Bureau Research Report - 011
Comparison of GLOSEA5 and POAMA2.4 Hindcasts 1996-2009 Harry H. Hendon, Mei Zhao, Andrew Marshall, Eun-pa Lim, Jing-Jia Luo, Oscar Alves and Craig MacLachlan
December 2015
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
i
Comparison of GLOSEA5 and POAMA2.4 Hindcasts 1996-2009
Harry H. Hendon1, Mei Zhao1, Andrew Marshall1, Eun-pa Lim1, Jing-Jia Luo1, Oscar Alves1
and Craig MacLachlan2 1Bureau of Meteorology Melbourne, Australia
2UKMO, Exeter, United Kingdom
Bureau Research Report No. 011
December 2015
National Library of Australia Cataloguing-in-Publication entry
Author: Harry H. Hendon, Mei Zhao, Andrew Marshall, Eun-pa Lim, Jing-Jia Luo, Oscar Alves and Craig MacLachlan
Title: Comparison of GLOSEA5 and POAMA2.4 Hindcasts 1996-2009 ISBN: 978-0-642-70671-3 Series: Bureau Research Report - BRR011
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
ii
Enquiries should be addressed to: Harry Hendon Bureau of Meteorology GPO Box 1289, Melbourne Victoria 3001, Australia Contact Email: [email protected]
Copyright and Disclaimer
© 2015 Bureau of Meteorology. To the extent permitted by law, all rights are reserved and no part of
this publication covered by copyright may be reproduced or copied in any form or by any means
except with the written permission of the Bureau of Meteorology.
The Bureau of Meteorology advise that the information contained in this publication comprises
general statements based on scientific research. The reader is advised and needs to be aware that such
information may be incomplete or unable to be used in any specific situation. No reliance or actions
must therefore be made on that information without seeking prior expert professional, scientific and
technical advice. To the extent permitted by law and the Bureau of Meteorology (including each of its
employees and consultants) excludes all liability to any person for any consequences, including but
not limited to all losses, damages, costs, expenses and any other compensation, arising directly or
indirectly from using this publication (in part or in whole) and any information or material contained
in it.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
iii
Contents
Abstract ......................................................................................................................... 1
1. Introduction .......................................................................................................... 1
2. Model Configurations and Hindcasts ............................................................... 4
3. Mean State Comparisons .................................................................................... 9
4. Australian precipitation and temperature skill................................................ 13
5. Climate modes and their teleconnections ....................................................... 25
6. Conclusions and Recommendations for POAMA Development ................... 43
7. References ......................................................................................................... 48
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
iv
List of Tables
Table 1 Accuracy scores averaged over all Australian land points for month 1 and months 2-4. .......................................................................................................................................... 18
List of Figures
Fig. 1 Mean rainfall for February, May, August, and November. Top row is based on AWAP. Middle row is month 1 of GLOSEA5 hindcasts. Bottom row is month 1 of P24 hindcasts. Units are mm/day. ................................................................................. 10
Fig. 2 Mean Tmax for February, May, August, and November. Top row is based on AWAP. Middle row is month 1 of GLOSEA5 hindcasts. Bottom row is month 1 of P24 hindcasts. Units are °C .......................................................................................... 10
Fig. 3 Mean Tmin for February, May, August, and November. Top row is based on AWAP. Middle row is month 1 of GLOSEA5 hindcasts. Bottom row is month 1 of P24 hindcasts. Units are °C .......................................................................................... 11
Fig. 4 Month 1 precipitation bias for all start months relative to GPCP (left column) and CMAP (right column) for GLSOEA5 (top row) and P24 (bottom row). Units are mm/day. ........................................................................................................................ 11
Fig. 5 SST bias for months 1 (left column) and month 2 (right column) for GLOSEA5 (top row) and P24 (bottom row). Units are °C. ............................................................... 13
Fig. 6 As in Fig. 5 but for months 3 and 4. ............................................................................... 13
Fig. 7 Anomaly correlation for precipitation for month 1 for February (top left), May (top right), August (bottom left) and November (bottom right) from GLOSEA5 and P24 hindcasts. Verification is against monthly AWAP ¼ degree analyses that was interpolated onto the native grids of each model. .......................................................... 15
Fig. 8 As in Fig. 7 except for Tmax. ......................................................................................... 15
Fig. 9 As in Fig. 7 except for Tmin. .......................................................................................... 16
Fig. 10 As in Fig. 7 except for months 2-4 for precipitation ........................................................ 16
Fig. 11 As in Fig. 7 except for months 2-4 for Tmax. ................................................................. 17
Fig. 12 As in Fig. 10 except for months 2-4 for Tmin. ................................................................ 17
Fig. 13 Accuracy score for predicting precipitation above median for month 1 for February, May, August, and November for GLOSEA5 (top row) and P24 (bottom row). AWAP analyses, interpolated to native model grids, is used for verification. Units are percentage correct. ........................................................................................ 19
Fig. 14 As in Fig. 13 except for Tmax. ....................................................................................... 19
Fig. 15 As in Fig. 13 except for Tmin. ........................................................................................ 20
Fig. 16 Accuracy score for precipitation above median for months 2-4 for MAM, JJA, SON, and DJF for GLOSEA5 (top row) and P24 (bottom row). ..................................... 20
Fig. 17 As in Fig. 16 except for Tmax ........................................................................................ 21
Fig. 18 As in Fig. 16 except for Tmin. ........................................................................................ 21
Fig. 19 Attributes diagram for predictions of above median rainfall for all land points over Australia or month 1 for February, May, August, and November from GLOSEA5 (top row) and P24 (bottom row). Shaded area indicates forecast contribute
SEASONAL DISTRIBUTIONS OF DAILY HEAVY RAIN EVENTS OVER AUSTRALIA
v
positively to the Brier skill score. Forecasts that fall on the sloping line would have perfect reliability. Size of dots is proportional to number of forecasts in each bin. Probability bins of width 10% were used. ...................................................................... 23
Fig. 20 As in Fig. 19 except for Tmax. ....................................................................................... 23
Fig. 21 Attributes diagram for SON (mnths 2-4) for above median prediction of precipitation (left hand column), Tmax (middle) and Tmin (right hand column) for all land points over Australia. GLOSEA5 is top row and P24 is bottom row. Plotting convention is as in Fig. 19. ............................................................................... 24
Fig. 22 Accuracy score for above median rainfall for GLOSEA5 and P24 verified against CMAP for month 1 for November (top row). Attributes diagram for above median rainfall using all points over Indo-Pacific domain is shown in bottom row. ..................... 25
Fig. 23 Anomaly correlation (top row) and rmse (bottom row) for monthly Nino34 and IOD indices using all start months. Correlation for persistence forecast is included. ........................................................................................................................ 26
Fig. 24 Predicted amplitude of the Nino34 and IOD indices from GLOSEA5 and P24 for all start months. Amplitude was computed using individual members and then averaged over all members. .......................................................................................... 27
Fig. 25 As in Fig. 24 except for Nino4 and Nino3 indices. ......................................................... 27
Fig. 26 Amplitude of the IOD index for forecasts starting in May, August and November. ......... 28
Fig. 27 Anomaly correlation for the Nino34 index for forecasts starting in May. ........................ 30
Fig. 28 Regression onto Nino34 index of observed rainfall (top row) and ensemble mean rainfall from GLSOEA5 (middle) and P24 (bottom) forecast for months 2-4 for SON. Units are mm/day. ............................................................................................... 30
Fig. 29 Regression onto Nino34 and IOD indices of observed rainfall (top), GLOSEA5 (middle) and P24 (bottom) for months 2-4 for JJA and SON. Units are mm/day. .......... 31
Fig. 30 As in Fig. 29 except for Tmax and Tmin in SON. Units are C. ....................................... 31
Fig. 31 Anomaly correlation for predicting SAM index for November using forecasts initialized on 1st November and 1st August (top). Persistence forecast is shown in blue. Anomaly correlation for months 2-4 (SON) for forecasts initialized on 1st August (bottom). ............................................................................................................ 33
Fig. 32 Time series of November SAM index (top) and SON SAM index (bottom) from observations and forecasts. .......................................................................................... 34
Fig. 33 Correlation (left) and rmse/spread (right) for predicting daily SAM index using all start months for days 1-35. Top row is for forecasts initialized on 9th, middle row combines forecasts from 1st and 9th and bottom row combines 25th, 1st, and 9th. ........... 36
Fig. 34 Comparison of SAM forecast skill using Gong SAM index (top) and by projecting onto observed EOF of SLP (bottom). ............................................................................ 38
Fig. 35 Bivariate correlation for Wheeler-Hendon MJO index for November starts (left) and August starts (right). ............................................................................................... 39
Fig. 36 As in Fig. 35 except for rmse and spread for November forecasts. ............................... 40
Fig. 37 As in Fig. 33 except for predicting the daily blocking index using all start months for days 1-30. ................................................................................................................ 42
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
vi
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
1
ABSTRACT
Multiweek, monthly and seasonal forecast performance of the UKMO GLOSEA5 systems is
compared to the BoM POAMA2.4 system using available hindcasts for the period 1996-2009.
The GLOSEA5 system is based on a state-of-the-art high resolution coupled model (60km
atmosphere and 25 km ocean), which is a marked improvement compared to P24 (250km
atmosphere and 200 km ocean). GLOSEA5 is based updated versions of the component UM
models that have been improved to better represent key climate processes such as the MJO and
to reduce climate drift.
Our assessment reveals a marked increase in forecast performance from GLOSEA5 for
prediction of regional Australian climate (precipitation and temperatures) compared to P24,
especially for the first month of the forecast. A more modest improvement is detected in months
2-4. GLOSEA5 also better predicts multiweek variations of the Southern Annular Mode and
blocking and El Nino at long lead, all of which are important for improved prediction of
regional Australian climate. Forecast reliability from GLOSEA5 is found to be good despite the
limited ensemble size available from the hindcast set.
Based on this assessment a number of recommendations for future development of the POAMA
system are made, including development of the next version of POAMA based on a coupled
model comparable to GLOSEA5, sharing of hindcasts with the UKMO so as to better sample
the full range of interannual variability, and implementation of the POAMA coupled
assimilation system into the GLOSEA5 model, and. A number of model deficiencies in
GLOSEA5 were flagged as well (eg, too strong IOD and ENSO variability, biased ENSO
teleconnection to Australian temperatures, rainfall bias over Maritime Continent) that are
degrading GLOSEA5 forecast performance for Australia and could be worked on jointly with
the UKMO.
1. INTRODUCTION
Operational multiweek-seasonal forecasting at the Bureau of Meteorology is currently based
on ensemble predictions provided by the POAMA seasonal forecasting system. The POAMA
system version P24 (Hudson et al. 2013), which has been operational since 2013, has proven
capability to predict the main drivers of Australian climate variations, including predicting the
occurrence of El Nino and La Nina 2-3 seasons in advance (Zhao et al. 2014) and the Indian
Ocean Dipole up to one season in advance (Shi et al. 2012), interannual variations of the
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
2
Southern Annular Mode (SAM) up to 2-3 seasons in advance (Lim et al. 2013), and the
Madden-Julian Oscillation up to 3 weeks in advance (Marshall et al. 2012). The forecast model
also faithfully reproduces the impact of these climate drivers on regional climate in Australia
(e.g., Langford and Hendon 2013). Therefore forecasts from P24 provide for skilful prediction
of some components of regional climate in Australia from weeks to seasons in advance
(Langford and Hendon 2013).
The POAMA seasonal forecast system consists of coupled models of the ocean, atmosphere
and land surface and an initialization/assimilation system that provides initial conditions with
perturbations so as to facilitate ensemble predictions that attempt to account for forecast
uncertainty arising from errors in the initial conditions. The P24 system uses a state of the art
ensemble assimilation/initialization system (Yin et al. 2011; Hudson et al. 2013; Zhao et al.
2014) and also accounts for forecast uncertainty due to model errors by running 3 different
versions of the model (Langford and Hendon 2013). However, the capability of the P24 system
is limited by the relatively coarse horizontal resolution of the component models (~250 km
horizontal resolution in atmosphere and 200 km in ocean; vertical resolution in both
component models is relatively coarse as well). The coarse resolution causes some aspects of
regional climate in Australia to not be well resolved. The coarse resolution of P24 also results
in some of the key climate drivers of relevance to Australian climate, such as the IOD and the
MJO, to not be faithfully represented. Furthermore, many of the physical parameterizations in
the component models in P24 are outdated, largely being based on model developments
upwards of 30 years ago.
The next version of the POAMA system is under development and will include upgrades to the
assimilation/initialization system and to the component models. A major change to the
POAMA system will be adoption of the model components provided by the ACCESS
program, which is a joint modelling partnership between the BoM, CSIRO, UKMO, and
Australian Universities. The first version of the ACCESS coupled model, referred to as CM1
(Bi et al. 2013), is based on the UKMO Unified Atmospheric Model equivalent to the versions
used in HadGEM2 or GA1, the AUSCOM ocean model that is derived from the GFDL
MOM4, the CICE sea ice model, and either the CABLE land surface model (ACCESS 1.3) or
the UKMO land surface model (ACCESS 1.0). More detail of these component models is
provided in Sec. 2. The atmosphere and ocean model resolutions in ACCESS CM1 are roughly
150km and 100km, respectively, which represent a marked increase over POAMA2.4 (250 km
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
3
and 200 km, respectively). Furthermore, ACCESS CM1 uses 38 vertical levels in the
atmosphere compared to 17 in POAMA2.4, meaning that the stratosphere, and tropospheric
boundary layer, and tropopause are much better resolved.
A preliminary version of POAMA has been developed based on ACCESS CM1 and some
seasonal forecast performance using a rudimentary initialization scheme has been reported as
part of the WIRADA project (Zhao et al. 2013). That assessment was based on a small set of
hindcasts that were initialized without making use of the full POAMA
assimilation/initialization scheme, and so the forecast performance of POAMA based on
ACCESS CM1 is probably a conservative indication of what might be achieved once the full
POAMA assimilation/initialization scheme is implemented. Nevertheless, a key finding of
that study is that at least for the first month of the forecast (before coupled initialization
matters) performance of POAMA based on ACCESS CM1 is expected to only be marginally
improved compared to POAMA2.4. The limited improvement of POAMA3 over POAMA2
was thought to be mainly due to an only modest increase in resolution for POAMA3 over
POAMA2 and the use of relatively older model components in ACCESS CM1. Zhao et al
(2013) further noted that some substantial biases, especially in the mean rainfall over the
Indian Ocean/Maritime Continent, remain in ACCESS CM1 that may hinder seasonal forecast
performance. And, although the ACCESS CM1 system experiences substantially less climate
drift than POAMA2.4, the representation of El Nino, the Indian Ocean dipole and the Madden
Julian Oscillation and their teleconnections to Australian climate were either on par or in some
instances worse than that in POAMA2.4. Substantial development of the UKMO Unified
atmospheric Model that occurred after configuration of ACCESS CM1 addressed some of
these shortcomings and Zhao et al (2013) suggested that these developments be trialled in
POAMA prior to finalizing a version of POAMA based on ACCESS. Furthermore, based on
experience at the UKMO with higher resolution models, Zhao et al (2013) also suggested
trialling POAMA seasonal forecasts systems with substantially higher resolution than used in
ACCESS CM1.
The development of a new, higher resolution version of the ACCESS coupled model is
ongoing and will use updated versions of the UM code. Work is also ongoing to implement the
POAMA assimilation/initialization scheme in the ACCESS CM1 model and forecast
performance will subsequently be investigated. In the meantime, the UKMO has moved to a
high resolution version of their coupled model (GC1) for their seasonal forecast system
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
4
(GLOSEA5), which now can be considered the state of the art for seasonal prediction
(MacLachlan et al. 2014). The GLOSEA5 system is based on more recent versions of the
UKMO component models than used in ACCESS CM1 (as well as using a different ocean
model). The UKMO intends to upgrade to GLOSEA6 using the most recent version of their
coupled model GC2 based on the Unified model GA6 in February 2015. Nevertheless, the
resolution of GLOSEA6 will be the same as GLOSEA5 system and this is much higher than
we can currently afford at BoM. However, new supercomputing upgrades are to become
available at BoM in late 2016, and the POAMA system should be able to be run at BOM at the
resolution of GLOSEA5/GLOSEA6. Hence, evaluation of the forecasts performance of the
GLOSEA5 system for seasonal prediction of Australian climate can provide a benchmark for
what might be achievable with a POAMA system based on the most recent model upgrades
available from the UKMO and run at the highest affordable resolution.
We report here on some comparison of forecast performance of the POAMA2.4 system with
the GLOSEA5 system based on available hindcasts for the period 1996-2009. The intent of
this comparison is to provide guidance for future development of POAMA based on improved
model components with higher resolution from the UKMO as provided by ACCESS. More
detail on POAMA2.4 and the GLOSEA5 systems, and their available hindcasts for study here
is provided in Sec. 2. Comparison of mean state biases is provided in Sec. 3. Assessment of
prediction of regional Australian climate is provided in Sec. 4. Assessment of prediction of key
drivers of Australian climate is provided in Sec. 5. Recommendations for future development
of POAMA, based on the findings of this study, are provided in Sec. 6.
2. MODEL CONFIGURATIONS AND HINDCASTS
a. POAMA2.4
POAMA is the BoM atmosphere-ocean coupled-model seasonal forecast system. The
atmospheric component of POAMA is the BoM’s Atmospheric Model version 3 (BAM3;
Colman et al. 2005), which is a spectral model with roughly ~ 250 km horizontal resolution on
17 vertical levels (T47L17). The land surface component carries three soil temperature levels
with simple bucket hydrology (Manabe and Holloway 1975, Hudson et al. 2011). The ocean
component is the Australian Community Ocean Model version 2 (Schiller et al. 2002), which
is based on the GFDL modular ocean model MOM2 (Pacanowski 1996). The ocean model has
horizontal resolution of 2° in the zonal direction and 0.5° in the meridional direction in the
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
5
tropics that increases to 1.5° towards the poles. ACOM2 has 25 vertical levels. BAM3 and
ACOM2 are coupled by the Ocean Atmosphere Sea Ice Soil (OASIS) software (Valke et al.
2000).
The POAMA version used in this study is version 2.4 (P24; e.g., Hudson et al. 2013), which
became the operational forecast model for seasonal outlooks in 2013. P24 forecasts are
initialised with realistic atmosphere and ocean conditions that are generated from separate
atmosphere/land surface (Hudson et al. 2011) and ocean (Yin et al. 2011) data assimilation
systems. The atmosphere-land surface assimilation consists of a nudging of the model
atmospheric fields (u, v, T, and q) towards an existing analysis every 6 hours. For the period
1980-2002, the ERA-40 reanalyses (Uppala et al. 2005) are used, and for the period 2003-
onward the analyses from the BoM’s operational numerical weather prediction are used.
Through this nudging process the land surface is realistically initialised in response to the
model generated surface fluxes.
The ocean assimilation is provided by the PEODAS system, which is a simplified form of an
Ensemble Kalman Filter (Yin et al. 2011). In situ temperature and salinity observations are
assimilated to a central run of the ocean model that is forced by re-analysis surface fluxes. An
ensemble of ocean states is generated by perturbing the surface fluxes within a possible range
of observational errors on intra-seasonal time scales. This ensemble of ocean states is used to
compute the background error covariances for temperature and salinity. Using these
background error covariances, observed data are assimilated to the central run. After each
analysis cycle (three days), the ensemble members are nudged toward the analysis from the
central run.
In P24, perturbed atmosphere-land-ocean initial conditions are created by running an ensemble
of coupled model simulations whereby the central member is constrained to the atmosphere-
land-ocean analysis from ALI and PEODAS. The other ensemble members are free to evolve
but are relaxed back to the central member once per day, with the strength of the relaxation
prescribed so as to allow the ensemble spread to span analysis uncertainty. In this fashion, an
ensemble of perturbed coupled initial conditions are created that span observational
uncertainty and which are suitable for generation of ensemble forecasts for multiweek-seasonal
lead times. Forecast uncertainty due to model error is addressed in P24 by running three
different versions of the coupled model, referred to as P24a, P24b and P24c. P24 uses an
ensemble of 33 members, whereby one unperturbed and 10 perturbed members are run from
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
6
each of the three model versions. Real time forecasts out to 270 days are initialized twice per
week and hindcasts are created every 5 days (6 times per month).
b. GLOSEA5
GLOSEA5 (MacLachlan. et al. 2014) is based on the UKMO coupled model HadGEM3,
which consists of the Unified Atmospheric Model version GA3 (Walters et al. 2011; Brown et
al., 2012), the JULES versions 3.0 Land surface ( Joint UK Land Environment Simulator
JULES; Best et al., 2011), the NEMO Ocean model version 3 (Madec, 2008; Megann et al.
2014), and the Los Alamos Sea Ice Model (CICE; Hunke and Lipscomb 2010) version 3.
Major upgrades over the GLOSEA4 system include improved model physics, a fully resolved
stratosphere, initialized sea ice and increased horizontal resolution.
A detailed description of the Global Atmosphere 3.0 configuration is given in Walters et al.
(2011). There have been numerous changes to the physical parametrizations since GA2 (and
its predecessor GA1 used in ACCESS CM1): introduction of cloud inhomogeneity, reduction
of spurious drizzle, reduction of spurious deep convection, introduction of the JULES land
surface model (Blyth et al., 2006), and the facility to read iceberg calving ancillary data. The
dynamical core is updated as well (called NewDynamics), which uses a semi-implicit semi-
Lagrangian discretization to solve the fully compressible, non-hydrostatic atmospheric
equations of motion. The stochastic physics scheme Stochastic Kinetic Energy Backscatter v2
is included to represent unresolved processes and provide small grid-level perturbations during
the model integration. Climate forcings for trace gases and insolation are set to observed
values up to the year 2005; after this point the emissions follow the Intergovernmental
Panel on Climate Change (IPCC) RCP4.5 scenario. Climatologies with a seasonal variation are
used for other aerosols (biogenic aerosols, biomass burning, black-carbon, sea salt, sulphates,
dust, and organic carbon fossil fuels). These climatologies have been generated from a climate
simulation using HadGEM2 (except dust which is from a HadGEM1a run). The observational
climatology is used for ozone, which includes a seasonal cycle.
The atmospheric resolution of HadGEM3 for GloSea5 has been increased from 1.88◦ longitude
×1.25◦ latitude in GLOSEA4 (and ACCESS CM1) to 0.83◦×0.56◦ (i.e. from approximately
150 km in midlatitudes to 50 km). The higher-resolution version of HadGEM3 requires a
reduced time step and altered diffusion settings to increase stability.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
7
HadGEM3 uses the updated NEMO ocean model version 3 run with horizontal resolution ¼
degree globally and 75 vertical levels (Megann et al. 2014). The higher horizontal resolution
means some of the major closed seas (Great Lakes, Lake Victoria, Caspian Sea and the Aral
Sea) are included. Increased horizontal resolution in the ocean model also means the western
boundary currents are better resolved leading to less bias in the mid latitude storm tracks
(Scaife et al. 2011). Tropical instability waves, which play a key role in the heat budget of the
equatorial Pacific cold tongue, are also better resolved , thereby leading to better of El Nino
and its teleconnections to the extratropics (Dawson et al. 2012). Vertical resolution of the
ocean model is also increased using 75 levels to better resolve the mixed layer and
thermocline (resolution is approximately 1 m near the surface and increases to 200 m at 6000
m depth). A detailed evaluation of the high resolution ocean model in GLOSEA5 will be
reported in a separate MCV study.
The initial atmospheric conditions for the GLOSEA5 forecast members are generated by the
Met Office operational numerical weather prediction (NWP) 4D-Var data assimilation system
(Rawlins et al., 2007). In hindcast mode, reanalyses from the European Centre for
Medium-Range Weather Forecasts (ECMWF) ERA-Interim project (Dee et al. 2011) are
used to initialize the atmosphere and land surface.
GloSea5 uses the Forecast Ocean Assimilation Model (FOAM) Ocean Analysis (Blockley et
al., 2013) to initialize the ocean and sea-ice components for the real-time forecasts. An
equivalent product, the GloSea5 Ocean and Sea Ice re-Analysis, a 23 year (1989 – 2011)
reanalysis, supplies initial conditions for the hind-casts. Both ocean analyses use the new
NEMOVAR (Mogensen et al., 2012; Waters et al., 2014) assimilation scheme developed
jointly by the UK Met Office, Centre Europe´en de Recherche et de Formation Avance´e en
Calcul Scientifique (CERFACS), ECMWF, and Institut National de Recherche en
Informatique et en Automatique/Laboratoire Jean Kuntzmann (INRIA/LJK).
The assimilation system (based on the NEMOVAR scheme) used to create the forecast and
hindcast analyses has the same ocean and sea-ice model (NEMO/CICE ORCA 0.25 L75) as
the coupled model used in GloSea5. The ERA-Interim atmospheric re-analysis (Dee et al.,
2011) is used as boundary forcing for the hindcast ocean reanalysis. NEMOVAR is a
multivariate, incremental 3D-Var first guess at appropriate time (FGAT) system. The system
implemented at the Met Office operates on a daily cycle with a 1 day time window and uses an
incremental analysis step. It assimilates satellite and in situ observations of sea-surface
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
8
temperature (SST), sea- level anomaly satellite data, sub-surface temperature and salinity
profiles, and satellite observations of sea-ice concentration. The temperature, salinity and sea-
level observations are assimilated in a multivariate fashion using balance relationships between
the variables (hydrostatic and geostrophic balance, plus preservation of the temperature –
salinity relationship in density), while the sea- ice concentration is assimilated as a univariate
field. Furthermore, bias correction schemes are implemented to reduce the bias inherent in
satellite measurements of SST (Martin et al., 2007; Donlon et al., 2012) and to reduce the bias
in the (supplied) mean dynamic topography correction required to convert measurements of
sea-level anomaly into sea-surface height (Lea et al., 2008).
Real-time forecasts from GLOSEA5 are created every day, whereby four members are
initialized from the 0000 UTC analyses from the NWP global data assimilation and the ocean
– sea-ice data assimilation system. Two of these members are run out to 210 days (seasonal
forecast members) and two are run out to 60 days (intraseasonal members). Spread between
members initialized on the same date is achieved only through the use of a stochastic physics
scheme (ie no perturbations to the initial conditions).
c. Monthly and Seasonal Hindcasts
For the analysis of month 1 and months 2-4 forecasts, output from the GLOSEA5 hindcasts
was used for the period 1996-2009 for forecasts starts on the 1st of February, May, August and
November, and on the 25th of each month prior. Four members are used from each start time.
We form an 8 member ensemble from GLOSEA5 by using the four forecasts from the 1st and
four forecasts from the previous 25th, throwing away the first 5-6 days of the forecasts. We
refer to the first month of the forecast as lead time 1 (or sometimes lead time 0, but usage
should be clear). We refer to the three-month mean formed from months 2-4 of these forecasts
as lead time 2-4. To facilitate comparison of monthly and seasonal forecasts with P24, we
make a similar 8 member ensemble from P24 using 4 members from a single model version
(P24a) on the 1st of the month and from the 26th of the preceding month.
d. Daily-Multiweek Hindcasts
For evaluation of daily-multiweek forecasts, we form an ensemble from GLOSEA5 as for the
seasonal assessment (eg 4 ensemble members from each start date) but use starts from the
25th, 1st and 9th, taking the 9th of the month as the first day of the forecasts (ie we throw
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
9
away the first 8 days of forecast from the start on the 1st and the first 13 days of forecast for
the start on the 25th). We note that for starts in May and November, GLOSEA5 hindcasts
provide 8 members for each start time. We include these additional 4 members for May and
November. For comparison with P24, we use the 10 member ensemble from the 11th of each
month from model version P24a.
3. MEAN STATE COMPARISONS
a. Australian Precipitation and Temperature
We begin by examining the mean precipitation (Fig. 1), Tmax (Fig. 2) and Tmin (Fig. 3) over
Australia from the GLOSEA5 and POAMA2.4 hindcasts for month 1 for times in February,
May, August and November for 1996-2009. Similar distributions are seen using lead times 2-4
(not shown). We compare to observations as provided by the 1/4 degree AWAP analyses of
Australian precipitation and temperatures (Jones et al. 2009). The higher resolution afforded
by GLOSEA5 is immediately apparent with Tasmania being resolved in contrast to it
appearing as ocean points in POAMA2.4 (only “land values” are plotted in Fig. 1, 2, and 3 and
so Tasmania is left blank in the P24 plots). For precipitation (Fig. 1), the strong dry bias over
northern Australia in the warm seasons in POAMA2.4 is greatly reduced in GLOSEA5, and
some key regional features, including the strong wintertime maximum on the western side of
Tasmania and along the Great Dividing Range in Victoria and New South Wales, the coastal
maximum along east coast, and the cool season maximum in south west Western Australia are
all resolved in GLOSEA5 and absent in POAMA2.4. Hence, the higher resolution of
GLOSEA5 greatly improves the depiction of mean precipitation throughout Australia.
For Tmax (Fig. 2), GLOSEA5 also provides improved depiction of regional features, including
Tasmania and the costal minimum along the east coast and Great Dividing Range. However,
GLOSEA5 typically suffers from a cold bias relative to the AWAP analyses and is no better
than POAMA2.4 in this regard. For Tmin (Fig. 3), POAMA2.4 exhibits a strong overall warm
bias, which is largely alleviated in GLOSEA5, implying that the diurnal temperature range is
much better depicted in GLOSEA5 as well as key regional features along the Great Dividing
Range and coastal regions. Overall, GLOSEA5 is clearly an enormous improvement over
POAMA2.4 for depiction of the mean climate over Australia.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
10
Fig. 1 Mean rainfall for February, May, August, and November. Top row is based on AWAP. Middle row is month 1 of GLOSEA5 hindcasts. Bottom row is month 1 of P24 hindcasts. Units are mm/day.
Fig. 2 Mean Tmax for February, May, August, and November. Top row is based on AWAP. Middle row is month 1 of GLOSEA5 hindcasts. Bottom row is month 1 of P24 hindcasts. Units are °C
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
11
Fig. 3 Mean Tmin for February, May, August, and November. Top row is based on AWAP. Middle row is month 1 of GLOSEA5 hindcasts. Bottom row is month 1 of P24 hindcasts. Units are °C
Fig. 4 Month 1 precipitation bias for all start months relative to GPCP (left column) and CMAP (right column) for GLSOEA5 (top row) and P24 (bottom row). Units are mm/day.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
12
b. Global Precipitation
The bias in global precipitation, using all 4 start months at lead time 1 month, is shown in Fig.
4 relative to the GPCP analyses (left hand column) and CMAP analyses (right hand column).
There are significant differences in the observed climatology based on these two analyses, with
the CMAP analyses generally being wetter in the mean over tropical latitudes than are the
GPCP analyses. Hence, the dry bias over the Maritime Continent in GLOSEA5 is more
prominent when assessed relative to CMAP than to GPCP. Similarly, the wet bias in the
western Indian Ocean in GLOSEA5 is more prominent when assessed relative to GPCP than to
CMAP. Overall, both GLOSEA5 and POAMA2.4 show large biases in the ITCZ’s, which
appear to result from positional errors, but in general, the biases in GLOSEA5 are less than in
POAMA2.4. For instance, the extreme dry bias over the Amazon in POAMA2.4 is absent in
GLOSEA5 and the overall biases in the tropical Pacific are lower in GLOSEA5. In summary,
GLOSEA5 has generally lower precipitation biases throughout the Tropics compared to
POAMA2.4 but significant biases remain in GLOSEA5 that will affect capability to predict
Australian climate (e.g the dry bias over the Maritime Continent means that the teleconnection
of the IOD to Australia will be underdone). That significant biases remain in GLOSEA5,
despite the use of high resolution and significant upgrades to model physics, emphasizes the
daunting task of alleviating precipitation biases. Continued development of the UM
atmospheric model, especially the parameterization of moist processes and clouds, is required.
c. Tropical SST Biases
Mean tropical SST is a key field to correctly simulate in order to faithfully predict ENSO and
its teleconnections. The mean biases at lead times 1 and 2 months are shown in Fig. 5 and at
lead times 3 and 4 months in Fig. 6. POAMA2.4 is well known to exhibit “global cooling”,
with significant tropical wide cold biases developing by lead time 2 and continuing to grow
through lead time 4. Importantly, however, POAMA2.4 exhibits little in the way of a tropical
eastern Pacific cold tongue bias (ie, the cooling is largely uniform). In contrast, GLOSEA5
exhibits little in the way of overall SST drift, which is a marked improvement over
POAMA2.4. However, GLOSEA5 suffers from a prominent, classical, cold tongue bias in the
eastern Pacific, which might affect depiction of the distinction of classical El Nino events from
Modoki or warm pool El Nino events. The cold tongue bias may also impact the depiction of
ENSO dynamics, thereby resulting in biases in the amplitude of El Nino variability (addressed
in Sec. 5 and in a forthcoming MCV report that focuses on GLOSEA5 ocean components).
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
13
Fig. 5 SST bias for months 1 (left column) and month 2 (right column) for GLOSEA5 (top row) and P24 (bottom row). Units are °C.
Fig. 6 As in Fig. 5 but for months 3 and 4.
4. AUSTRALIAN PRECIPITATION AND TEMPERATURE SKILL
We begin by assessing predictions of Australian precipitation and temperatures in the first
month of the forecast (lead time 1 month). We validate using correlation and accuracy score
against the AWAP ¼ degree analyses. We interpolate the AWAP analyses onto the respective
model grids so that the detail afforded by the high resolution in GLOSEA5 and low resolution
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
14
in POAMA2.4 are maintained. For correlation we use the ensemble-mean forecast anomaly
and correlate with observed anomaly. For accuracy score, we first create a probability forecast
for being above median. The accuracy score is defined as the number of hits plus correct
negatives divided by total number of forecasts, where an ensemble forecast is scored as a hit if
the ensemble forecast was for greater than 50% of the members predicting above median
rainfall and observed rainfall exceeded median. Similarly, a correct negative is if the ensemble
forecast was for less than 50% chance of exceeding median rainfall and observed rainfall was
less than median. Median thresholds for the forecast are based on the hindcast climatology,
while median values for verification are based on the observed climatology. In this fashion the
forecasts are “calibrated” to have correct median values. The ensemble forecast is formed by
counting the number of forecast members that predict above median rainfall and creating a
probability by dividing by the total number of ensemble members.
a. Correlation
We show correlation skill for the first month of the forecasts for precipitation (Fig. 7), Tmax
(Fig. 8) and Tmin (Fig. 9). Note now that the forecast skill over Tasmania is plotted for P24
results, even though there is no land point in P24 for Tasmania. This score reflects the P24
forecast for the ocean point that covers Tasmania. For precipitation, Tmax and Tmin,
GLOSEA5 exhibits a clear advantage over P24 across Australia, both in terms of higher
correlations and more spatial detail. Rainfall skill tends to be higher in both systems in the cool
seasons especially in the east. For Tmax and Tmin, GLOSEA5 exhibits uniformly higher skill
in all seasons.
Correlation skill for seasonal forecasts (months 2-4 of the forecasts) are shown for
precipitation (Fig. 10), Tmax (Fig. 11) and Tmin (Fig. 12). Forecast skill is lower than for the
first month of the forecasts, and precipitation and Tmin skill drops faster than Tmax skill for
both systems. Nonetheless, there are large areas of positive skill, with GLOSEA5 still
exhibiting distinctive advantages over P24. It is interesting to note than some areas of
relatively strong negative skill show up, which could be due to model systematic errors in
representing remote impacts of ENSO, for instance, or maybe due to sampling errors as a
result of the relatively short hindcast record analysed here (1996-2009).
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
15
Fig. 7 Anomaly correlation for precipitation for month 1 for February (top left), May (top right), August (bottom left) and November (bottom right) from GLOSEA5 and P24 hindcasts. Verification is against monthly AWAP ¼ degree analyses that was interpolated onto the native grids of each model.
Fig. 8 As in Fig. 7 except for Tmax.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
16
Fig. 9 As in Fig. 7 except for Tmin.
Fig. 10 As in Fig. 7 except for months 2-4 for precipitation
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
17
Fig. 11 As in Fig. 7 except for months 2-4 for Tmax.
Fig. 12 As in Fig. 10 except for months 2-4 for Tmin.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
18
b. Accuracy score
Accuracy scores (correct positive and correct negatives) for above median prediction of
precipitation (Fig. 13), Tmax (Fig. 14) and Tmin (Fig. 15) for the first month of the forecast
also reveal a distinctive advantage of GLOSEA5 over P24. Table summarizes the mean
accuracy score over all Australian land points. For month 1, GLOSEA5 accuracy typically
exceeds P24 accuracy by 5-10%. However, for months 2-4 (Figs. 16-18 and Table 1) the
advantage of GLOSEA5 over P24 is much more modest. Accuracy scores for month 2-4 for
Tmax (Fig. 17) are consistently better than for precipitation, with GLOSEA5 showing
consistent improvement over P24. In both systems, scores for Tmin (Fig. 18) are lower than
for Tmax.
Prc Tmax Tmin
e24a glosea5 e24a glosea5 e24a glosea5
0201 0.54 0.62 0.62 0.73 0.62 0.67
0501 0.54 0.62 0.65 0.74 0.64 0.69
0801 0.57 0.64 0.56 0.66 0.54 0.67
1101 0.59 0.63 0.65 0.70 0.61 0.67
MAM 0.54 0.54 0.60 0.60 0.53 0.52
JJA 0.51 0.54 0.55 0.60 0.51 0.51
SON 0.59 0.62 0.58 0.63 0.53 0.51
DJF 0.53 0.50 0.61 0.63 0.56 0.59
Table 1 Accuracy scores averaged over all Australian land points for month 1 and months 2-4.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
19
Fig. 13 Accuracy score for predicting precipitation above median for month 1 for February, May, August, and November for GLOSEA5 (top row) and P24 (bottom row). AWAP analyses, interpolated to native model grids, is used for verification. Units are percentage correct.
Fig. 14 As in Fig. 13 except for Tmax.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
20
Fig. 15 As in Fig. 13 except for Tmin.
Fig. 16 Accuracy score for precipitation above median for months 2-4 for MAM, JJA, SON, and DJF for GLOSEA5 (top row) and P24 (bottom row).
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
21
Fig. 17 As in Fig. 16 except for Tmax
Fig. 18 As in Fig. 16 except for Tmin.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
22
c. Reliability
Forecast reliability is assessed by constructing attributes diagrams based on predictions of
above median. Forecast probability bins of width 0.10 are used and compiled over all
Australian land points. Comparison of forecast reliability between the two systems is
problematic because the GLOSEA5 hindcasts have limited start times and members and so
reliability estimates are subject to sampling uncertainty (ie the plots are noisy). In order to
facilitate comparison of the two systerms, we chose to use only 8 members from P24 (4 from
the 1st and 4 from 26th) and from only one model version, which will severely reduce the
reliability attained in the operational P24 system that is based on a 33 member ensemble (11
members from 3 model versions) using 2 starts per week. Nonetheless, this comparison is
primarily to give an indication of how well the combination of lagged starts (25th and 1st of the
month) and stochastic physics is performing in GLOSEA5 for providing reliable forecasts.
Reliability for month 1 forecasts is shown in Fig. 19 for precipitation and in Fig. 20 for Tmax
(Tmin is similar). For this comparison, GLOSEA5 outperforms P24 for all 3 parameters and
for all months. This presumably stems from the use of stochastic physic in GLOSEA5 but
only one model version in P24. The reliability achieved with P24 using the full 33 member
ensemble based on 3 model versions is pronouncedly better than that depicted in Figs. 19 and
20. Nonetheless, the GLOSEA5 results are encouraging in that the combination of an
improved model, lagged starts and stochastic physics (even though ensemble is small) are
resulting in uniformly reliable forecasts.
The results from month 1 carry over to seasonal forecasts. We show results for spring season
in Fig. 21. GLOSEA5 forecasts of precipitation and Tmax exhibit good reliability, but Tmin is
overconfident (emphatic) and with little sharpness, perhaps reflecting the systematic error in
mean Tmin in GLOSEA5 discussed in Sec. 3. Reliability of the GLOSEA5 seasonal forecasts
is also uniformly better than P24a, again acknowledging that reliability of the POAMA system
that is run in operations is much better than depicted here due to the use of three model
versions and a larger ensemble.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
23
Fig. 19 Attributes diagram for predictions of above median rainfall for all land points over Australia or month 1 for February, May, August, and November from GLOSEA5 (top row) and P24 (bottom row). Shaded area indicates forecast contribute positively to the Brier skill score. Forecasts that fall on the sloping line would have perfect reliability. Size of dots is proportional to number of forecasts in each bin. Probability bins of width 10% were used.
Fig. 20 As in Fig. 19 except for Tmax.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
24
Fig. 21 Attributes diagram for SON (mnths 2-4) for above median prediction of precipitation (left hand column), Tmax (middle) and Tmin (right hand column) for all land points over Australia. GLOSEA5 is top row and P24 is bottom row. Plotting convention is as in Fig. 19.
d. Global context
Predictions of Australian precipitation are placed in a global context by examining forecast
skill across the Indo-Pacific. Both forecast systems show highest skill in over the tropical
Pacific, where ENSO variations dominate but GLOSEA5 shows a distinctive advantage over
P24 in the equatorial central and eastern Pacific. An example of the accuracy score for above
median rainfall for November is shown in Fig. 22. GLOSEA5 clearly outperforms P24
throughout the central and eastern Pacific, which presumably reflects better depiction of ENSO
variations in GLOSEA5 despite the cold tongue bias as discussed in Fig. 5. Importantly,
precipitation in this region affects tropical circulation and so can result in improved prediction
of the evolution of ENSO as well as remotely forced teleconnections, although probably not of
direct relevance to Australia because they are too far east. Forecast reliability for above median
rainfall across the Indo-Pacific domain is good in both systems, with GLOSEA5 again
showing a slight advantage.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
25
Fig. 22 Accuracy score for above median rainfall for GLOSEA5 and P24 verified against CMAP for month 1 for November (top row). Attributes diagram for above median rainfall using all points over Indo-Pacific domain is shown in bottom row.
5. CLIMATE MODES AND THEIR TELECONNECTIONS
The capability to predict some key modes of climate variability and some of their
teleconnections, specifically ENSO, the IOD, SAM and the MJO are now assessed
a. ENSO and IOD
Forecast skill for ENSO and the IOD are assessed using monthly means for all 4 available start
months from GLOSEA5 (Feb, May, Aug, and Nov). Correlation (top row) and root mean
square error (rmse; bottom row) for the Nino34 index (left column) and DMI (right column)
are shown in Fig. 23. GLOSEA5 shows slightly improved correlation but slight worse rmse for
Nino34 as compared to P24. GLOSEA5 shows a distinct improvement for correlation of IOD
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
26
compared to P24 but similar rmse. The discrepancy between improved correlation in GLOSEA
but worsened or similar rmse is explained by the simulation of the amplitude of these indices
in the respective systems (Fig. 24, based on the mean standard deviation using each ensemble
member). The GLOSEA5 system simulates stronger then observed Nino34 and IOD
amplitude, whereas the P24 system is slightly damped. Hence, the improved correlations in
GLOSEA5 indicate better depiction of temporal variability, but the worse rmse indicates
poorer performance with amplitude. The overestimation of the Nino34 index in GLOSEA5 is
symptomatic of ENSO variability that is too strong in the eastern Pacific, as seen in Fig. 25
that shows amplitude of Nino4 in GLOSEA is only slightly above normal but Nino3 is much
stronger than observed. This stronger-than-observed ENSO variability in GLOSEA5 may stem
from the strong cold tongue bias in GLOSEA (eg. Figs. 5 and 6) but further investigation is
required.
Fig. 23 Anomaly correlation (top row) and rmse (bottom row) for monthly Nino34 and IOD indices using all start months. Correlation for persistence forecast is included.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
27
Fig. 24 Predicted amplitude of the Nino34 and IOD indices from GLOSEA5 and P24 for all start months. Amplitude was computed using individual members and then averaged over all members.
Fig. 25 As in Fig. 24 except for Nino4 and Nino3 indices.
Similarly, for the IOD, GLOSEA5 simulates stronger than observed behaviour which is
primarily due to over-growth of the IOD during the austral winter season (Fig. 26). GLOSEA5
faithfully captures the peak amplitude of the IOD in austral spring and the demise of the IOD
in summer. The stronger than observed IOD variability in GLOSEA5 from early winter might
stem from the rainfall bias in the Indian Ocean (Fig. 4), which would promote an easterly wind
bias in the eastern Indian Ocean, resulting in a spuriously shallow thermocline along
Java/Sumatra, thereby acting to make SST in the eastern pole of the IOD too sensitive to
thermocline variations. However, this speculation needs to be further investigated.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
28
Fig. 26 Amplitude of the IOD index for forecasts starting in May, August and November.
The improved predictions of Nino34 in GLOSEA5 (Fig. 23) are seen to primarily occur for
forecasts initialized in May (Fig. 27). This is an outstanding improvement over P24 because
forecast starts in May are at the time of the year when forecast skill for ENSO is the lowest (ie
the "spring predictability barrier"). For this reason alone GLOSEA5 appears to offer a
substantial advantage over P24 for predicting Australian climate because ENSO impacts in
Australia are strongest in winter and spring and so improved prediction of ENSO from May
suggests potential increasing the lead time of skilful predictions of Australian climate.
Of course, it is not sufficient to simply predict the SST anomalies associated with ENSO and
IOD in order to translate into skilful prediction of regional climate: the teleconnections of
ENSO must also be well captured. The depiction of the tropical rainfall anomalies during
ENSO and IOD are shown in Fig. 28 for the SON season based on forecast starts on 1 August.
The rainfall anomalies are shown as the regression onto the predicted Nino34 and IOD indices.
Compared to the observed anomalies (top row of Fig. 28), GLOSEA5 shows a westward shift
of the primary anomaly in the western/central Pacific, despite the fact that the ENSO SST
anomaly in GLOSEA5 shows limited westward bias (Dawson et al. 2012). This westward shift
of the rainfall probably stems from the cold tongue bias in GLOSEA, so that although the SST
anomaly during ENSO is well predicted, the cold tongue bias will act to shift the rainfall
anomalies westward. P24 shows a similar westward bias of the rainfall anomaly, despite the
fact that the P24 ENSO SST anomaly is shifted west (Hendon et al. 2009). However, P24 has
little cold tongue bias (rather the SST bias is more uniform cooling), so the mean state bias in
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
29
P24 will not act as strongly as in GLOSEA5 to shift the ENSO rainfall anomaly westward.
However, P24 shows limited capability to capture the contracted ITCZ in the eastern Pacific
during El Nino, whereas this feature is well captured in GLOSEA5. The opposite signed
rainfall anomaly in the Maritime Continent and eastern Indian Ocean is weaker in GLOSEA5
than observed or in P24. This is further highlighted by the weaker than observed rainfall
anomalies in GLOSEA5 in these regions associated with the IOD. These weaker anomalies
may reflect the dry bias in this region in GLOSEA5, but further investigation is required.
The rainfall teleconnections to Australia are shown in Fig. 29 for the key winter (JJA) and
spring (SON) seasons when ENSO and IOD impacts are most significant. Both GLOSEA5
and P24 under-represent the impacts of ENSO and IOD during winter, however, it is
encouraging to see a good depiction of the IOD impact along the Great Dividing Range in the
high resolution GLOSEA5. In spring, the observed impacts of ENSO and IOD are stronger and
similar in both systems, although the stronger signal in the SE associated with the IOD
compared to ENSO during SON seems better depicted in P24. This might reflect a better
depiction of the tropical IOD rainfall anomalies in P24 compared to GLOSEA5, which are the
source of the teleconnection in SE Australia depicted in Fig. 29.
Temperature teleconnections of ENSO and IOD for Tmax and Tmin during SON, which is the
season of strongest impact, are shown in Fig. 30. For Tmax, both systems show equally good
teleconnections (strong warm anomalies during both El Nino and IOD), but GLOSEA5 does
provide some improved distinction of the impacts between the far-east and central Australia. In
contrast, for Tmin, GLOSEA5 shows a serious error, with colder than normal Tmin being
depicted across most of Australia during El Nino and IOD, when in reality Tmin should be
above normal across the southern 2/3 of the country. This is well captured in P24. The lack of
land surface initialization in the GLOSEA5 hindcast set may contribute to this error, but
further investigation is required. The sense of the error (Tmin too low during dry condition due
to El Nino and IOD) may indicate too much night time cooling (ie due to a dry land bias)
compared to subsidence warming that would act to keep Tmin higher than normal during dry
El Nino/IOD conditions.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
30
Fig. 27 Anomaly correlation for the Nino34 index for forecasts starting in May.
Fig. 28 Regression onto Nino34 index of observed rainfall (top row) and ensemble mean rainfall from GLSOEA5 (middle) and P24 (bottom) forecast for months 2-4 for SON. Units are mm/day.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
31
Fig. 29 Regression onto Nino34 and IOD indices of observed rainfall (top), GLOSEA5 (middle) and P24 (bottom) for months 2-4 for JJA and SON. Units are mm/day.
Fig. 30 As in Fig. 29 except for Tmax and Tmin in SON. Units are C.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
32
b. SAM
Skill for predicting seasonal variations of the SAM are assessed using the Gong and Wang
(1999) SAM index, which is based on the difference of zonal mean surface pressure between
40S and 65S. We concentrate on the spring season when seasonal skill for the SAM is the
highest. The skill for predicting November- mean SAM from 1 Nov and from 1 Aug is shown
in top row of Fig. 31. Forecast skill is comparable from each system but with GLOSEA5 being
slightly better than P24 from 1 Nov and and P24 better than GLOSEA5 from 1 Aug. The
prediction of seasonal SAM for the SON season from 1st Aug is shown in the lower panel of
Fig. 31. As previously reported in Seviour et al. (2014), GLOSEA5 can provide skilful
predictions of seasonal SAM during SON with at least one month lead time. P24 is marginally
worse, which Seviour et al. (2014) attribute to the lack of resolution in the stratosphere in P24.
However, examination of the time series of observed and predicted SAM indices for
November and SON (Fig. 32) suggest that the improved depiction of the stratosphere in
GLOSEA5 may not be of primary importance for improved prediction of the SAM. For
November SAM predictions from 1st Nov (top of Fig. 32), the predictions from P24 track very
closely the predictions from GLOSEA5. In fact, the correlation of the P24 SAM predictions
with GLOSEA5 is 0.87, which is much higher than either is with reality. This is also true for
the SON prediction from 1st August (correlation P24/GLOSEA5= 0.74). This suggests
predictability is similar in both systems, but can’t be coming from the stratosphere as it is
absent in P24.
Importantly, both systems over-predict low SAM in November 2002, which followed the
strong stratospheric warming and extreme negative SAM in September-October 2002. Seviour
et al. (2014) speculate that improved initialization of the stratosphere should provide
predictability of SAM in November, but without a resolved stratosphere P24 equally over-
predicts the SAM in November 2002. The observed SAM value in October 2002 was strongly
negative, and so both systems may have over-represented persistence in their November 2002
forecasts. We also note that P24 provides a better prediction of the low SAM seasonal value in
SON 2002 after the stratospheric warming in September 2002 than does GLOSEA5, again
suggesting that the resolved stratosphere is probably not the source of predictability.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
33
Fig. 31 Anomaly correlation for predicting SAM index for November using forecasts initialized on 1st November and 1st August (top). Persistence forecast is shown in blue. Anomaly correlation for months 2-4 (SON) for forecasts initialized on 1st August (bottom).
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
34
Fig. 32 Time series of November SAM index (top) and SON SAM index (bottom) from observations and forecasts.
The strong correlation of the ensemble-mean predictions from GLOSEA5 and P24 at short
lead time suggest some other source of predictability is acting in the forecast models (perhaps
spuriously) but it is not due to a resolved stratosphere. No doubt some of the strong
correlation of the ensemble mean predictions from the two systems reflects the elimination of
unpredictable noise by forming ensemble means. Research is ongoing to explore whether
individual forecast members are no better correlated from each system than with reality, which
then could explain the strong correlation of the two ensemble mean forecasts with each other.
The SAM decorrelates with about a 10 day timescale so it is of interest to investigate the daily
predictions of the SAM in the two systems. The GLOSEA5 hindcast set is not optimal for
assessing daily predictions because the start dates are 7-8 days apart and there are only 8
(maximum) ensemble members available for each start date but without any perturbations
(recall GLOSEA5 relies on stochastic physics to provide spread from an individual start time).
Therefore, we assess three sets of ensembles from GLOSEA5 that contain i) 8 members from
the 9th of the month, ii) 8 members from the 9th and 8 members from the 1st, and iii) 8 members
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
35
from the 9th, 8 members from the 1st and 8 members from the 25th. We use the 10 member
ensemble from the 11th of each month from P24a. Use of less ensemble members for P24 at
longer lead times as compared to GLOSEA5 means that the comparison to GLOSEA5 is not
necessarily fair, but for most of our purpose here, this should not be an issue.
Correlation of the SAM index based on all available start times is shown in Fig. 33. We also
include comparison to the 10 member lagged ensemble from the ACCESS1.3 version of
POAMA (Zhao et al. 2014), which is based on a lag spanning 2.5 days (ie 10 members that are
each 6 hours apart). For GLOSEA5, the effect of using a large-span lagged ensemble is
clearly seen, with forecast skill dropping in going from 8 members from the 9th, 8 members
from the 1st and the 9th, and 8 members from the 25th, 1st and 9th. GLOSEA5 is seen to
provide the most skilful predictions of the SAM for forecasts initialized on the 9th (ie
correlation of 0.5 is achieved at 12-13 days in GLOSEA5 but only at 10 days in P24).
However, for the lagged forecasts, GLOSEA5 is worse than P24 for the first 3-6 days.
More insight into the impact of using long lags in the GLOSEA5 ensemble is gained from
looking at rmse and spread (Fig. 33). For forecasts from a single start on the 9th (top panel in
Fig. 33), ensemble spread in GLOSEA5 is small and grows slowly, only achieving the spread
from P24 (which uses only 1 start time but with perturbations) after ~20 days. Furthermore,
GLOSEA5 is severely under-dispersive (spread is much smaller than rmse). For the lagged
ensemble combining starts on 9th and 1st the spread from GLOSEA is now comparable to P24
from a single start date, with spread roughly tracking rmse. For the lagged ensemble
combining starts on the 25th, 1st and 9th, GLOSEA5 is over-dispersive, and this is reflected in
reduced forecast skill. We also compare to the spread from the lagged ensemble from
ACCESS-POAMA, which is based on 10 members lagged by 6 hours each (so up to 2.5 days).
The ACCESS-POAMA spread is seen to be smaller than P24, as documented in Zhao et al.
(2013), but is seen to be greater than GLOSEA5 based on 1 start time but less than GLOSEA5
based on 2 start times that are 7-8 days apart.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
36
Fig. 33 Correlation (left) and rmse/spread (right) for predicting daily SAM index using all start months for days 1-35. Top row is for forecasts initialized on 9th, middle row combines forecasts from 1st and 9th and bottom row combines 25th, 1st, and 9th.
From 9th: Correlation RMSE & spread
From 1st & 9th: Correlation RMSE & spread
From 25th, 1st & 9th: Correlation RMSE & spread
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
37
In summary, stochastic physics is not an effective mechanism to generate spread for the first
~3 weeks of forecast but appears to be adequate for seasonal forecasts. A lag of ~8 days as
afforded by the GLOSEA5 hindcasts does produce adequate spread (ie spread is comparable to
rmse) but at the expense of drop in forecast skill. The conclusion is that the ensemble strategy
in GLOSEA5 hindcast is not optimal for generating mutliweek forecasts but may be acceptable
for seasonal forecasting. And, the assessment of reliability and spread/rmse in GLOSEA5
based on the hindcasts may not give a true indication of the capability of the real time
GLOSEA5 system, which uses daily lags to form the ensemble, and so by forming ensemble
bases on a few days start times can produce sufficient spread without degrading forecast skill.
However, there are clear advantages of the perturbation approach used in P24, so this
technique should be explored in conjunction with stochastic physics in future versions of
POAMA based on updated model code from UKMO.
We also explored assessing SAM forecasts by projecting predicted sea level pressure (SLP)
onto the leading EOF of monthly mean SLP from observations, which depicts the SAM (e.g.,
Lim et al. 2013). The comparison of forecast scores using the Gong index and the observed
SLP EOF is shown in Fig. 34. Forecast scores for P24 and ACCESS-POAMA are similar
using both methods but the score drops dramatically in GLOSEA5 when the observed EOF is
used. This suggests that there is a significant bias in the structure of the SAM anomalies in
GLOSEA5, but further work is required to fully understand its cause.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
38
Fig. 34 Comparison of SAM forecast skill using Gong SAM index (top) and by projecting onto observed EOF of SLP (bottom).
c. MJO
Forecast skill for the MJO is assessed by projecting forecast output of u200, u850, and OLR
onto the leading pair of observed EOFS as defined by Wheeler and Hendon (2004). A bivariate
correlation (e.g., Marshall et al. 2012) is computed for November and August starts (Fig. 35).
For November starts from the 9th (and 1st and 9th), GLOSEA5 and P24 are comparable (ie
correlation of 0.5 is achieved out to ~25 days) but GLOSEA5 based on starts from 25th, 1st, and
9th displays a drop in skill after 10 day. This again indicates that the lag in the GLOSEA5
hindcasts is too large for multiweek forecasts. We also see that GLOSEA5 from starts on the
9th and on the 1st and 9th is appreciably better than ACCESS1-POAMA in November
(ACCESS1 drops to 0.5 correlation at about 15 days compared to 25 days for GLOSEA5).
Gong Index
SLP EOF
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
39
Fig. 35 Bivariate correlation for Wheeler-Hendon MJO index for November starts (left) and August starts (right).
From 9th: Nov Start Aug Start
From 1st & 9th: Nov Start Aug Start
From 25th, 1st & 9th: Nov Start Aug Start
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
40
Fig. 36 As in Fig. 35 except for rmse and spread for November forecasts.
From 9th
From 1st & 9th
From 25th, 1st & 9th
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
41
This increase in skill of GLOSEA5 compared to ACCESS1 is a good outcome for
development of the UKMO models because the MJO was poorly simulated and predicted in
the earlier versions of the UKMO atmospheric models model ( upon which ACCESS1-
POAMA was based). We also see that GLOSEA5 is better than P24 for August starts (and
better than ACCESS-POAMA), which suggests the potential for improved prediction of Asian
summer monsoon variability.
d. Blocking
Skill for predicting atmospheric blocking is assessed using a simple index developed by the
Bureau of Meteorology for use in the Australian region. The index describes the degree of
splitting of the 500 hPa westerly airstream (Pook and Gibson 1999). Here we focus on
blocking at 140°E, which is regarded as a typical longitude for blocking in the Australian
region with widespread influences on seasonal (Risbey et al. 2009) and intra-seasonal
(Marshall et al. 2014) rainfall variability over the continent. Correlation score for prediction of
the blocking index is shown in Fig. 37 (left column). The effect of using a large-span lagged
ensemble is clearly seen over the first ~9 days of the forecast, with skill again dropping in
going from 8 members from the 9th, 8 members from the 1st and the 9th, and 8 members from
the 25th, 1st and 9th. GLOSEA5 provides the most skilful predictions of blocking out to about
15 days initialized on the 9th, however for the lagged forecasts using all three start dates
GLOSEA5 is no better than P24.
RMSE and spread (Fig. 37, right column) again reflect the impact of using long lags in the
GLOSEA ensemble, and bear similar results to those seen for the SAM analysis in Fig. 33. For
forecasts from a single start on the 9th, GLOSEA5 is severely under-dispersive (spread is much
smaller than rmse, and smaller than spread in P24 and ACCESS-POAMA), whereas for the
lagged ensemble combining starts on the 25th, 1st and 9th, GLOSEA5 is over-dispersive. For the
lagged ensemble combining starts on 9th and 1st, the spread from GLOSEA is comparable to
P24 and greater than ACCESS-POAMA, and spread is also roughly equal to rmse. As for the
SAM and MJO analyses, these results for blocking indicate that the lag in the GLOSEA5
hindcasts is too large for multiweek forecasts.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
42
Fig. 37 As in Fig. 33 except for predicting the daily blocking index using all start months for days 1-30.
From 9th: Correlation RMSE & spread
From 1st & 9th: Correlation RMSE & spread
From 25th, 1st & 9th: Correlation RMSE & spread
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
43
The increase in skill out to about 15 days for GLOSEA5 compared to P24 and ACCESS-
POAMA is a good outcome for development of the UKMO models because weather and
climate models have known deficiencies in simulating the observed location and frequency of
blocking highs, in both hemispheres. This can be linked to the climatological biases of the
models, as shown for P24 (Marshall et al. 2014) and for a predecessor to GLOSEA5 (Scaife et
al. 2010), however it is unclear whether vertical, and in particular stratospheric, resolution is
also important for a model’s depiction of blocking (Scaife and Knight 2008, Woollings et al.
2010). Further work is required to fully understand the impact on blocking of mean state biases
and increased stratospheric resolution in GLOSEA5, compared to P24 and ACCESS-POAMA.
6. CONCLUSIONS AND RECOMMENDATIONS FOR POAMA DEVELOPMENT
Based on comparison of available hindcasts for the period 1996-2009 from the high resolution
GLOSEA5 system, we have come to a number of conclusions regarding the expected benefit
of a GLOSEA5-like system over the current P24 system. We have also uncovered a few
shortcomings of the GLOSEA5 system that could benefit from further development of the
model and ensemble systems. Taken together, these form the basis for some recommendations
on the development of future versions of POAMA.
a. Expected benefits of the high resolution GLOSEA5
1. Improved resolution of Australian climate and reduced drift
The increased horizontal resolution of GLOSEA5 shows a clear advantage over the existing
POAMA2.4 system for resolving Australian climate, including depiction of Tasmania, coastal
contrasts, and impacts of the Great Dividing Range. Benefits of high ocean model resolution
are being assessed in a separate study. Adoption of a high resolution POAMA system would
thus be expected to result in provision of improved seasonal forecast products on spatial and
temporal scales that are demanded by POAMA user, which include stream flow forecasts and
crop models that are driven by downscaled precipitation and temperature forecasts.
GLOSEA5 also exhibits reduced dry bias over the Australian continent, meaning that the
model will more realistically simulate impacts of climate variability on regional rainfall,
especially the north-south contrast during the summer monsoon. GLOSEA5 has appreciably
less drift than P24 so that predicted climate variability at longer lead time will be more
faithfully represented.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
44
GLOSEA5 delivers appreciably better regional forecasts of precipitation and temperature over
Australia in the first month of the forecast, with more modest gains over P24 for months 2-4.
In this study, we are not able to conclusively attribute this benefit to increased resolution or
other model improvements afforded by GLOSEA5 but the outcome is more important than the
cause. Similarly, GLOSEA5 exhibits acceptable forecast reliability, even based on the
inadequate hindcast ensemble that was available for this study. Again, we do not know how
much contribution to good reliability comes from good forecast skill and how much comes
from the ensemble methods (combination of lagged and stochastic physics). However, we may
anticipate even bigger improvements in forecast performance with GLOSEA5 once the
POAMA assimilation/ensemble generation strategy is implemented (or if lagged ensembles
were formed with shorter lags than provided by the hindcast set).
2. Improved prediction of climate modes
GLOSEA5 exhibits a substantial improvement over P24 in long lead prediction of ENSO/IOD
and ENSO-impacts on tropical rainfall, which is fundamental to improved long lead prediction
of Australian climate because ENSO and the IOD are the greatest source of predictable
Australian climate variability. The increased skill in predicting IOD and ENSO from late
autumn is one of the most outstanding advantages of the GLOSEA5 over P24. The
improvements in predicting and simulating ENSO and its impacts in GLOSEA5 was attributed
by Dawson et al (2013) to increased resolution of both the atmosphere and ocean models used
in GLOSEA5. However, the reason for the pronounced improvement over P24 in autumn
needs further investigation. Furthermore, the improved prediction of the IOD over P24 needs
investigation because GLOSEA5 over-estimates the IOD amplitude for forecasts in the crucial
autumn season when the IOD is most rapidly growing. GLOSEA5 also exhibits improved
prediction of the MJO during winter, which is another outstanding advantage over P24. The
improvements in representation and prediction of the MJO in GLOSEA5 compared to
GLOSEA4 (and compared to ACCESS1) was attributed to atmospheric model improvements
afforded by UM GA3 (MacLachlan et al. 2014), primarily due to improvements to convective
parameterization.
GLOSEA5 also provides for improved short lead prediction of the SAM and blocking (ie for
days 1-15), which could translate into improved week 2 and week 3 forecasts for Australian
rainfall and temperatures due to their strong influences on regional Australian climate. The
reason for improved short lead prediction of the SAM was speculated to be due to improved
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
45
atmospheric initial conditions afforded by GLOSEA5 as compared to P24. However,
GLOSEA5 only marginally beats P24 for seasonal prediction of the SAM, which we speculate
is not due to improved resolution of the stratosphere because GLOSEA5 forecast correlated
much more strongly with P24 forecasts (that don’t have a stratosphere) than either does with
reality, but further work is required to verify this speculation. The improved prediction of
blocking also needs investigation. For instance, does it stem from improved representation of
the climatology of blocking and if so, is this due to increased resolution?
b. Areas requiring development in GLOSEA5
Although GLOSEA5 exhibits many outstanding improvements over P24, we have uncovered a
few shortcomings of the GLOSEA5 system that could benefit from continued development.
1. Too short hindcast: Although not a model deficiency, the available hindcast set from the
GLOSEA5 system (1996-2009, four start months per year and lagged over 3 start days
with 4 (or 8) members each) is too short to fully assess predictions of extremes and the
full gamut of ENSO variability. It is also too short to calibrate follow-on predictions
(e.g., streamflow forecasts). The short hindcast set stems from the trade-off of going to
a higher resolution (more expensive) system that is updated frequently. But, the
possibility of sharing hindcasts across institutions could be considered in order to create
a longer hindcast set that would be more useful for end-users.
2. Cold tongue bias: GLOSEA5 exhibits a pronounced tropical Pacific cold tongue bias.
Although there is little overall drift in GLOSEA5, the cold tongue bias affects ENSO
teleconnections and so can be contributing to reduced forecast skill over Australia.
Mediating the cold tongue bias is a challenging problem but could be pursued jointly
between CAWCR researchers and the UKMO for inclusion in a future version of
POAMA based on UKMO/ACCESS coupled models.
3. ENSO amplitude too strong: ENSO amplitude, as measured by the variability of the
Nino3 index is substantially overestimated in GLOSEA5, perhaps due to the cold tongue
bias. Too strong ENSO variability will act to degrade Australian climate forecasts
because of the primary role that ENSO plays in Australian climate. Tackling the ENSO
amplitude problem is challenging and might also be best pursued jointly between
CAWCR and UKMO scientists.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
46
4. ENSO teleconnection to Australian Tmin: GLOSEA5 simulates negative Tmin
anomalies over Australia during El Nino, which is a serious error (it should be warmer
over Australia during El Nino). Further work is required to understand what is
contributing to this error (eg lack of soil initialization or a dry bias leading to too much
longwave cooling) but needs to be mediated in order to deliver good forecasts of Tmin
over Australia during winter and spring.
5. Indian Ocean-Maritime Continent Rainfall Bias: GLOSEA5 has a distinctive wet bias in
the tropical western Indian Ocean and dry bias in the eastern Indian Ocean/Maritime
Continent, which can contribute to degrading forecast skill of the MJO and the IOD and
their remote impacts in Australia. This rainfall bias is recognized as a serious issue and a
PEG at the UKMO, led by CAWCR researchers, has formed to address it.
6. Non-optimal ensemble strategy: The lagged/stochastic physics approach to ensemble
generation for GLOSEA5 hindcasts, while apparently adequate for seasonal forecasts, is
inadequate for multiweek forecasts (ie too little spread is generated, or sufficient spread
is generated at the loss of forecast skill, thereby resulting in either over-emphatic
forecasts or degraded accuracy). Although the reliability of the real time GLOSEA5
system may be better than the hindcasts set up, we cannot assess this. Clearly adoption
of the POAMA assimilation/ensemble generation into the GLOSEA system would
alleviate some of this problem; however, other simpler methods of perturbing the initial
conditions could also be explored.
7. SAM bias: GLOSEA5 apparently has a biased depiction of the spatial structure of the
SAM, as evidenced by lower skill scores when observed EOF patterns are used to verify
SAM rather than a simple zonal mean index of pressure differences. The nature and
cause of the bias needs to be explored.
Most of the deficiencies in the GLOSEA5 system outlined above stem from deficiencies in the
atmospheric model, which is similar (but more updated) to ACCESS. Tackling these
deficiencies should be a priority for future development projects.
c. Recommendations for POAMA development
A clear outcome of this intercomparison is that the future version of POAMA should be based
on a high resolution coupled model that is comparable in performance to GLOSEA5.
GLOSEA5 sits apart from all other seasonal forecast systems that were operational in 2014,
being by far the highest resolution system. GLOSEA5 is appreciably higher resolution
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
47
(vertical and horizontal directions) in both atmosphere and ocean model components than the
current ACCESS1 coupled model.
GLOSEA5 also exhibits some significant improvements over models based on earlier versions
of the UKMO atmospheric model, such as ACCESS1.0 (e.g., used in the POAMA3
prototype), including improved depiction of the MJO, improved regional climate over
Australia, and better simulation and prediction of ENSO. Future development of POAMA
should take advantage of these recent developments at the UKMO that went into the
component models of GLOSEA5 (and its successor), including the increased capability
afforded by running both high resolution atmosphere and ocean components (Dawson et al.
2012). Importantly, the ACCESS-POAMA program should position itself so that
improvements to the component models, either developed at UKMO or locally, will get into
future versions of the seasonal forecast model in a more timely fashion.
The stochastic physics option used in GLOSEA5 appears to offer some benefit for seasonal
ensemble generation. This option has yet to be implemented in the ACCESS coupled model.
Future versions of POAMA should implement this option, but further work is also needed to
quantify the additional benefit over that achieved from the POAMA perturbation approach.
The possibility of an expanded collaboration with the UKMO that would be facilitated by
using the same updated coupled model could be explored, for instance, by sharing generation
of a larger hindcast set (members and earlier start dates). A larger hindcasts set is critical to
allow for proper assessment of the potential benefit of seasonal forecasts for agricultural and
other applications. A larger hindcast set would also allow better sampling of predictive
capability of the “flavours” of ENSO and their regional impacts, as well as better sampling of
predictive skill of extreme events, such as heat waves and floods.
Finally, the greatest forecast benefit of GLOSEA5 over P24 occurred in the first month (and
even first 2 weeks) of the forecast. To fully capitalize on this benefit, operational seasonal
climate prediction, with whatever forecast model we end up using, should be done with the
shortest lead possible. Forecast skill drops off monotonically with lead time, so minimizing
the lead time is a guaranteed method to increase forecast skill.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
48
7. REFERENCES
Bi, D., and co-authors, 2013: The ACCESS Coupled Model: Description, Control Climate and
Evaluation. Aust. Met. Oceanog. J. 62,4.
Brown, A., S. Milton, M. Cullen, B. Golding, J. Mitchell, and A. Shelly, 2012: Unified
modeling and prediction of weather and climate: a 25-year journey. Bull. Amer. Meteor. Soc.,
93, doi: 10.1175/BAMS-D-12-00018.1.
Colman, R., L. Deschamps, M. Naughton, L. Rikus, A. Sulaiman, K. Puri, G. Roff, Z. Sun, and
G. Embery, 2005: BMRC Atmospheric Model (BAM) version 3.0: Comparison with mean
climatology. BMRC Research Report No. 108, Bureau of Meteorology Research Centre, 32 pp.
(available from http://www.bom.gov.au/bmrc/pubs/researchreports/researchreports.htm).
Dawson A, Matthews AJ, Stevens DP, Roberts MJ, Vidale P. , 2012. Importance of oceanic
resolution and mean state on the extra-tropical response to El Nino in a matrix of coupled
models. Clim. Dyn. 41: 1439–1452, doi: 10.1007/s00382-012-1518-6.
Dee, D. P. and co-authors, 2011: The ERA-Interim reanalysis: configuration and performance
of the data assimilation system. Quart. J. Roy. Meteorol. Soc. 137, 553–597.
Gong, D., and Wang, S., 1999: Definition of Antarctic Oscillation index. Geophys Res
Lett 26(4): 459–462. doi:10.1029/1999GL900003.
Hendon, H.H., E. Lim, G. Wang, O. ALves, and D. Hudson, 2009: Prospects for predicting two
flavors of El Nino . Geophys. Res. Lett.,36, L19713, doi:10.1029/2009GL040100
Hudson, D, A G. Marshall, Y Yin, O Alves, H H. Hendon, 2013: Improving Intraseasonal
Prediction with a New Ensemble Generation Strategy. Mon Wea Rev 141: 4429–4449
Hudson, D., Alves, O., Hendon, H.H. and Wang, G., 2011: The impact of atmospheric
initialisation on seasonal prediction of tropical Pacific SST. Climate Dynamics. 36, 1155-1171.
Hunke, E.C., and W. H. Lipscomb, 2010: CICE: The Los Alamos sea ice model documentation
and software users manual, version 4.1. Technical Report LA-CC-06-012, Los Alamos National
Laboratory.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
49
Jones, D.A., Wang, W. and Fawcett, R. 2009: High-quality spatial climate data sets for
Australia. Aust. Met. Oceanogr. J., 58, 233-248.
Langford, S., and H.H. Hendon, 2013: Improving reliability of coupled model forecasts of
Australian seasonal rainfall. Mon.Wea.Rev., 141, 728-741.
Lim, E.-P., H. H. Hendon and H. A. Rashid, 2013: Seasonal predictability of the Southern
Annular Mode due to its association with ENSO. J. Climate, http://dx.doi.org/10.1175/JCLI-D-
13-00006.1
Madec, G., and co-authors, 2008: NEMO ocean engine. Technical Report Note du Pole de
od´elisation No 27, ISSN No 1288-1619, Institut Pierre-Simon Laplace (IPSL), France.
MacLachlan, C., and co-authors, 2014: Global Seasonal forecast system version 5 (GloSea5): a
high-resolution seasonal forecast system. Quart. J. Roy. Meteor. Soc., DOI: 10.1002/qj.2396.
Manabe S, Holloway J, 1975: The seasonal variation of the hydrological cycle as simulated by a
global model of the atmosphere. J Geophys. Res. 80, 1617–1649
Marshall,A. G., D. Hudson, M. C. Wheeler, H. H. Hendon and O. Alves, 2010: Assessing the
simulation and prediction of rainfall associated with the MJO in the POAMA seasonal forecast
system . Climate Dyn., doi:10.1007/s00382-010-0948-2.
Megann, A., Storkey, D., Aksenov, Y., Alderson, S., Calvert, D., Graham, T., Hyder,
P., Siddorn, J. and Sinha, B.(2014) GO5.0: The joint NERC-Met Office NEMO global ocean
model for use in coupled and forced applications.Geoscientific Model
Development, 7, (3), 1069-1092. (doi:10.5194/gmd-7-1069-2014).
Pacanowski, R. C., 1996: MOM2. Documentation, user’s guide and reference manual. Tech.
rep. GFDL Ocean Group Tech. Rep. 3.2, 328 pp.
Reynolds, R. W., N. A. Rayner, T. M. Smith, D. C. Stokes, and W. Wang, 2002: An improved
in situ and satellite SST analysis for climate. J Climate 15, 1609–1625.
Scaife, A. A., and Coauthors, 2011: Improved Atlantic winter blocking in a climate
model. Geophys. Res. Lett., 38, L23703, doi:10.1029/2011GL049573.
COMPARISON OF GLOSEA5 AND POAMA2.4 HINDCASTS 1996-2009
50
Schiller, A., J. S. Godfrey, P. C. McIntosh, G. Meyers, N. R. Smith, O. Alves, G. Wang, and R.
Fiedler, 2002: A New Version of the Australian Community Ocean Model for Seasonal Climate
Prediction. CSIRO Marine Research Report No. 240.
Seviour, W.J. M., Steven C. Hardiman, Lesley J. Gray, Neal Butchart, Craig MacLachlan, and
Adam A. Scaife, 2014: Skillful Seasonal Prediction of the Southern Annular Mode and
Antarctic Ozone. J. Climate, 27, 7462–7474.
Shi, L., H. H. Hendon, O. Alves, J. -J. Luo, M. Balmaseda, D. Anderson, 2012: How
predictable is the Indian Ocean Dipole? Mon. Wea. Rev., 140, 3867-3884.
Uppala, S.M. and Coauthors, 2005. The ERA-40 re-analysis. Q J Roy Met Soc ,131, 2961-3012.
Valke, S., Terray, L. and Piacentini, A. 2000. The OASIS coupled user guide version 2.4,
Technical Report TR/ CMGC/00-10, CERFAC.
Walters, D.N., M. J. Best, A. C. Bushell, D. Copsey, J. M. Edwards, P. D. Falloon, C. M.
Harris, A. P. Lock, J. C. Manners, C. J. Morcrette, M. J. Roberts, R. A. Stratton, S. Webster, J.
M. Wilkinson, M. R. Willett, I. A. Boutle, P. D. Earnshaw, P. G. Hill, C. MacLachlan, G. M.
Martin, W. oufouma-Okia, M. D. Palmer, J. C. Petch, G. G. Rooney, A. A. Scaife, and K. D.
Williams, 2011: The Met Office Unified Model Global Atmosphere 3.0/3.1 and JULESglobal
land 3.0/3.1 configurations. Geoscientific Model Development, 4, 919–941.
Wheeler, M.C., and H.H. Hendon, 2004: An all season real-time multivariate MJO index:
Development of an Index for monitoring and prediction. Mon. Wea. Rev., 132, 1917-1932.
Yin, Y., O. Alves, and P. R. Oke, 2011: An ensemble ocean data assimilation system for
seasonal prediction. Mon. Wea. Rev., 139, 786-808.
Zhao, M., H. Hendon, O. Alves, Y. Yin, 2014: Impact of Improved Assimilation of
Temperature and Salinity for Coupled Model Seasonal Forecasts. Climate Dynamics, Vol. 42,
No.9, 2565-2583. DOI: 10.1007/s00382-014-2081-0.
Zhao, M., G. Roff, H.H. Hendon, P. Okely, X. Zhou, X., A. Marshall, G. Liu, F. Tseitkin, and
O. Alves, 2013: Improving Multiweek Rainfall Forecasts: Experimentation with the ACCESS
climate models CAWCR Technical Report No. 064.