Meteorol. Z, 22, 2013 
N.H. Schade et al.: Regional Evaluation of ERA-40 Reanalysis Data 
679 
Downloaded from www.schweizerbart.de 
Unauthorized distribution of this copyrighted material is strictly forbidden! 
sets presented in the following. Nevertheless, they surely 
point out the need to correct GZS data for climate 
research studies accordingly, if they want to be compared 
with the regularly sampled ERA-40 (or any reanalyses) 
data for that matter. 
Finally, the uncertain fraction variability has to be 
addressed. Since we have to assume, that observational 
errors occur, we have to account for the representative 
ness of GZS. Stoffelen (1998) used a triple-collocation 
method to validate anemometer, scatterometer and NCEP 
winds. This can not be done with only two datasets avail 
able, unless one is divided into subsamples, which would 
not be practical for our analyses. Therefore, we have to 
assume the GZS temperature und sea level pressure 
errors by what we know. This is, as stated above, the ran 
dom measurement error (KENT and BERRY, 2005), which 
is used as “expected variance” a 2 . Secondly, the “error 
variances” e 2 are computed as variance of the GZS val 
ues of each time step per box. Now, an error estimate, or 
pseudobias, can be obtained that is the difference 
between the GZS value x and the biased mean 
<y>- xct 2 / (a 2 + e 2 ). This pseudobias is small if e 2 
is small, which means the variance of measured GZS val 
ues is small. Mean pseudobiases of 1.4 ± 1.9 K (Box 1) 
and 1.7 ± 2.5 K (Box 2) for air temperatures can be 
found, with minimum values of about 1 = 0.5 K in win 
ter (December) and maximum values of about 3 = 1.2 K 
in summer (August). Sea level pressure biases are found 
to be 1.9 ± 2.1 hPa (Box 1) and 2.4 ± 2.9 hPa (Box 2). 
Nevertheless, to correct the GZS data this way would 
imply to do so for every ship, variable and measurement 
device separately, which is not possible in the context of 
this work. Additional information about e.g. the shielding 
of thermometers, the exposure times, etc. would have to 
be investigated for every single measurement for the 
whole time period. Often this data not available at all 
and to correct only a part of the data and assume that this 
correction applies to the rest as well would introduce 
biases also. Instead we decided to use the “raw” GZS 
data after they passed HQC and the manual inspection 
for box 1 und 2 and to keep in mind that measurement 
errors and pseudobiases exist and to account for them 
in the discussion. As mentioned above, the sampling bias 
does not effect the differences between both datasets, if 
we compare GZS-like sampled ERA-40 with GZS. 
3 Results 
The distributions of air temperatures at 2 m height above 
ground (AT) are presented in Fig. 3 through box plots. 
They show the percentile values of both ERA-40 and 
observed ATs for all months of the reference period 
1961-2000, and for the winter season (DJF) only. The 
upper row includes both grid boxes, the lower row shows 
the four subgrid boxes of Box 2 (see Fig. 2). For the 
overall data, ERA-40 (grey boxes) and observations 
(black boxes) are very close together and the differences 
in the respective percentiles are only marginal. For DJF 
however, differences in Box 2 are obvious. The median 
value of the ERA-40 data is 0.6 K below the respective 
one of the observations and the 99% percentile differs 
by 1 K. The percentiles 1 and 5 are in better agreement. 
A closer look at the four sub grid boxes of Box 2 reveals 
an interesting feature: Both eastern boxes facing to the 
Danish coast (even numbers) indicate large differences 
in the ERA-40 temperatures in the sea side part, com 
pared to the GSZ, as well as to ERA-40 itself. Differ 
ences are about 1-2 K in the respective higher (wami) 
percentiles and the median, whereas the lower (cold) per 
centiles differ less. This reflects the characteristic of the 
reanalysis model not to differentiate exactly between land 
and sea and a land-influence reaching far into the open 
sea areas. The western, sea-side boxes (odd numbers) 
are much closer together, but show also colder maximum 
temperatures compared to GZS. Again, this can be 
observed only in DJF, all other periods are in good agree 
ment. Even, if measurement errors and biases are 
accounted for, this would not explain the differences 
between the ERA-40 values in the eastern boxes to those 
in western boxes. 
However, these differences are only apparent for the 
AT investigations. Comparisons of the SLP data show 
no such differences between ERA-40 and the observa 
tions (not displayed), neither in the median nor in the 
percentile values. This leads to the conclusion that the 
ERA-40 land-facing grid points are influenced by colder 
temperatures occurring over land in the winter months 
DJF, whereas the pressure fields are independent of the 
subsurface and land-sea effects can be disregarded. This 
independency of the land-sea distribution leads to a better 
consistency between observations and reanalysis. The 
assumption is supported by the results presented in 
Fig. 4: All ERA-40 monthly means of August and 
December during the period 1961-2000 for Box 2 are 
plotted against the respective means basing on observa 
tions of AT (left graph) and SLP (right graph). These 
months show the best and worst fitting data pairs. While 
in August all AT pairs are relatively close around the 
optimum accordance line, despite the high uncertain frac 
tion variability, all ERA-40 mean ATs in December are 
systematically smaller than the respective means basing 
on observations, with a maximum difference of 
2.2 = 0.2 K. This is about twice the size of the monthly 
mean random measurement error of 1.2 ± 0.3 K found 
by Kent and Berry (2005). Therefore, the colder 
ERA-40 temperatures can not result from measurement 
errors in the observations alone, especially in December 
where the pseudobias due to the uncertain fraction vari 
ability is low. In fact, measurement errors resulting from 
radiative heating should have a greater effect in the sum 
mer months. Here, ERA-40 means are smaller during the 
whole span from October to April (not shown), until in 
May the mean values become closer to the optimum 
accordance line and begin to differ again in September.