BENCHMARK DATASET HOMOGENIZATION. A first test.

BENCHMARK DATASET HOMOGENIZATION. A first test.

Contribution to the Working Groups Meeting of COST-HOME action, Tarragona, 9-11 March 2009

Enric Aguilar, Center on Climate Change, URV, Tarragona.

C3-SNHT APPLICATION New software (F95)

Detection (done) and taylored correction modules (almost done)

Reference selection (distance, correlation, overlap)

One homogenization for station: for each station/element/month-season-annual a special “network” is created and only the results for the main candidate are retained

Series are split until the most recent significant break is found and the remaining series is tested until no more breaks are found or the segment is too small

Breaks need to be inspected and a correction pattern created. This is better done with annual and seasonal averages (problem with benchmark; run over anomalies?)

R-code for output inspection

WORK IN PROGRESS:

feed breaks to HOM; use weather types to calculate daily factors

Network TEMP/SUR1/000001

VISUAL INSPECTIONSTATISTICAL DETECTION

BREAKS VALIDATION

VISUAL INSPECTION

Saint-Cornier (06): large break around 1959 & missing 1945

Chartres (03): large break around 1950

Saint Georges (01): break around 1970 & missing 1945

Groix (02) : break around 1955 & and missing 1945

Rennes (04): break around 1985 & missing 1940

STATISTICAL DETECTION

Monthly results are confusing and noisy

Seasonal are better (only 4!)

Annual, optimum to start

PLOTS FOR BREAK VALIDATION

HOMOGENIZATION FOR SAINT-CORNIER Data series for main candidate and references. Red lines indicate breaks


HOMOGENIZATION FOR SAINT-CORNIER Zeta-series for main candidate and references. Red lines indicate breaks


HOMOGENIZATION FOR SAINT-CORNIER Data and Zeta-series for main candidate


HOMOGENIZATION FOR SAINT-CORNIER test-values and break location


HOMOGENIZATION FOR SAINT-CORNIER Histogram of Detected breaks, in all homogenizations where this station has participated


HOMOGENIZATION FOR SAINT-CORNIER Detected breaks, drawn proportional to the largest one

SAINT-GEORGES

Large break in 1948

Smaller breaks after 1970

Pattern: BREAK 1948 TREND 1948-1981 or

BREAK 1970

GROIX Annual detects nothing

Z-plots point to 1950-1960

Histogram plot indicates 1949 and 1955 or 1956

They're detected in summer

Pattern:

BREAK 1949

BREAK 1956

GROIX Question: is my decision

influenced by the existence of larger breaks?

Plot 1 = no

Plot 2 = maybe!

CHARTRES

Looking at annual + Spring (plotted)

BREAK 1980 BREAK 1948 (Clear) BREAK 1926 BREAK 1919 (?)

CHARTRES

SUMMER: non detected outlier (level is 3.5 sd) going in the oposite direction of a potential break: no detection!

RENNES

ANNUAL: BREAK 1990 BREAK 1951

OTHER POSSIBLE BREAKS:

1919 & 1926 ... (remember last slide?) Influenced by others or margin effect?

~ 1979 -1984 ==> forced by outlier

GIEVRES

Breaks are small

BREAK 1985

BREAK 1943

MAYBE TREND 1943-END?

SAINT CORNIER

BREAKS 1994

BREAK 1959

BREAK 1942 (missing?)

BREAK 1928 (?)

Annual

Summer: detection around 1935, but seems to be ~ 1928

ILE-YEU Annual does not

detect breaks, but there is an evident (short) problem between 1931-1934 (2 breaks?)

Winter (shown bottom pannel) and Spring identify:

BREAK 1963

LA MOTE-ACHARD

Breaks have small amplitude. Visual inspecetion of z-series suggests:

TREND 1951-1980?

BREAK 1951

BREAK 1939

BIARD

BREAK ~ 1970?

BREAK 1943

MANY OUTLIERS FOUND

BENCHMARK DATASET HOMOGENIZATION. A first test.

Documents