The ENSEMBLES high-resolution gridded daily observed dataset

1
The ENSEMBLES high-resolution gridded daily
observed dataset

• Malcolm Haylock, Phil Jones, Climatic Research
  Unit, UK
• WP5.1 team KNMI, MeteoSwiss, Oxford University

2
Outline

• The gridded dataset who, why, when and what?
• The station network
• Interpolation
• method comparison
• two-step interpolation of monthly and daily data
• kriging and extremes
• Point vs interpolated extremes
• implication for RCM validation
• Uncertainty
• Then finally some analyses comparing with GCM
  simulations from RT2B with ERA-40 forcing

3
The dataset

• Who
• Four groups in WP5.1
• KNMI data gathering and data quality and
  homogenisation
• MeteoSwiss homogeneity of temperature data
• UEA and Oxford interpolation
• Why
• Validation of RCMs
• Climate change studies
• Impacts models
• Many data providers do not allow distribution of
  station data

4
The dataset

• When
• Daily 1950-2006
• Available now from ENSEMBLES web site plus ECAD
• Two papers submitted to JGR, one on the
  comparison of methods, and one on the final
  gridded dataset with the chosen methods , which
  differ by variable
• What
• Five variables
• precipitation
• mean, minimum and maximum temperature
• mean sea level pressure (early 2008)
• Europe
• 0.250 and 0.50 CRU grids
• common RCM rotated-pole grid0.220 and 0.440
  rotated pole (-162.00, 39.250)

5
No. of stations
6
Precipitation Stations
2050
7
Tmean Stations
1231
8
Interpolation

• Need to match observations to model grid for
  direct comparison
• Therefore need to estimate observations at
  unsampled locations
• Compare several methods to find most accurate at
  reproducing observations in a cross validation
  exercise see more in Nynke Hofstras
  presentation tomorrow
• Largest QC problem is that date of observations
  do not match day is day when values occurred,
  but sometimes it is day when measured

9
Interpolation Methods

• Natural neighbour interpolation
• Angular distance weighting
• Thin-plate splines
• 2-D and 3-D
• Kriging
• 2-D and 3-D
• 4-D Regression
• lat, lon, elevation and distance to coast
• Conditional Interpolation important for
  precipitation

10
Stochastic or Deterministic

• Stochastic
• assumes that an interpolated surface is just one
  of many, all of which could produce the
  observations
• models the data with a statistical distribution
  to determine the expected mean at unsampled
  locations
• probabilistic model allows uncertainty estimates
• Deterministic
• assumes only one possible interpolated surface
• adopts a particular geographical model
• e.g. bilinear, inverse distance, Thiessen polygons

11
Cross Validation

• . For each station, interpolate to that station
  using its neighbours and compare with the
  observed value.
• Repeat for all days.
• Do for monthly averages and daily anomalies

Daily precipitation ( of monthly total)
compound relative error (cre) rms / s critical
success index (csi) hits/(false
alarmhitsmisses)
12
Cross Validation
Daily pressure (anomaly from monthly mean)
precip 2-D kriging with separate occurrence
model pressure 2-D kriging Tmean, Tmin, Tmax
3-D kriging
Daily Tmean (anomaly from monthly mean)
13
Interpolation methodology

• Grid monthly means using 2-D (pressure) and 3-D
  (temp and precipitation) thin-plate splines
• Determined to be the best method using cross
  validation
• Grid daily anomalies using kriging
• Combine the interpolated monthly means and the
  interpolated anomalies as well as their
  uncertainty
• Create a high resolution master grid (10km
  rotated-pole grid) and do area averaging to
  create different coarser resolution products.

14
Kriging and extremes

• Kriging estimates the mean and variance of the
  distribution at unsampled locations
• The best guess is the mean but extremes are
  usually a combination of a high local signal
  superimposed on a high background state
• Therefore kriging will tend to underestimate
  extremes and produce results similar to the area
  mean

15
Precipitation interpolation extremes reduction
factor
2yr
5yr
10yr
50
75
90
95
99
16
Tmax interpolation extremes -reduction in anomaly
50
75
90
95
99
2yr
5yr
10yr
17
Gridded Extremes
Precipitation Extremes
Extremes of Gridded
precipitation 10-year return period
18
Uncertainty

• Interpolation uncertainty only

19
(No Transcript)
20
Conclusions

• We have created a European daily dataset very
  much improved over previous products, with a
  detailed comparison of interpolation methods
• Kriging gives the best estimate of a point
  source, but when the interpolated grid (25km) is
  smaller than the average separation (45km for
  precipitation), the interpolated point will be
  more an area average
• Therefore validation of RCMs using the gridded
  data assumes the RCMs represent area-averages
• Kriging can be extended to produce more realistic
  simulations of point precipitation at unsampled
  locations, with a better estimate of uncertainty
  of the extremes, but this is computationally very
  expensive

21
ENSEMBLES WP5.4 and ETCCDI Meeting KNMI De Bilt
13-16 May 2008

• Extremes of temperature and precipitation as seen
  in the daily gridded datasets for surface climate
  variables (D5.18 Haylock et al.) and in the RCM
  model output from the (RT3) 40-year experiments
  driven by ERA-40 reanalysis data
• Phil Jones and David Lister Climatic Research
  Unit

22
Gridded Data

• Available on ENSEMBLES web site
• Two papers submitted to JGR
• One on a comparison of gridding techniques
  (Hofstra et al.)
• One on the final gridded dataset (Haylock et al.)
• A simple comparison shown here

23
The location and period of coverage of
station-series which went into the
interpolation/gridding exercise
24
Extreme Measures

• Trends of mean maximum and minimum temperatures
• Trends of 5th percentile of Tn
• Trends of 95th percentile of Tx
• Compare gridded trends with station trends
• Trends patterns over various periods

25
Testing of extreme values in a fairly flat part
of the region covered by the observed grids
Lubny, Ukraine
26
As earlier, but JJA
27
Trends (C/decade) in the (gridded/observed) 05th
percentile Tmin. series 1950-2006
28
Trends (C/decade) in the CRU 0.5 grids (CRU
TS3.0) Tmin. series 1950-2006
29
Trends (C/decade) in the (gridded/observed) 95th
percentile Tmax. series 1950-2006
30
Trends (C/decade) in the (gridded/observed) 05th
percentile Tmin. series 1961-2006
31
Trends (C/decade) in the (gridded/observed) 95th
percentile Tmax. series 1961-2006
32
Tn05 histogram of differences compared to
gridded observations
33
Tx95 histogram of differences compared to
gridded observations
34
(No Transcript)
35
(No Transcript)
36
(No Transcript)
37
(No Transcript)
38
(No Transcript)
39
(No Transcript)
40
(No Transcript)
41
(No Transcript)