Forecasting the Aegean Sea POSEIDON-II operational

Korres_et_al_JOO_Feb.qxd
2/25/10
2:46 PM
Page 37
Forecasting the Aegean Sea hydrodynamics within the POSEIDON-II operational system
Forecasting the Aegean Sea
hydrodynamics within the
POSEIDON-II operational system
G Korres, K Nitti, L Perivoliotis, K Tsiaras, A Papadopoulos, G Triantafyllou, all of the Hellenic
Centre for Marine Research, Greece
I Hoteit, King Abdullah University of Science and Technology, Kingdom of Saudi Arabia
The development and first assessment of a high resolution eddy-resolving forecasting
system for the Aegean Sea hydrodynamics, developed as part of the POSEIDON-II
European Economic Area (EEA) Grants project and the European Coastal Sea Operational
Observing and Forecasting (ECOOP) European Union (EU) project, is presented.The system uses an assimilation scheme based on a localized version of the Singular Evolutive
Extended Kalman (SEEK) filter with partial evolution of its correction directions.The filter is
o
used to correct the forecast state of a 1/30 Princeton Ocean Model (POM) of the Aegean
Sea on a weekly basis. Data assimilation experiments are performed over a 12-month
period to estimate the impact of observations insertion on the forecasting skill of the
modelling system.The accuracy of the hydrodynamic state identification is assessed by the
relevance of the assimilation system in fitting the data, and in forecasting future variability.
The results of the assimilation system are also inter-compared with the forecasts estimated
by an identical forecasting system that is re-initialized on a weekly basis from the basin wide
MFS model analysis using a proper downscaling technique.
AUTHORS’ BIOGRAPHIES
Dr Gerasimos Korres is a Research Associate at the Institute of
Oceanography of HCMR. He has a PhD in Physical
Oceanography (University of Athens, Greece). His research
interests and experience include ocean and wave modelling,
data assimilation techniques and air-sea interaction processes.
Dr Kostas Nittis is a Research Director at the Institute of
Oceanography of HCMR. He has a PhD in Physical Oceanography
(University of Athens, Greece).
Mr Leonidas Perivoliotis is a Research Associate Scientist of the
Institute of Oceanography of HCMR. He has an MSc in Physical
Oceanography (University of Athens, Greece).
INTRODUCTION
he Aegean Sea is located to the northeast of the
Ionian and to the northwest of the Levantine Sea. It is
the third major sea of the Eastern Mediterranean
basin and has a complex topography. It is bounded to
T
Volume 3 No. 1 2010
Journal of Operational Oceanography
the east by the Turkish coasts (Asia Minor), to the north and
west by the Greek mainland and to the south by the island of
Crete. Its coastline is very irregular and hundreds of islands are
scattered all over the Aegean. The Aegean Sea is composed of
three major basins: the North Aegean, the Chios basin in
the central part, and in the south the Cretan basin which is the
largest and deepest one.
The water budget and the phenomenology of the Aegean
Sea is largely controlled by the exchange of heat and salt with
the Levantine basin to the south and the exchange of salt
with the Black Sea through the Dardanelles Straits to the
north-east. The Dardanelles Straits circulation consists of a
two-layer flow regime. At the surface, the low salinity waters
outflowing from the Black Sea, a dilution basin where there
is excess precipitation and river runoff over evaporation, are
transported into the Aegean Sea while the saltier Aegean
waters flow as an undercurrent into the Black Sea. On an
annual basis ~1257km3 of low salinity waters (25–35psu
37
Korres_et_al_JOO_Feb.qxd
2/25/10
2:46 PM
Page 38
Forecasting the Aegean Sea hydrodynamics within the POSEIDON-II operational system
depending on the season) are flowing into the Aegean while
957km3 are flowing into the Black Sea.1, 2
The net outflow into the Aegean exhibits maximum values
during mid-summer. The amount of equivalent fresh water
input from the Black Sea can be estimated to be ~1.88my-1 on
an annual basis. It is possible to observe variations between
1my-1 during the winter period (outflow minimum) and
2.77my-1 during summer (outflow maximum). The annual
amount of precipitation (P), evaporation (E) and river-runoff
(R) for the Aegean are 0.5my-1, 1.3my-1 and 0.1my-1 respectively.3, 4 Thus although the climatological value of E-(P+R) is
positive for the Aegean Sea, the Black Sea net outflow gives a
negative overall freshwater budget which in turn drives a net
volume outflow through the Cretan Straits.
The inflow/outflow regime at the Cretan Straits is
complex and shows important seasonal and interannual variability. The Asia Minor Current (AMC) carrying warm and
saline water masses enters the Aegean Sea through the
Kassos, Rhodes and Karpathos Straits and moves northwards.
It has a strong seasonal variability with maximum transport
during winter and minimum during summer. Modified water
masses of Atlantic origin can sporadically enter the Aegean
Sea through the western Cretan Straits and can be detected in
various areas as a subsurface salinity minimum.5
The most frequent winds prevailing in the Aegean Sea
blow from the northern sector. During the winter these
northerlies account for the cold and dry conditions that
prevail over the Aegean. These conditions in turn favor dense
water formation in the north Aegean and in the Cretan Sea
due to extensive heat loss.6
During the summer, dry northerly winds called the
Etesians, which sometimes reach gale force, prevail. They
strengthen from May through June and are well established
from July to September. The wind field during summer is the
superposition of the large scale Etesian wind system with
the localized diurnal sea breeze systems developing on both
the Greek and Turkish coasts.7 The main axis of the Etesian
wind system goes from the North Aegean, passes slightly east
of Limnos and Skyros islands, through Cyclades, past
Karpathos island and towards the center of the eastern
Levantine basin. It is curved cyclonically almost parallel to
the Asia Minor coasts. The Etesians cause upwellings over
the eastern part of the Aegean that in turn cause the formation
of a cool water zone in the same area except in the shallows
along the Turkish coasts. This surface cooling also causes
also high static stability in the atmosphere and operates as a
feedback for the increase of Etesian wind system.8
The Aegean and Ionian Seas surrounding Greece
constitute one of the main links from Europe to the Eastern
Mediterranean and Russia. Therefore, these waters serve as
the main routes of oil transportation to and from Europe. This
results in a continuous pressure to the marine environment, as
indicated by the following remarks:
1.
2.
38
During 2000, the Greek authorities reported 379
oil-pollution incidents, most of them caused by ships.
According to the map published by the Joint Research
Centre (JRC) for illicit vessel discharges in the
Mediterranean during 1999, over 400 oil-spills were
detected by SAR imagery around Greece.
Recent developments in naval architecture have led to a
growth in the amount of oil transport throughout the Aegean
Sea. The expected installation of the Burgas–Alexandroupolis
large-capacity pipeline in the near future, connecting the harbor
of Alexandroupolis (close to the river Evros delta region) to oilproducing countries will allow, for the first time, large-volume
ships to travel through the Aegean, greatly increasing the potential of large-scale pollution accidents. The expected increase of
oil traffic in the area coupled with the increased size of tankers
and the large number of scattered islands and ports situated at
short distances from international shipping routes results in a
high risk of a serious accidental pollution in the Aegean Sea. At
the same time accurate forecasts for the fate of the oil in the sea
require realistic information on the atmospheric and oceanic
conditions for the whole water column.
Complex topography and various exchanges taking place
between the Levantine, Ionian and Black Sea, turn the
development of a capable nowcast/forecast modelling system
for the Aegean Sea into a rather challenging effort. This
forecasting system will operationally provide valuable information for safe marine navigation, coastal management
activities, monitoring of pollution events and control of emergency events etc. Moreover, recent studies9, 10 have shown that
the Aegean Sea constitutes an important component of the
whole Mediterranean climate system and thus understanding
its role to the overall climate dynamics shares its own merit.
Monitoring and forecasting at the regional and coastal scale
are key aspects of the POSEIDON system. The monitoring
basis of the system is an operational network of ten oceanographic buoys including two multi-parametric deep sea stations.
The network is complemented by a Ferry-Box system measuring the surface properties of the sea11 and a High Frequency
(HF) radar current measuring system.12 Additionally, an Argo
floats observing component is under development in the framework of Euro-Argo project in which the Hellenic Centre for
Marine Research (HCMR) actively participates. Euro-Argo was
one of the 35 projects selected in 2006 for the European
Roadmap for Research Infrastructure by ESFRI (European
Strategy Forum on Research Infrastructures).
Pre-operational monitoring and forecasting capacity has
been established in the Aegean Sea during the past years
through a number of research and technology development
projects. The most important contribution was the
POSEIDON-I (1997–2000) and POSEIDON-II (2004–2008)
national projects that established an integrated real-time
monitoring and forecasting system for the marine environmental conditions of the Aegean Sea.13 Both the observing and
modelling components of the POSEIDON system are
primarily focused on the physical processes (atmospheric conditions, waves and hydrodynamics) of the marine environment
while its biochemical monitoring components are rather limited and at a research level. The system description, its main
limitations and a preliminary evaluation of its forecasting skill
are given by Nittis et al (2001).14 The quality of biochemical
measurements and the lack of data from deeper layers of the
water column have been identified as main limitations of the
observing system. The meteorological forecasting model was
found to have increased skill for air temperature and
atmospheric pressure while the wind speed was found to be
underestimated by 10–20% during extreme events.
Journal of Operational Oceanography
Volume 3 No. 1 2010
Korres_et_al_JOO_Feb.qxd
2/25/10
2:46 PM
Page 39
Forecasting the Aegean Sea hydrodynamics within the POSEIDON-II operational system
The POSEIDON operational system validation is still
under development. Zervakis et al (2002)15 inter-compared
hydrodynamic model results with XBT observations
collected along transects across the Eastern Mediterranean.
This evaluation showed that the model adequately predicts
the seasonal cycle of the thermocline evolution but tends to
generate less steep thermocline than observed. These authors
concluded that several mesoscale circulation features were
not accurately predicted by the model.
This paper describes the three different versions of the
Aegean Sea hydrodynamics forecasting system starting from
version V0 of the POSEIDON-I system. It further provides an
assessment of the forecasting skills of the most recent versions
(V1 (-) and V1) over a 12-month period (January–June 2008)
respectively based on a proper initialization technique and a
Singular Evolutive Extended Kalman filter (SEEK) assimilating in-situ (T/S Argo profiles and XBTs) and remotely sensed
Sea Surface Height (SSH) and Sea Surface Temperature (SST)
observational data sets. This is the first work that uses an
advanced assimilation scheme such as the SEEK filter to
forecast the ocean state at a very high resolution.
A brief description of version V0 of the Aegean Sea model
is provided. This is the initial version of the system developed
during POSEIDON-I project. System versions V1 (-) and V1
are also discussed and the assimilation system developed for
version V1 is described. The inter-comparison of forecasts
produced by model versions V1 (-) and V1 and the assessment
of the forecast skill is also carried out. A brief discussion and
summary are offered in conclusion.
AEGEAN SEA – POSEIDON BASIC
SYSTEM DESIGN (V0)
The Aegean Sea hydrodynamic model16 developed during the
POSEIDON-I project13 is based on the Princeton Ocean model
(POM). POM is a primitive equation free surface ocean model
Fig 1: Aegean Sea model version
V0 and Eastern Mediterranean
model bathymetry maps
Volume 3 No. 1 2010
Journal of Operational Oceanography
39
Korres_et_al_JOO_Feb.qxd
2/25/10
2:46 PM
Page 40
Forecasting the Aegean Sea hydrodynamics within the POSEIDON-II operational system
Version name
Area
V0
20 E–29 E
o
o
30 N–41 N
o
o
19.5 E–30 E
o
o
30.4 N–41 N
o
o
19.5 E–30 E
o
o
30.4 N–41 N
V1(-)
V1
o
o
Horizontal/vertical
resolution
o
o
1/20 x 1/20 /29
sigma layers
o
o
1/30 x 1/30 /24
sigma layers
o
o
1/30 x 1/30 /24
sigma layers
Nesting
Initialization
method
Data
assimilation
EMED 1/10o
-
-
MFS/OGCM
o
1/16
MFS/OGCM
o
1/16
VIFOP
-
VIFOP
SEEK
Table 1:The different versions of the Aegean Sea modelling system
which operates under the hydrostatic and Boussinesq
approximations.17 The POM model equations are written in
sigma-coordinates and discretized using the centered secondorder finite differences approximation in a staggered ‘Arakawa
C-grid’ with a numerical scheme that conserves mass and energy.
o
The model domain covered the geographical area 20 E –
o
o
o
29 E and 30 N – 41 N (Fig 1) with a horizontal resolution of
o
1/20 and 29 sigma layers through the vertical with a logarithmic distribution near the surface and the bottom. The model
was forced with 3-hourly surface fluxes of momentum, heat
and water provided by the POSEIDON - Eta high resolution
o
(1/10 ) regional atmospheric model18, 19 issuing forecasts for
72 hours ahead (Table 1).
Boundary conditions at the western and eastern open boundaries of the Aegean Sea hydrodynamic model were provided by a
large scale model based on POM code, covering the whole
o
Eastern Mediterranean Sea with a resolution of 1/10 (Fig 1). The
nesting between the two models involved the zonal/meridional
external (barotropic) and internal velocity components, the tracer
profiles and the free surface elevation following the method
described in the literature.20 A detailed description of the modelling system and a preliminary evaluation of its forecasting skill
are provided by a number of authors.14, 15, 21 The latest model setup
(before the transition to system version V1) involved the
implementation of a freshwater flux surface boundary condition
instead of the virtual salt flux boundary condition22 and the
implementation and pre-operational testing of a multivariate data
assimilation system based on the SEEK filter.23, 24
AEGEAN SEA MODEL VERSION V1
The Aegean Sea POSEIDON operational system V0 was
recently upgraded towards the V1 (-) and V1 versions
(Table 1). The new system was enhanced with the following
features:
1.
o
The horizontal resolution was increased from 1/20 x
o
o
o
1/20 to 1/30 x 1/30 in order to resolve better the complex topographic features of the Aegean Sea and the
complex coastline due to the presence of numerous
islands. Moreover the new resolution better resolves the
small spatial scales of the circulation characteristics
within the northern part of the Aegean Sea where the
internal Rossby radius can reach values as low as 1–2km.
The vertical resolution of the model was decreased with
respect to V0 version to 24 sigma layers. The new
configuration was successfully tested through various
hindcast experiments.
Fig 2: Aegean Sea model version
V1 bathymetry (m)
40
Journal of Operational Oceanography
Volume 3 No. 1 2010
Korres_et_al_JOO_Feb.qxd
2/25/10
2:46 PM
Page 41
Forecasting the Aegean Sea hydrodynamics within the POSEIDON-II operational system
2.
3.
4.
5.
o
High resolution (1/20 ) atmospheric forcing.
The upgraded Aegean Sea model has been one-way
nested to the Mediterranean Forecasting System (MFS)
OGCM model of the Mediterranean Sea at a horizontal
o
resolution of 1/16 and 72 levels in the vertical.25
A downscaling variational initialization (V.I.) technique
was used to initialize forecasting runs of the Aegean Sea
model on a weekly basis from the analysis of the MFS
model as described by Auclair et al (2000).26
The implementation of a data assimilation scheme based
on a localized version of the SEEK filter.
The upgraded version V1(-) of the Aegean Sea forecasting
system uses the features described above in numbers 1 to 4.
V1 uses the features numbered 1 and 3 and 5. These new
attributes are described in more details below.
MODEL CONFIGURATION AND
ATMOSPHERIC FORCING
o
The model domain covers the geographical area 19.5 E –
o
o
o
30 E and 30.4 N – 41 N (Fig 2) with a horizontal resolution
o
of 1/30 and 24 sigma layers along the vertical with a logarithmic distribution near the surface and the bottom. The
model includes parameterization of fresh water discharge
from major Greek rivers (Axios, Aliakmonas, Nestos, Evros)
while the inflow/outflow at the Dardanelles is treated with
open boundary techniques.16 The Aegean Sea model is forced
with hourly surface fluxes of momentum, heat and water
o
provided by the POSEIDON - Eta high resolution (1/20 )
regional atmospheric model27 issuing forecasts for 5 days
ahead. The net shortwave and the downward long-wave
radiation terms are provided directly by the atmospheric
model while the upward long-wave radiation and the turbulent fluxes are calculated by the hydrodynamic model using
its own SST and the relevant atmospheric parameters (air
temperature, relative humidity and wind velocity).
Nesting procedures
Boundary conditions at the western and eastern open
boundaries of the Aegean Sea hydrodynamic model are
provided on a daily basis (daily averaged fields) by the MFS
system. The nesting between the two models involves the
zonal/meridional external (barotropic) and internal velocity
components, the temperature/salinity profiles and the free
surface elevation following the nesting procedures described
in Korres and Lascaratos (2003).20 Additionally, volume
conservation constraints between the two models are applied
at both open boundaries of the Aegean Sea model.
Model initialization
The Aegean Sea model is re-initialized from the
Mediterranean Forecasting System-Oceanic General
Circulation Model (MFS-OGCM) results once every week. In
order to filter out spurious oscillations that may occur during
the re-initialization procedure, the Variational Initialization
and Forcing Platform (VIFOP) optimization tool has been
implemented26 in the forecasting system. VIFOP is a variational initialization technique based on the minimization of a cost
function involving data constraints as well as a dynamical
Volume 3 No. 1 2010
Journal of Operational Oceanography
penalty involving the tangent linear model. The spurious
oscillations affect mainly the barotropic variables of the model
(depth integrated velocities and free surface elevation)
contaminating the model forecasts. They can be efficiently
suppressed in magnitude and duration by optimizing the
global divergence of the external mode and the free surface
tendency and by using enforcing a strong constraint on the
free surface time tendency. This tendency usually consists of a
physical and a spurious numerical part which during the optimization phase, is minimized without affecting the physical
time tendency, which is the real signal.
Data assimilation
The assimilation scheme is based on the Singular Evolutive
Extended Kalman (SEEK) filter which is an error subspace
extended Kalman filter that operates with low-rank error
covariance matrices as a way to reduce the prohibitive computational burden of the extended Kalman filter.23 The filter
is additionally implemented with covariance localization and
partial evolution of the correction directions. The approximation of ‘partial evolution’ of the correction directions was
first proposed by Hoteit et al (2002, 2003)28, 29 and subsequently by Korres et al (2009)11 who found that this approach
has a limited impact on the filter performance. Partial evolution of the correction directions significantly reduces the
computational cost of the SEEK filter making it more suitable for operational systems. Its use in the high resolution
Aegean Sea model can be further supported by the fact that
the evolution of the last (ie, least significant) empirical
orthogonal functions (EOF) in the SEEK filter can be problematic and might introduce noise in the filter correction
directions, and this was observed in preliminary experiments. This happens because the tangent linear-based
evolution equation of the SEEK filter might fail to track fast
fine-scale variations similar to the ones resolved by the last
EOFs of the Aegean Sea high resolution coastal model,
bearing in mind that the filter was designed to follow slow
dynamical changes only.30
The Aegean Sea model was first integrated for one year
period (Sept 2007–Sept 2008) re-initialized on a weekly basis
o
from the MFS model analysis and forced with the Eta 1/20
atmospheric model 24-hour forecasts in order to sample a set
of 366 daily model realizations. This set of model outputs is
used to determine ‘multivariate’ EOFs needed for the
initialization of the SEEK filter. Before applying the EOF
analysis, state variables were normalized by the inverse of the
square-root of their domain-averaged variances as they are of
different nature. The EOF analysis suggests that more than
85% of the model variability during the one year period is
represented by the first 60 EOFs.
o
o
1/8 gridded AVISO SSH data, 1/16 gridded AVHRR SST
data (Gruppo di Oceanografia da Satellite (GOS) Optimally
Interpolated SST), T/S Argo profiles and temperature profiles
from any available XBTs in the Aegean Sea were assimilated
into the system. These data were used on a weekly basis to correct the forecast state of the Aegean model by projecting the
forecast error (or innovation) into the state space using the time
evolving filter statistics. Data assimilation is performed on a
sequence of 7-day assimilation cycles. Once a week, (Day J,
Monday) an analysis is produced for Day J-6 for which the
41
Korres_et_al_JOO_Feb.qxd
2/25/10
2:46 PM
Page 42
Forecasting the Aegean Sea hydrodynamics within the POSEIDON-II operational system
gridded near real time SSH product is made available from the
AVISO group. The observational data set that is assimilated at
Day J-6, apart from SSH consists also of the satellite SST gridded data of the same date and the in-situ T and T/S profiles
from XBTs and ARGO floats respectively covering the period
from Day J-13 to Day J-6. The latter data (XBTs and ARGO
profiles) are assimilated all at once by calculating their misfit
with model’s forecast at Day J-6. After the analysis for Day
J-6 is produced, the model and the filter error statistics are integrated forward in time until Day J, forced with atmospheric
analysis fields. Finally, at Day J, a 5-day forecast is produced.
For the dates J+1->J+6 the model produces 5-day forecasts on
a daily basis restarted each time from the previous day’s forecast. As for the error statistics, during the days that intervene
until the next analysis phase (produced at Day J+1) they are
propagated in time following the model dynamics.
The observational data are assimilated with a nominal
o
accuracy of 3cm for the SSH data, 0.8 C for the SST data,
0.04psu for the salinity profiles and a depth varying error for
o
o
temperature profiles (0–5m: 0.8 C; 5–20m: 0.6 C; 20–100m:
o
o
o
0.4 C; 100–500m: 0.2 C; 500-bottom: 0.1 C). In the assimilation runs, the rank of the filter error covariance matrix was set
to 60. A higher rank of the error covariance matrix did not lead
to any significant improvement in the filter performances.
After several assimilation experiments with different number
of evolving correction directions, the first 10 modes were
allowed to evolve with the tangent linear model as in the SEEK
filter while the rest of the modes, corresponding to the EOFs
generally associated with the small eigenvalues,31 were kept
invariant in time. An influence radius of 200km was chosen
to localize the filter correction.32 The filter correlations were set
to zero outside this radius to ensure no influence of any data
beyond this distance. A further decrease of the radius of influence in the present system introduced dynamically imbalanced
structures in the analysis state subsequently triggering instabilities in the model forecast
a)
b)
c)
d)
Fig 3: a) Mean SSH field corresponding to Aegean model version V1(-) 2008 simulation (cm); b) Mean SSH field corresponding
to Aegean model version V1 2008 simulation (cm); c) AVISO mean sea level for year 2008 (cm); d) MFS OGCM mean SSH field
for year 2008 (in cm)
42
Journal of Operational Oceanography
Volume 3 No. 1 2010
Korres_et_al_JOO_Feb.qxd
2/25/10
2:46 PM
Page 43
Forecasting the Aegean Sea hydrodynamics within the POSEIDON-II operational system
ASSIMILATION RESULTS
Two experiments were performed over the period Jan–Dec 2008
in order to evaluate the performance of the assimilation system
and estimate the impact of the data assimilation process in the
model. In the first experiment the Aegean Sea model (system
Version V1 (-)) is periodically (on a weekly basis) initialized
from the MFS analysis model results, thus benefiting from the
assimilation of the data into the larger domain Mediterranean
model. Every week, the MFS model assimilates all available
satellite data of SLA (AVISO) and SST (GOS AVHRR SST
fields), in-situ profiles from Argo floats and XBTs.33 In the second experiment the Aegean Sea model is integrated for the same
period of time nested with the MFS model and assimilating
satellite SSH, SST data and Argo T/S profiles on a weekly basis
using the SEEK filter (system Version V1).
An important point of difference between the two versions of the Aegean operational system V1 (-) and V1 that
needs to be discussed is related to the mean sea surface
height. As previously stated, system version V1(-) is periodically re-initialized from the MFS model assimilating along
track satellite Sea Level Anomalies (SLA) data that have been
converted to sea surface height measurements using the mean
dynamic topography (MDT) as estimated by Dobricic
(2005)34 from the diagnostics of an existing operational
assimilation system in the Mediterranean (SSH=SLA+MDT).
The system version V1 directly assimilates AVISO gridded
SSH data that rely on the Mediterranean synthetic mean
topography (SMDT) as calculated by Rio et al (2007)35 using
a first guess 7-year Mediterranean model output, observed
drifting buoy velocities and altimetric data. The one year
(2008) mean SSH fields corresponding to Aegean Sea system
versions V1(-) and V1 are shown in Fig 3a and Fig 3b respectively, while the AVISO SSH mean for the same period is
plotted in Fig 3c. The SSH mean of year 2008, as it results
from the V1 assimilation system, compares very well with the
AVISO mean SSH. In the Ionian Sea both mean fields bear a
o
strong signature of the Pelops anticyclone located at 21.5 E
o
35.5 N. This gyre is only faintly represented in the V1 (-)
2008 mean SSH picture partially due to the fact that this gyre
does not bear a strong signal in the MFS model SSH mean for
year 2008 (shown in Fig 3d). The evolving error statistics of
the V1 system being consistent with the model dynamics
seem to be more efficient in inserting and sustaining the signal of the Pelops gyre in this version of the model.
The time evolution of the basin averaged root mean
square (rms) errors for the sea level anomalies (SLA) corresponding to model analyses is shown in Fig 4 for both systems V1(-) and V1. The SLA rms error of Version V1,
remains for the whole time period much lower than the one
corresponding to version V1(-) indicating a positive impact of
the assimilation performed at the level of the Aegean Sea
model. On average, the SLA rms error of V1 (-) is about
4.88cm. It goes down to 3.62cm for version V1. Considering
the total signal (SLA+MDT), system version V1 presents an
analysis rms error of 2.7cm on average, which is already
below the prescribed observational error uncertainty of 3cm
that has been specified.
An inter-comparison of the forecast SLA rms error of version V1 with the analysis SLA rms error of version V1 (-) is
shown in Fig 5. The forecast rms error is computed just before
Volume 3 No. 1 2010
Journal of Operational Oceanography
Fig 4:Time evolution of the SLA misfits (cm) as they result
from system versions V1 (-) and V1 analyses
the analysis update and can be used as an indicator of the
consistency between the model dynamics and the filter statistics, and of the ability of the filter in assimilating the data into
the model. This is actually one form of comparison of the
assimilation solution with independent data. It can be seen that
the V1 forecast SLA error is higher than the V1 (-) analysis
error during the initial 2–3 months period when the filter
statistics are being adjusted to the model dynamics. After this
initial period, the V1 forecast error becomes gradually equal
and then lower (especially during autumn and winter) than the
V1 (-) analysis error. The fact that the system V1 forecast error
is quite close to the analysis error of the same system, indicates that the statistical correction introduced by the filter is
consistent with the model dynamics. On average, the forecast
SLA RMS error of system V1 is 4.6 cm and it goes to 4.1 cm
when the total signal (ie, SLA+MDT) is taken into account.
Another way to assess the system predictive skill is through
the correlation skill score which is defined as
ρ=
Cov ( M , O )
Var ( M )Var (O )
Fig 5:Time evolution of the SLA misfits (in cm) as they result
from system versions V1 (-) analysis and V1 forecast
43
Korres_et_al_JOO_Feb.qxd
2/25/10
2:46 PM
Page 44
Forecasting the Aegean Sea hydrodynamics within the POSEIDON-II operational system
where M is the model analysis/forecast and O is the verifying
observation field. This statistic is applied to quantify the
predictive skill of system V1 in terms of SSH and SLA.
Fig 6a presents the time series of the correlation skill score
for SSH corresponding to system V1 forecast and analyses
(ie, the model skill exactly before and exactly after the assimilation of SSH observations). The overall skill being equal to
0.92 for model analyses and 0.83 for model forecast, shows a
very good performance of the system in predicting correctly
the SSH signal. The SSH analyses skill score is almost always
above 0.9. The forecast skill, which is 0.66 after the first
assimilation cycle, improves gradually and reaches a value of
0.85 after the first 7–8 assimilation cycles. The system V1
correlation skill score for SLA is shown in Fig 6b for model
analyses and forecast over the year 2008 period. The model
analyses skill score is on average 0.68 and shows little variations around this value with the worst cases reaching 0.5–0.4.
On the other hand, the model forecast skill score for SLA
shows larger variations around an average value of 0.37.
It is interesting to examine the time evolution of the
analysis SST rms error (with respect to GOS AVHRR
Optimally Interpolated Sea Surface Temperature (OISST)) as
they result from the systems V1 (-) and V1 (Fig 7a). For both
systems, the evolution of the SST error shows seasonality
with lower values during the winter and spring period and
abrupt error increase during summer and autumn period
(June–Sept 2008). Overall the SST rms error is comparable
for both systems although system V1 (-) error presents more
pronounced peaks during the time period from April to June
2008 with respect to system V1. The bias times series defined
as <A – O> (where A is the model SST analysis, O the GOS
OISST observations and <..> denotes a spatial mean) is
shown in Fig 7b for both system versions. System V1 (-)
shows a systematic negative bias (ie, the model’s mixed layer
temperature is colder than the observations) which becomes
more pronounced during the warm season (June–Sept). The
o
annual mean value of this bias is -0.26 C while during the
o
warm season reaches -0.5 C. On the other hand, system V1
shows an improvement with respect to V1 (-) as the overall
o
bias is only -0.04 C.
It is interesting to define three time periods where the
system V1 exhibits different behavior in terms of the SST bias.
a)
a)
b)
b)
Fig 6: a) Time evolution of the SSH (SLA+MDT) correlation
skill score corresponding to system version V1 analyses and
forecast; b) Time evolution of the SLA correlation skill score
corresponding to system version V1 analyses and forecast
Fig 7: a) Time evolution of the SST misfits (in oC) as they
result from system versions V1 (-) and V1 analyses; b) Time
evolution of the SST bias (in oC) as they result from system
versions V1 (-) and V1 analyses
44
Journal of Operational Oceanography
Volume 3 No. 1 2010
Korres_et_al_JOO_Feb.qxd
2/25/10
2:47 PM
Page 45
Forecasting the Aegean Sea hydrodynamics within the POSEIDON-II operational system
a)
b)
c)
d)
e)
f)
g)
Fig 8: a) MEDARGO float 1900603 track during 30 Nov-20
Dec 2008 yielding 10 sets of profiles; b) MEDARGO float
1900604 track during 31 Dec 2007-28 Aug 2008 yielding 13
sets of profiles; c) MEDARGO float 1900606 track during 14
Jan 2008-10 Sep 2008 yielding 23 sets of profiles; d)
MEDARGO float 1900630 track during 31 Dec 2007-30
Dec 2008 yielding 19 sets of profiles within the model area;
e) MEDARGO float 6900098 track during 31 Dec 2007-30
Dec 2008 yielding 79 sets of profiles within the model area;
f) MEDARGO float 6900284 track during 1 Jan 2008-8 June
2008 yielding 19 sets of profiles within the model area; g)
MEDARGO float 6900455 track during 31 Dec 2007-14
Nov 2008 yielding 58 sets of profiles within the model area.
Volume 3 No. 1 2010
Journal of Operational Oceanography
45
Korres_et_al_JOO_Feb.qxd
2/25/10
2:47 PM
Page 46
Forecasting the Aegean Sea hydrodynamics within the POSEIDON-II operational system
During the warm season, the model mixed layer is colder by
o
0.24 C in average than the observations. This behaviour is partially similar with model results (using system version V0)21
which showed a systematic underestimation of SST being
more pronounced during the late spring and summer period
probably due to the heat flux parameterization scheme used by
the model and the overestimation of the mixed layer depth.
However, the situation in system V1 is different from V0 as in
V1 the model SST is slightly warmer than the observations
during the cold season while the overall bias is close to zero.
The rms differences between system V1 SST forecasts and
GOS OISST are shown in Table 2 for lead times +0h, +24h and
+48h. Overall, V1 system nowcasts SST with an accuracy of
o
o
o
0.79 C which is further increased to 0.8 C and 0.81 C for +24h
and +48h forecast lead times respectively. The summer period
(June–Sept) is characterized by an increase of the SST rms difo
ference with respect to the overall mean. It becomes 0.92 C for
model nowcasts while for +24h and +48h lead times forecast
o
o
is 0.94 C and 0.95 C respectively. During the cold season
(Nov–March), the system shows an improved forecasting skill
in terms of SST and the rms errors associated with +0h, +24h
and +48h forecasts are generally lower.
Lead time
Overall (oC)
0h
24h
48h
0.79
0.80
0.81
Jun–Sep
period (oC)
0.92
0.94
0.95
Nov–Mar
period (oC)
0.67
0.68
0.69
Table 2: RMS difference between model version V1 SST and
GOS SST
A further assessment of the results of the systems V1 (-) and
V1 is achieved by comparing several Argo floats temperature
and salinity profiles with the relevant model forecasted profiles. The floats, referred to as MEDARGO, are part of the
global ocean Argo project36 which is a significant component
of the global ocean observing system (GOOS). T/S profiles
from Argo floats are nowadays assimilated in many operational systems at global and regional scale. Particularly, in the
Mediterranean Sea they provide, together with the XBT data
collected along the Voluntary Observing Ships (VOS) tracks,
a significant component of the data assimilation and forecast
system at basin scale.37, 38 As within the Mediterranean Sea
density is significantly controlled by salinity, the concurrent
assimilation of temperature and salinity profiles is expected
to be of particular relevance.
In the time frame studied here, a total of seven Argo
floats were continuously or transiently present within the
geographical area covered by the Aegean Sea model domain.
Unfortunately all of these floats made their tracks within the
central Levantine basin between Crete and the African coasts.
Thus, the assessment in terms of efficiency in forecasting the
Argo floats T/S profiles involves only the southern part of the
model domain as there are no Argo tracks available elsewhere. The tracks of the seven floats are shown in Fig 8a to
Fig 8g. In the case of system version V1, the Argo profiles are
assimilated directly to the model on a weekly basis together
with satellite SSH and SST data. On the other hand, system
version V1 (-), implicitly benefits from Argo profiles through
46
the weekly assimilation process carried out at the coarse
resolution model level (MFS-OGCM) from which the system
V1 (-) is periodically re-initialized. Thus the rms error profiles calculated for both system versions along the tracks of
the Argo floats constitute a quasi independent estimator in the
sense that differences between model T/S profiles and ARGO
observations are calculated taking into account model analysis and model forecasts as one follows the track of the float.
Fig 9a to Fig 9g shows the rms error T/S profiles (with
respect to the observed Argo profiles) corresponding to system
versions V1 (-) and V1 for the Argo tracks presented in Fig 8.
These error profiles have been calculated by considering the
differences between Argo profiles at the different ascending
positions on each Argo track and the respective model profiles
corresponding either to model analysis or forecast at zero lead
times. In almost all floats presented, system version V1 seems
to be more efficient in correcting the T/S profiles error over
large parts of the vertical column. The most significant impact
of direct assimilation of Argo profiles into the model can be
clearly seen within the upper 300–400m of the water column
and this is more pronounced for the temperature profiles. The
temperature rms error for both systems, ranges between 0.5o
o
1.5 C within the first 150m and reduces gradually to 0.5–0.1 C
below this depth. On average, the direct insertion of Argo
profiles done in system V1 induces an error reduction for temperature profiles in the first 200m that varies between
o
o
0.1–0.5 C and can reach even 1 C as in the case of Argo float
tracks shown in Fig 8b and Fig 8e respectively. Below 200m
the error reduction with respect to system version V1 (-) is
o
between 0.1–0.2 C which is also significant considering that
the original RMS error for this system varies between
o
0.1–0.5 C. The salinity RMS error for the two system runs
varies between 0.1–0.2psu in the first 150m of the water column. In this layer, the direct insertion of Argo salinity profiles
into the model can reduce the error by 0.02–0.05psu or even
by 0.1psu as in the case of error profiles shown in Fig 9g
corresponding to an Argo float which covers an extensive area
to the southeast of the modelling domain (Fig 8g). However,
there are also cases of error intensification as observed in the
case of float tracks error profiles of Fig 9a and Fig 9b. Below
150m, there are three cases (error profiles shown in Fig 9c, 9d
and 9g) where the error reduction can reach or even surpass
0.05psu in the depth range between 200–500m. For the rest of
the floats the salinity rms error of system V1 is slightly better
or comparable to system version V1 (-) error.
SUMMARY AND CONCLUSIONS
The different development phases of an advanced forecasting
system of the Aegean Sea hydrology and hydrodynamics and a
first assessment of its forecasting capability have been
described and presented. The most recent system is composed
o
of a high resolution (1/30 ) eddy resolving implementation of
POM model within the Aegean Sea nested with the MFS
OGCM model. The forecasting system is forced with hourly
o
1/20 non-hydrostatic Eta atmospheric model surface momentum, heat and freshwater fluxes, and an assimilation subsystem based on a localized version of the singular evolutive
extended Kalman (SEEK) filter. In this application the SEEK
filter was designed to apply the filter correction to the model
Journal of Operational Oceanography
Volume 3 No. 1 2010
Korres_et_al_JOO_Feb.qxd
2/25/10
2:47 PM
Page 47
Forecasting the Aegean Sea hydrodynamics within the POSEIDON-II operational system
a)
b)
c)
d)
e)
f)
g)
Fig 9: Average temperature (in oC) and salinity (in psu) RMS
error profiles corresponding to model Versions V1(-) and V1
with respect to: a) MEDARGO float 1900603 profiles;
b) MEDARGO float 1900604; c) MEDARGO float 1900603;
d) MEDARGO float 1900630; e) MEDARGO float 6900098;
f) MEDARGO float 6900284; g) MEDARGO float 6900455
Volume 3 No. 1 2010
Journal of Operational Oceanography
47
Korres_et_al_JOO_Feb.qxd
2/25/10
2:47 PM
Page 48
Forecasting the Aegean Sea hydrodynamics within the POSEIDON-II operational system
solution along 60 directions of correction (modes) 10, of which
(the most dominant ones) are allowed to evolve in time to track
large scale changes in the evolution of model dynamics. For
high resolution applications such as the Aegean Sea discussed
here, the partial evolution approach which helps avoiding the
propagation of errors associated with higher order modes, can
be beneficial to enhance the model stability and to simplify the
computing resources needed for the operational run of the system. Moreover, the localization of the filter error covariance
matrix generally increases the filter efficiency by allowing
more degrees of freedom during the analysis step. Experiments
with and without data assimilation were performed in order to
validate the system and to demonstrate the effectiveness of the
filter while assimilating a multivariate data set of SSH, SST
and T/S profiles measurements. Multivariate data were
properly assimilated by the model. Weekly assimilation of
satellite SSH and SST data and T/S profiles from ARGO floats
into the model, conducted over 2008, showed the ability of
the system to provide a reliable forecast. More specifically, the
accuracy of the sea level analyses is on average below the nominal observational error of 3cm characterized by an average
correlation skill score of 0.92. Almost the same conclusion can
be drawn for the sea level anomaly analysis rms error. In this
case the average correlation skill score is on average equal to
o
0.68. At the same time the average SST rms error of 0.79 C is
again lower than the specified SST observational error. On the
other hand, the inter-comparison of system version V1 with
V1(-) results (in terms of SLA, SST and vertical T/S profiles)
showed the beneficial role of directly blending model solutions
with observations at the high resolution scale rather than
periodically downscaling the solution of a coarser model (MFS
OGCM) that carries the data assimilation exercise.
The overall behavior of the present forecasting system is
very encouraging. Some key improvements can be implemented in the near future to enhance the system behavior. Such an
example is a recent installation of a HF radar system that
periodically measures the surface currents distribution in front
of Dardanelles Straits (the geographical area east of Limnos
Island). Data from this observation system can be used to
better parameterize the Dardanelles outflow into the Aegean
Sea model and/or to be assimilated directly into the model
solution. Another example is the enhancement of the observational network within the Aegean Sea (and the implied
strengthening of the assimilation sub-system) with the addition
of a Ferrybox system that has been used in the past in order to
continuously measure SSS, SST, chlorophyll a and turbidity
data along a ferryboat route from port Piraeus to HeraklionCrete and is currently under maintenance. SSS data from this
system has been assimilated in the past into the V0 version of
the model. Finally another line of improvement can be
obtained by assimilating more accurate SSH on a shorter time
scale and the mean SSH products through the utilization of an
extended network of tide gauges within the Aegean Sea which
is currently under development.
ACKNOWLEDGEMENTS
This work has been supported by the POSEIDON-II project
co-funded by EEA Grants and the Hellenic Ministry of
National Economy, and the ECOOP project co-funded by
48
EU-FP6 and the Hellenic Ministry of Development (General
Secretariat for Research and Technology). The altimeter
product were produced by Ssalto/Duacs and distributed by
AVISO, with support from CNES (http://www.aviso.
oceanobs.com/duacs). Optimally interpolated SST AVHRR
data were provided by the Gruppo di Oceanografia da
Satellite, Institute of Atmospheric Sciences and Climate of
the Italian National Research Council (GOS-CNR-ISAC).
REFERENCES
1. Unluata U, Oguz T, Latif MA and Oszoy E. 1990. On
the physical oceanography of the Turkish Straits. The
Physical Oceanography of Sea Straits. Kluwer Academic
Publishing, Dordrecht.587pp.
2. Besiktepe S, Ozsoy E, and Unluata U. 1993. Filling the
Marmara Sea by the Dardanelles lower layer inflow. Deep
Sea Research I 40 (9): 1815–1838.
3. Drakopoulos PG, Poulos SE and Lascaratos A. 1998.
Buoyancy fluxes in the Aegean Sea, Rapp. Comm. Int. Mer
Medit 35: 134–135.
4. Poulos SE, Drakopoulos PG and Collins MB. 1997.
Seasonal variability in sea surface oceanographic conditions
in the Aegean Sea (Eastern Mediterranean): an overview.
J. Mar. Sys 13: 225–244.
5. Karageorgis A, Gardner D, Georgopoulos A. Mishonov
E, Krasakopoulou and Anagnostou C. 2008. Particle dynamics in the Eastern Mediterranean Sea: A synthesis based on
light transmission, PMC, and POC archives (1991–2001).
Deep Sea Research I 55: 177–202.
6. Georgopoulos D, Theocharis A, and Zodiatis G. 1989.
Intermediate Water formation in the Cretan Sea (S. Aegean
Sea). Oceanol. Acta 12: 353–359.
7. Meteorological Office. 1962. Weather in Mediterranean,
Vol. I, General Meteorology, H.M.S.O., London, Second Edition.
8. Metaxas DA. 1973. Air-Sea interaction in the Greek
Seas and resulted Etesian characteristics. University of
Ioannina Technical Report No. 5: 1–32.
9. Roether W, Manca BB, Klein B, Bregant D, Georgopoulos
D, Beitzel V, Kovac¡evic V, and Luchetta A, 1996. Recent changes
in Eastern Mediterranean Deep Waters. Science 271: 333–335.
10. Theocharis A, Balopoulos E, Kioroglou S,
Kontoyiannis H and Iona A. 1999. A synthesis of the circulation and hydrography of the South Aegean sea and the Straits
of the Cretan Arc (March 1994 – January 1995). Prog.
Oceanogr 44 (4): pp. 469– 509.
11. Korres G, Nittis K, Hoteit I and Triantafyllou G. 2009.
A high resolution data assimilation system for the Aegean Sea
hydrodynamics. J. Mar. Sys. 77: 325–340.
12. Gurgel KW, Antonischki G, Essen H and Schlick T.
1999. Wellen Radar (WERA): a new ground-wave HF radar
for ocean remote sensing. Coastal Engineering 3–4: 219–234.
13. Soukissian T, Chronis G, Nittis K and Diamanti C.
2002. Advancement of operational oceanography in Greece:
the case of the POSEIDON system. Global Atmos. Ocean
Sys 8: 119–133.
14. Nittis K, Zervakis V, Perivoliotis L, Papadopoulos A
and Chronis G. 2001. Operational monitoring and forecasting in the Aegean Sea: system limitations and forecasting skill
evaluation. Mar. Pol. Bul 43: 154–163.
Journal of Operational Oceanography
Volume 3 No. 1 2010
Korres_et_al_JOO_Feb.qxd
2/25/10
2:47 PM
Page 49
Forecasting the Aegean Sea hydrodynamics within the POSEIDON-II operational system
15. Zervakis V, Nittis K, Perivoliotis L and Tziavos C. 2002.
A comparison of model predictions to observations of seasonal
variability and circulation in the Eastern Mediterranean. The
Global Atmosphere Ocean System 8: 141–162.
16. Korres G, Lascaratos A, Hatziapostolou E and
Katsafados P. 2002. Towards an ocean forecasting system for
the Aegean Sea. Global Atmos. Ocean Syst. 8: 191–218.
17. Blumberg AF, and Mellor GL. 1987. A description of
a three-dimensional coastal ocean model, vol.4. AGU,
Washington D.C. 208pp
18. Papadopoulos A, Kallos G, Katsafados P and
Nickovic S. 2002. The Poseidon weather forecasting system:
an overview. Glob. Atmos. Ocean Sys. 8: 218–237.
19. Janjic ZI. 1994. The step-mountain Eta coordinate model:
Further developments of the convection, viscous sublayer and
turbulence closure schemes. Mon. Wea. Rev. 122: 927–945.
20. Korres G and Lascaratos A. 2003. A One-way nested
eddy resolving model of the Aegean and Levantine basins:
Implementation and climatological runs. Analles
Geophysicae, MFSPP – Part I Special Issue 21: 205–220.
21. Nittis K, Perivoliotis L, Korres G, Tziavos C and
Thanos I. 2006. Operational monitoring and forecasting for
marine environmental applications in the Aegean Sea.
Environmental Modelling and Software 21: 243–257.
22. Triantafyllou G, Korres G, Hoteit I, Petihakis G and
Banks AC. 2007. Assimilation of ocean colour data into a
Biochemical Flux Model of the Eastern Mediterranean Sea.
Ocean Science 3: 397–410.
23. Pham D, Verron J and Roubaud MC. 1998. A singular evolutive extended Kalman filter for data assimilation in
oceanography. J. Mar. Sys. 16 (3–4): 323–340.
24. Korres G, Hoteit I and Triantafyllou G. 2007. Data
assimilation into a Princeton Ocean Model of the
Mediterranean Sea using advanced Kalman filters. J. Mar.
Sys. 65 (1–4): 84–104.
25. Tonani M, Pinardi N, Dobricic S, Pujol I and
Fratianni C. 2008. A high resolution free-surface model of the
Mediterranean Sea. Ocean Sci. 4: 1–14
26. Auclair F, Casitas S, Marsaleix P. 2000. Application
of an inverse method to coastal modelling. Journal of
Atmospheric and Oceanic Technology 17: 1368–1391.
Volume 3 No. 1 2010
Journal of Operational Oceanography
27. Papadopoulos A and Katsafados P. 2009. Verification
of operational weather forecasts from the POSEIDON system
across the Eastern Mediterranean. Natural Hazards and
Earth System Sciences 9: 1299–1306
28. Hoteit I, Pham DT and Blum J. 2002. A simplified
reduced Kalman filtering and application to altimetric data
assimilation in the Tropical Pacific. J. Mar. Sys. 36: 101–127.
29. Hoteit I, Pham DT and Blum J. 2003. A semi-evolutive
filter with partially local correction basis for data assimilation
in oceanography. Oceanol. Acta 26: 511–524.
30. Hoteit I, Triantafyllou G and Petihakis G. 2005.
Efficient data assimilation into a complex 3D physical–biogeochemical model using partially local Kalman filters.
Ann. Geophys. 23: 1–15.
31. Hoteit I, Pham DT and Blum J. 2001. A semi-evolutive partially local filer for data assimilation. Mar. Poll. Bul.
43: 164–174
32. Houtekamer PL and Mitchel HL. 2001. A Sequential
Ensemble Kalman Filter for Atmospheric Data Assimilation.
Mon. Wea. Rev. 129: 123–237.
33. Tonani M, Pinardi N, Fratianni C, Pistoia J, Dobricic S,
Pensieri S, de Alfonso M and Nittis K. 2009. Mediterranean
Forecasting System: forecast and analysis assessment through
skill scores. Ocean Sci. 5: 649–660
34. Dobricic S. 2005. New mean dynamic topography of
the Mediterranean calculated from assimilation system
diagnostics. Geophys. Res. Lett. 32 (11).
35. Rio MH, Poulain PM, Pascual A, Mauri E, Larnicol
G and Santoleri R. 2007. A Mean Dynamic Topography of the
Mediterranean Sea computed from altimetric data, in-situ
measurements and a general circulation model. Journal of
Marine Systems 65 (1–4): 484–508
36. Poulain PM. 2005. A profiling float program in the
Mediterranean. Argonautics, 6:2
37. Dobricic S, Pinardi N, Adani M, Tonani M, Fratianni
C, Bonazzi A and Fernandez V. 2007. Daily oceanographic
analyses by Mediterranean Forecasting System at the basin
scale. Ocean Sci. 3: 149–157.
38. Korres G, Tsiaras K, Nittis K, Triantafyllou G and
Hoteit I. 2008. The POSEIDON-II system: Forecasting at the
Mediterranean scale. 5th EuroGoos Conference. Exeter, UK.
49