SIXTH SEMINAR FOR HOMOGENIZATION AND QUALITY CONTROL IN CLIMATOLOGICAL DATABASES AND COST ES0601 HOM

1
Different approaches for the homogenisation of
the Spanish Daily Temperature Series (SDATS)

• Aguilar, E., Brunet, M., Sigró, J.
• Climate Change Research Group, Universitat Rovira
  i Virgili, Tarragona, Spain

2
MOTIVATION

• SDATS dataset included only the longest and most
  reliable series, leading to a low density
  network
• CCRG is involved in a coordinated project
  (EXPICA) that wants to relate temperature and
  precipitation extrems to circulation patterns
  over the Iberian Peninsula
• Can our current homogenization procedure for
  daily data feed temperatures to EXPICA?
• Can we apply other procedures with the current
  network? (i.e. HOM)
• Do we have to expand it?
• CAFIDEXPI subproject ? re-homogenization on a
  daily bases of SDATS and calculation of extreme
  indices

3
Spanish Daily Temperature Series

• 22 Stations
• Unevenly distributed across Spain

4
HOMOGENIZATION STEPS
QCd daily data of TMax and TMin
Calculation of Monthly Values of TMax and TMin
Screen Bias Minimisation over monthly series of
TMax and TMin
Blind break-point detection over annual, seasonal
TMax, Tmin, Tmean with automated SNHT (1997)
Breakpoint validation (metadata, plot checks, )
Generation of correction pattern
Application to monthly Tmax and Tmin (As
described in Aguilar et al, 2002)
Monthly, Seasonal, Annual Tmax, Tmin, DTR, TMean
Series (STS)
Validation of daily corrected values
SDTS
Interpolation to daily data (Vincent et al., 2002)
5
SCREEN BIAS MINIMIZATION
Large effect on TMax
Much smaller effect on TMin
CCRGs SCREEN project (CICYT) ? 2 replicas of
Montsouris Screen, on operation since 2003
6
SCREEN BIAS MINIMIZATION
New Estimation (Murcia) TMaxStev -0.508
TMaxMont0.975
7
The homogenization methods. SNHT
Automated Software by Enric Aguilar. Available
under request
8
INTERPOLATION TO DAILY DATA
9
THE HOM METHOD CONCEPT

• 1) DEFINE HSPs for the candidates and reference
  stations
• 2) Identify highly correlated ref station that
  overlaps HSP1 and HSP2 of the reference
• 3) Model (LOESS) the relations in HSP1
• 4) Predict the temperature at the candidate in
  HSP2 using observations from the reference series
  in HSP2
• 5) Create a paired difference between predicted
  and observed temperatures in HSP2
• 6) Find the probability distribution (L-Moments,
  6 distributions) of the candidate in HSP1 and
  HSP2
• 7) Bin each difference in 5) according to the
  associated predicted temperature according the
  distribution of HSP1
• 8) Fit a smoothly varying function between the
  binned differences to obtain adjustments for each
  percentile
• 9) Using the probability distribution of the
  candidate in HSP2 , determine the percentile of
  each observation and adjust accordingly to the
  value obtained in 8)

10
(No Transcript)
11
PRELIMINARY APPLICATION OF HOM METHOD TO LA
CORUÑA, MADRID, MURCIA

• We compare the results obtained with CCRG
  procedure with the HOM method
• HOM is applied to raw data (with no screen
  adjustments) using the breakpoints detected
  through the CCRGs procedure.
• We use 3 series Madrid, Murcia and La Coruña,
  analyzing the impacts of the different approaches
  over annual trends in TMIN and TMAX and on four
  extreme indices warm days (TX90p) cold days
  (TX10p), warm nights (TN90p) and cold nights
  (TN10p

12
LA CORUÑA

• The method cannot be applied to this station with
  the current dataset
• Correlations with other series are too low
• Best candidates do not have overlapping HSPs. For
  example, San Sebastian
• Introduction of new stations (Gijón, Oviedo,
  shorter Galician stations) should improve this
  situation

13
MADRID

• Changes in screen around 1893 ? can HOM capture
  this kind of problems?
• Artificial trend (urban) between 1893 and 1960 ?
  this can be a problem for HOM, as were modelling
  HSPs and 1893-1960 wont be exactly an HSP. To
  try to tackle this we are using to schemes for
  Madrid
• 1893,1960
• -1893, 1920,1940 (understanding the urban trend
  as a succession of same sign shifts)
• Jump in 1960

14
Black raw Red CCRG Blue HOM-1break Green
HOM-3breaks
15
Model and CDF. Inhomogeneity in 1893.
HOM-1break. TMAX. August.
Larger values are evident in HSP2 (pre-1893)
represented by dashed lines. The adjustments
capture this jump
16
Model and CDF. Inhomogeneity in 1893.
HOM-1break. TMAX. April
Change in variance and in mean. Lower percentiles
need more correction than upper percentiles. Is
this what we should expect from the source of
inhomogeneity we know (i.e. change in screen)?

17
SOMETHING IVE HIDDING FROM YOU!

• Reference chosen among the available stations
  with a reasonable number of pairs and a
  reasonable correlation
• Reference for April is Badajoz
• Reference for August is Cádiz (!)
• This is far from optimum there is little chance
  to find closer neighbors for this part of the
  record

18
Trends for annual TMAX compared to trends from
CCRG original approach (bold italic, different
sign of point estimate bold different sign in
the confidence interval)
19
Same for TX90p
20
Same for TX10p
21
MURCIA

• Murcia presents a change in SCREEN around 1912
• And relocations
• 1939
• 1954
• 1984

22
Annual values derived from daily homogenized
data. Black lines original data red lines CCRG
procedure (correcting change of screen in 1912
and relocations in 1939, 1954 and 1984) green
lines HOM adjustments using 1863-1912 1913-1939
1940-1954 and 1955-2006 as HSPs. Notice the
excellent agreement between methods in the
highlithed area of the plot
23
ADJUSTMENTS FOR MURCIA. Break 1984. May (USING
ALICANTE, now this is good!!)
Wide range of adjustments from slightly negative
to about 1ºC in the higher percentiles
24
Histograms of differences between CCRG
adjustments and ORIGinal data (left) HOM
adjustments and ORIginal data (center) and CCRG
and HOM adjustments (right) for different months
(rows). Due the nature of the two sets of
adjustments, notice a largest gamma of adjustment
values when HOM is implied in the differencing.
The pairs of series, show significant changes in
variance.
25
CONCLUSIONS AND FUTURE WORK

• There is a strong consensus about the need of
  improving the homogenization of climatological
  time series, specially on daily and sub-daily
  scales The CCRG has been homogenizing daily
  values using an effective combination of an
  adapted version of SNHT interpolation of
  monthly factors to daily values
• The HOM method provides a powerful tool to adjust
  daily datasets accounting for Higher Order
  Moments inhomogeneities
• Although HOM method and CCRG procedures can show
  very similar adjustments when annual values are
  re-computed from homogenized daily values, in
  some ocasions adjustments can show large
  differences. This differences enlarged when
  seasonal or monthly series are analyzed, can be
  partially attributed to the lack of good
  references to produces overlapping HSPs or in
  other cases to non identified breakpoints. But
  they could also derive from the larger range of
  corrections applied to daily values for each
  month
• In the near future, several projects by the CCRG
  specially the CAFIDEXPI (Changes in Frequency
  Intensity and Duration of EXtremes in the Iberian
  Peninsula) and CLICAL - will introduce new series
  to SDATS for the compilation of a new version of.
  The application HOM method when applicable
  will continue to be explored.