15  May  2014

DANGENDORF  ET  AL.

3591

Fig.  6.  Reconstruction  of  surges  based  on  20CRv2  winds  and
SLP.  The  10-yr  moving  averages  of  the  99.9th  percentiles  of  observed ­
  surges  ±  standard  deviations  (black  line  with  gray  shaded
area;  the  standard  deviation  has  been  computed  as  a  measure  of
variability  over  each  10-yr  window)  and  their  reconstructions  based
on  20CRv2  (light  red:  individual  ensemble  members;  dark  red:
ensemble  mean;  both  normalized  to  a  common  period  from  1950  to
2011:  i.e.,  the  mean  has  been  removed)  are  presented.  Differences
between  both  are  shown  in  blue.  Annual  efficiency  criteria  between
observed  and  reconstructed  daily  surges  are  presented  in  black.
The  red  shaded  area  marks  the  period  for  which  significant  differences ­
  between  observations  and  20CRv2  are  detected.  The
different  shades  demonstrate  the  gradual  increase  of  inconsistencies
before  the  1910s.

for  the  year  1883),  exceeding  the  mean  of  the  calibration ­
  period  by  approximately  150%.  As  shown  before
with  the  stationary  correlation  of  observed  surge  levels
with  the  NSCI  since  1850  (Fig.  5b),  the  deviations  of
observed  surge  levels  with  those  predicted  through
20CRv2  are  unlikely  to  be  caused  by  the  observational
record  (the  NSCI  and  the  storm  surge  record  are  measures ­
  independently).
A  similar  picture  is  retrieved  by  comparing  the  10-yr
moving  averages  of  the  99.9th  percentile  time  series  of
observed  and  statistically  reconstructed  surges  (Fig.  6).
Over  the  past  100  yr,  the  reconstruction  fits  well  to  the
observations.  The  model  predicts  the  known  decline  in

storminess  in  the  mid-twentieth  century,  the  rapid  increase ­
  until  the  mid-1990s,  and  the  downturn  afterward.
Nevertheless,  in  the  early  1910s,  the  prediction  starts  to
decrease  in  a  manner  not  visible  in  the  observations
(for  both  the  ensemble  mean  as  well  as  the  ensemble
spread).  This  decrease  finally  results  in  significant  positive ­
  long-term  trends  over  the  entire  reanalysis  period
from  1871  to  2010  if  20CRv2  is  taken  as  predictor  (note
that  a  similar  behavior  was  also  observed  for  the  95th
percentile;  Fig.  SI  of  the  supplementary  material).
Related  to  this,  Bronnimann  et  al.  (2012)  demonstrated ­
  that  the  ensemble  mean  appears  to  be  biased
toward  lower  wind  speeds  during  earlier  decades.  They
recommended  the  use  of  single  ensemble  members
rather  than  the  ensemble  mean  when  investigating  longterm ­
  changes.  To  examine  whether  the  results  from
Fig.  6  are  influenced  by  such  biases,  we  additionally
evaluated  long-term  changes  in  each  ensemble  member
separately.  First,  we  calculated  the  differences  between
the  percentile  time  series  from  each  20CRv2  ensemble
member  prediction  and  the  observed  time  series.  Then,
in  a  second  step,  we  computed  linear  trends  for  each  of
the  residual  time  series.  The  results  show  that  for  each
percentile  all  ensemble  members  point  to  significant
positive  long-term  changes,  which  are  further  significantly ­
  different  from  the  observations  (Fig.  7a).  Additionally, ­
  we  found  that  the  residual  trends  are  generally
increasing  with  the  order  of  the  percentiles:  that  is,  highest
deviations  are  found  within  the  highest  percentiles.
To  determine  the  exact  timing  from  which  the  20CRv2
generated  surges  start  to  deviate  significantly  from  the
observations,  we  further  computed  linear  trends  for
the  residual  time  series  over  30-yr  moving  windows.  The
results  are  shown  in  Figs.  7b,c  for  the  99th  and  99.9th
percentiles,  respectively.  While  the  trend  estimates  scatter ­
  around  zero  back  to  approximately  1910,  before  that
time  statistically  significant  differences  are  found  for  the
ensemble  mean  as  well  as  each  individual  ensemble
member.  While  we  can  confirm  the  bias  of  the  ensemble
mean  reported  by  Bronnimann  et  al.  (2012)  (Fig.  7b),  our
results  also  illustrate  that  using  individual  members  cannot ­
  improve  the  results  significantly  (when  assessing  the
long-term  behavior  of  storm  surges  in  the  German  Bight).
The  reanalysis  is  significantly  biased  toward  a  lower  occurrence ­
  of  extreme  values  in  the  period  prior  to  1910  in
both  the  ensemble  mean  as  well  as  all  members  (in  this
region).
The  decreasing  coherence  between  reanalysis  forcing
and  observed  surges  is  generally  in  line  with  increasing
uncertainties  in  the  reanalysis  because  of  fewer  assimilated ­
  observational  data  in  the  earlier  periods  (Compo
et  al.  2011;  Krueger  et  al.  2013b).  The  results  therefore
partly  confirm  the  inconsistencies  between  storm