of interpolating climate statistics at irregularly-

spaced station locations to a regular grid

“Geospatial Climatology” The study of the spatial patterns of climate on the earth’s

surface and their causes

03/09/11

- Generates gridded estimates of Generates gridded estimates of climatic parameters (e.g., P, T, DP)climatic parameters (e.g., P, T, DP)

- Moving-window regression of Moving-window regression of climate/weather values vs. climate/weather values vs. elevation or climatology/radar for elevation or climatology/radar for each grid celleach grid cell- Uses nearby station observationsUses nearby station observations

- Spatial climate knowledge base Spatial climate knowledge base weights stationsweights stations in the regression in the regression function by their physiographic function by their physiographic similarity to the target grid cellsimilarity to the target grid cell

PRISMParameter-elevation Regressions on Independent Slopes Model

03/09/11

Prism Interpolated Weather Data

03/09/11

Products Guided by PRISM ClimatologiesNew USDA Plant Hardiness Zone MapNew USDA Plant Hardiness Zone Map

9

10

May Average, 1971-2000 May, 2058

13.310.8

11

2.5

12

13

Observed May 1971-2000 average tmp Hadley May 2058 tmp(observed + interpolated anomalies)

14

May, 2058

15

-study investigates regression-based SDSM model in its ability to simulate changes in mean temperature values, as well as extreme temperaturess, in northern Canada.

-used 4 sets of climate predictors: A2 and B2 from CGCM2 and HadCM3

-SDSM was able to capture the temperature change signal, with plausible higher warming in winter than summer and in A2 vs. B2

16

17

Overview-Uses multiple regression to select from a list of potentially correlated variables at regional scales-Calibrates the regression model(s) against the observed data, using GCM output-Applies the model(s) to future GCM data

Predictors to use-should have understandable physical relationship to the predictands-should avoid using GCM variables known to be inaccurate (here, surface temperatures)

-GCMs are thought to give a more realistic description of the free tropospheric variables and large-scale circulation than of surface parameters-in northern Canada, the biases of 2-m GCM simulated temperature over inland or arctic seas can reach more than 12˚C in winter and fall - due to poor simulation of timing/magnitude of sea-ice extent and thickness

-they used: 500 and 800 hPa geopotential heights, vorticity, zonal velocity component, specific humidity. The authors pride themselves on using specific humidity, because it represents a relatively independent predictor from the others to result in better accuracy.

18

-PROs-computationally inexpensive and thus can be applied to different data.-good for heterogeneous environments, like island or land/sea contexts, where there are strong relationships to synoptic-scale forcings

CONs-assumes that statistical relationships developed for present day climate also hold in future.-Climate change proceeds on much longer time scales than those to which SD method is fit.-can’t do local feedbacks not captured by GCM

19

20

Downscaled Raw GCM

21

Downscaled tmax extremes

22

-using upper atmosphere predictors, the SD method produces temperature values more physically consistent with the current conditions and what can be expected in the future

-the values are also more seasonally consistent and reduce inter-model differences

-winters will warm more than summers

-regression method is inherently conservative in the presence of non-stationarity and will underestimate the warming signal (not sure how they know that)

-the local scale feedback of sea-ice retreat is not dynamically simulated but taken from the broader GCM, so it is missed here

-they make the argument that SD passes the stationarity test because the same predictors and predictands (mostly) work in all stations for present day (substituting space for time)

23

-Two RCMs forced by 3 sets of initial boundary conditions for 10 years for US at 50 km resolution.

-Three boundary conditions: ‘reanalysis’, GCM current climate, future scenario

-RCMs do well in simulating orographic precipitation, E-W transcontinental gradients, and annual cycles. However, they miss the cool-season precipitation in the lower Mississippi River basin.

24

-Two RCMs: RegCM2, HIRHAM.-RegCM2:

-18 categories of land use, 12 soil types-3 soil layers: top layer (10cm), root zone(1-2 m, depending on veg type), deep soil (10m).-14 vertical atmospheric layers based on pressure levels.-rainwater is not a predictive variable, so water rains out immediately with no re-evaporation

-GCM used = HADCM2-assumed 1% per year increase in GHGs after 1990-aerosol effects not included-simulation window corresponds to 2040-2049 with 480 ppm-reanalysis (observations) came in at 28 pressure levels for 1979-1988

25

26

27

RCM performance bias: difference between reanalysis-driven RCM simulation and observations (depends on parameterization details, calibration, etc., models’ systematic errors and drift, and errors in re-analysis data)

Boundary forcing bias: difference between RCM run driven by GCM current and by reanalysis (how the GCM is doing for the current climate)

Intermodel bias: difference between runs from different RCMs, both driven by reanalysis data

GCM-RCM downscaling bias: difference between GCM current and corresponding RCM driven by GCM output (includes potential improvement offered by RCM, as well as RCM errors)

28

29

30

31

Climate Change bias

32

-They used t-tests of separate means and variances to compute the statistical significance of climate change (2x10 years of simulations for each model). In most of the continental U.S., the t-value is larger than 6 (1.73 significant at 95% confidence level). It was only not significant in Texas/New Mexico, Northern Minnesota. Annual relative increase is 14-56% depending on the region. The dry west has larger increases in both relative and absolute magnitudes

-GCM-RCM downscaling bias is by far the largest among all biases, implying a potential for RCMs to improve GCM simulations.

-Bias analysis shows: 1) climate change is substantially larger than largest biases in several seasons and regions, 2) summer ratios of climate change vs. bias are always less than 1, and 3) the climate change:bias ratio is especially large in California.

-in SE, RCM performance bias is relatively small, indicating that the GCM needs to provide better boundary conditions

-West coast inter-model bias is largest in winter, suggesting that RCM interactions between onshore flow and topography need improvement.

33

Dynamical Downscaling

-uses regional climate models (RCMs) or limited-area models (LAMs)

-typically resolved at ~.5 degree latitude and longitude scale

-parameterize physical atmospheric processes

-can realistically simulate regional climate features such as orographic precipitation, extreme events, and regional scale anomalies, or non-linear effects such as El Nino

-model skill depends strongly on biases inherited from driving GCM, presence/strength of regional scale forcings like orography, land-sea contrast, vegetation cover, and particular RCM

-places with large topographic features (western US, Europe, New Zealand) report more skillful dynamical downscaling than places like Great Plains b/c regional forcings are stronger

-variability in internal parameterizations also provides uncertainty (important and bad ones can include large-scale condensation and convection schemes, soil parameterization, lack of moisture advection into region, snow-albedo feedbacks)

34

Dynamical Downscaling

-most RCMs are computationally expensive and only run for a few decades at a time – they will pick slices of time and just use ‘pattern scaling’: changes are scaled according to temperature signal modeled for intervening period, assuming linear rate of change

-using an RCM can also add uncertainty to that inherent in the GCM output:-for temperature projections, uncertainty introduced by RCM is less than that from emissions scenario, but opposite is true for precipitation.-GCM output variability is generally larger than RCM variability

-studies (Leung et al, 2004) have stated RCMs provide added value because the signals from RCMs can be significantly different from GCMs because of orographic forcing and rain-shadow effects. RCMs can also provide improved simulation of meso-scale precipitation processes

-RCM model inter-comparisons are becoming more common.

35

Statistical Downscaling

-all assume regional climates are a function of the large-scale atmospheric state

-all use R = F(X), where R is local climate variable, X is set of large-scale climate variables, and F is function relating the two. F is typically derived by calibration using point observations or gridded reanalysis data.

-validate using correlation coefficients, root mean square error, or explained variance (possibly low frequency variability important for climate change)

-the trick is what set of variables, X, to use to predict R. This may not be easy and may not be stationary in time (ex. Humidity, pressure, to predict precip). Geopotential heights and specific humidity are found to be useful in all locations and seasons

-choice of predictor domain also important (spatial/temporal extent of GCM outside of study area)

36

Statistical Downscaling

-statistical methods are more straightforward than dynamical downscaling but tend to poorly represent extreme events and underestimate variance because only part of the regional and local climate variability is related to large scale variations

-There are approaches to add variability: variable inflation and randomization-variable inflation: increases variability by multiplying by a suitable factor-randomization is variable inflation but with an added in the form of ‘white noise’ (found to give good results in European surface temperatures)

37