INTERNATIONAL JOURNAL OF CLIMATOLOGY

Int. J. Climatol. 24: 1963–1972 (2004)

Published online in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/joc.1099

SKILL OF A SUPERENSEMBLE FORECAST OVER EQUATORIAL

SOUTHEAST ASIA
MASTURA MAHMUD*
Geography Programme, School of Social, Development and Environment, Universiti Kebangsaan Malaysia, Bangi, 43600 Selangor,
Malaysia
Received 15 May 2003
Revised 29 July 2004
Accepted 29 July 2004

ABSTRACT
The performance of the multianalysis–multimodel superensemble weather forecast and its comparison with several other
operational prediction models is investigated in this study for the Southeast Asian region. The superensemble technique
has been proven to exhibit exceptional skills compared with individual or even ensemble average forecasts for parameters
such as precipitation, motion and mass fields. Precipitation forecasts from the superensemble technique for the year
of 2001 from this study show that its skill scores are better than the other global operational models employed. Its
precipitation root-mean-square errors (RMSEs) are consistently lower than the other member models for the annual
period examined. The equitable threat score, which indicates the association of the common rainfall threshold distribution
between the forecast and observed best rainfall analysis from the Tropical Rainfall Measurement Mission and Special
Sensor Microwave Imager dataset, is slightly underpredicted for light rain thresholds, but it still outperforms the other
individual members or even the ensemble mean forecasts. Correlation analysis of the superensemble forecast against the
observed is far higher than either the ensemble mean or the individual forecast member models and is maintained for up
to 3-day forecasts.
The motion field comparisons also exhibit that the superensemble forecasts are closer to the observed data than the
other member models due to its low systematic errors at the 850 hPa pressure level. Both the precipitation and wind
field analyses have shown that the Florida State University multianalysis–multimodel superensemble forecast provides
the lowest RMSE, the highest correlation against the best-observed data and lowest systematic errors compared with
the other participating operational models. These forecasts have the potential to provide better daily weather predictions
over the Southeast Asian region, particularly during the early northeast monsoon that often causes heavy rainfall in the
equatorial part of the Southeast Asian region. Copyright  2004 Royal Meteorological Society.
KEY WORDS: superensemble; multimodel–multianalysis; precipitation forecast; Southeast Asia; motion field analysis; root-mean-square
error; precipitation bias

1. INTRODUCTION

Advances in numerical weather prediction have been possible through the utilization of fast computations that
simulate the physics of the atmosphere from complex mathematical equations. High-speed supercomputers
have made it feasible for numerical models to perform complex parameterizations of physical processes and
include increased horizontal resolutions that generate high temporal and spatial details of the evolution of
events at the synoptic and sub-synoptic scales (Brooks and Doswell, 1999).
Despite these supercomputers, inaccurate weather forecasts still exist due to errors from the chaotic nature
of the atmosphere, the inexact equations of motions, and gaps in specifying the initial state of the atmosphere
that can jeopardize the forecasts disseminated to the public. To date, many approaches have been undertaken
by researchers to improve the accuracy of forecasts that can predict reasonably well to more than 3 days in
advance.

* Correspondence to: Mastura Mahmud, Geography Programme, School of Social, Development and Environment, Universiti
Kebangsaan Malaysia, Bangi, 43600 Selangor, Malaysia; e-mail: [email protected]

Copyright  2004 Royal Meteorological Society

1964 M. MAHMUD

One such method that has been attempted to improve the accuracy of weather forecasts is ensemble
forecasting, and it is gaining popularity as a technique in numerical weather prediction due to its improved
forecast results. The ensemble approach is different from the trend of increasing horizontal resolutions that
amplifies the computational time required to produce a forecast.
The forecast technique investigated in this paper is called the superensemble, which is a unique type of
ensemble developed at Florida State University (FSU; Krishnamurti et al., 2002). The superensemble is a
collective bias estimation that produces the highest skill of weather and climate prediction compared with
member models of the ensemble mean and the bias-removed ensemble mean (Krishnamurti et al., 1999,
2001, 2002, 2003). The superensemble has produced better forecasting results for weather and climate that
are closer to observed values over the globe and in many specific areas, including the USA, Africa, Asia and
the Pacific, and for various events, such as forecasting typhoons, hurricanes and floods (Krishnamurti et al.,
1999, 2000a,b, 2002, 2003). The superensemble forecast skill is superior in terms of fewer errors, lower bias
and higher correlations compared with other individual operational models, ensemble mean or bias-removed
mean forecasts (Krishnamurti et al., 1999, 2000a,b, 2002, 2003; Vijaya Kumar et al., 2003).
This study describes an exercise performed to assess the comparison of the superensemble multi-
model–multianalysis real-time weather forecasts and analysed data over a selected region of Southeast Asia
for the year 2001. The multimodel superensemble in this study refers to the combination of six weather
models that constitute the member models of the Japan Meteorological Agency (JMA), National Centers
for Environmental Prediction (NCEP), Bureau Meteorology Research Centre (BMRC), FSU, Recherch é en
Prévision Numérique (RPN), and the Naval Research Laboratory (NRL). Past forecasts are employed to derive
statistics on the previous performance of these models from the operational European Centre for Medium-
Range Weather Forecasting (ECMWF) analysis that is utilized as the benchmark (Krishnamurti et al., 2003;
Vijaya Kumar et al., 2003). A superensemble forecast is then created from these statistics. Verification of
the wind field forecasts will be made on the superensemble against the six individual member models and
the ensemble mean of these member models. The data analysed are the ECMWF–FSU output data incor-
porated within the Tropical Rainfall Measurement Mission (TRMM) and Special Sensor Microwave Imager
(SSM/I) dataset in the FSU global spectral model for precipitation verification and wind field evaluation. The
precipitation verification analysis will be the superensemble forecast, which consists of the multimodels plus
an additional multianalysis component that consists of five other physical initialized components of the FSU
model of different initial rain rates from different satellite-based algorithms (Krishnamurti et al., 1999, 2001).
These five different rain-rate algorithms constitute the multianalysis components of the superensemble, specif-
ically the different initializations run on the FSU model. These analyses utilize: the Ferraro (1977; Ferraro
et al., 1998) algorithm and SSM/I rain rates; the Olson et al. (1990) algorithm and SSM/I rain rates; blended
geostationary and microwave rain-rate (GEO) algorithm (Turk et al., 2001); combined TRMM, SSM/I and
geostationary satellite-based algorithm (Turk et al., 2001); and combined TRMM (2A12) and SSM/I (GPROF)
algorithm (Kummerow et al., 2000).

2. OBJECTIVE

The performance of the prediction skill of the superensemble technique on forecasting of the rainfall and
wind field over the equatorial region of Southeast Asia is investigated. More accurate weather or climate
forecasts than the present operational precipitation prediction are sought to assist local forecasters to prepare
the nation better for any imminent heavy rainfall events that may cause flooding. Such examples are cold-surge
events that may intensity any existing vortex within the vicinity of the South China Sea, or the expectation
of heavy rainfall from a developing tropical cyclone or that from easterly waves, particularly during the
northeast monsoon, when heavy rainfall can cause flooding to some countries, such as Malaysia, Indonesia
and southern Thailand.
A comparison between the superensemble and the individual operational models, as well as with the
ensemble mean, will be performed in this investigation. The ensemble mean forecast consists of six different
operational member models, as stated previously. This type of ensemble is different from the ensemble
Copyright  2004 Royal Meteorological Society Int. J. Climatol. 24: 1963–1972 (2004)
 SUPERENSEMBLE FORECASTING OVER SOUTHEAST ASIA 1965

forecast of the single model that utilizes a set of slightly different initializations (Gel et al., 2002), where
these different initializations constitute the member models from which an ensemble of evolutions of the
atmosphere is generated (Brooks and Doswell, 1999).
The verification of the forecast skill of the superensemble is analysed thorough standard statistics, such as
the root-mean-square error (RMSE) and correlation. The precipitation analysis will also include comparisons
for the forecast of individual multianalysis components and the superensemble multimodel–multianalysis,
where additional skill scores assessed will include the equitable threat score and bias analysis. The formulae
for the statistical parameters utilized in the verification analysis are shown in Appendix A, following Ross
and Krishnamurti (submitted).

3. THE SUPERENSEMBLE TECHNIQUE

The superensemble multiple model, also called the multimodel superensemble, requires a minimum number of
seven to eight member models to produce a superior performance of forecast (Krishnamurti et al., 2002). The
superensemble forecast system utilizes an evaluation period of 120 days to capture the characteristics between
the observed and the forecast fields (Krishnamurti et al., 2001; Ross and Krishnamurti, submitted). Better
forecasts are generated when higher quality evaluation data sets are utilized compared with an evaluation of
poorer datasets.
The uniqueness of the superensemble lies in the evaluation or training phase, which is divided into two
parts. The first portion of the evaluation phase consists of a comparison between the forecasts of the member
models and the observed field. At this stage, a least-squares regression of the individual model forecasts to
observed values is obtained.
The superensemble forecast is represented by

N
S=O+ ai (Fi − F i )
i=1

where S is the superensemble forecast, O is the time mean of the observed state during the evaluation period,
ai is the regression coefficient by a minimization process for model i, i is the model index, N is the number
of models, F i is the mean of prediction by model i, and Fi is the forecast by model i (Ross and Krishnamurti,
submitted).
The regression coefficient is determined using the Gauss–Jordan elimination technique (Krishnamurti et al.,
2002). Minimization of the mean-square error of the forecast is achieved and regression coefficients a i are
derived from estimating the minimum of the function

teval
G= (St − Ot )2
t=1

where O is the observed state, t is the time and teval is the length of the evaluation period.
The second portion of the evaluation phase consists of the member-model predictions. The superensemble
combines each of the forecasts according to the weights obtained from the evaluation period. Weightings are
generated from the least-squares minimization of the difference between the analysis and model (Krishnamurti
et al., 2002). Different weightings are given to the member models that produce the poorer results compared
with the member models that produce better forecasts. Higher weightings are given to better models, whilst
lower weightings are designated for the poorer member models. This makes this ensemble superior to other
ensembles, such as the normal ensemble mean or the bias-removed ensemble mean, or even individual
member models. The selective weightings from each model contribute to the success of the superensemble
in reducing errors and bias against the observed values. This effective weighting is the crux of the success
of the superensemble in improving the weather and climate forecasts (Krishnamurti et al., 1999).
Copyright  2004 Royal Meteorological Society Int. J. Climatol. 24: 1963–1972 (2004)
1966 M. MAHMUD

4. AREA OF RESEARCH

The area chosen for the study covers the Southeast Asian region, whose domain is within the longitudes 90-
120 ° E and from 5 ° S to 20 ° N. The countries of this region include Malaysia, Indonesia, Singapore, Indonesia,
Brunei and Indochina, which comprises countries such as Thailand, Myanmar, Cambodia, Laos and Vietnam.
The area falls within the maritime continent, where the latent heating is considered one of the important
energy sources of the global circulation.
There are two main climate types over the Southeast Asian region. Malaysia and Indonesia experience
an equatorial climate that is characterized by warm and humid weather throughout the year. The region is
influenced by the prevailing southwesterly winds from May to September during the southwest monsoon and
by northeasterly winds from November to March. Transitional periods occur in April and October, when the
winds do not have distinct prevailing directions. The climate over the mainland Indochina region is tropical,
where the warm, rainy season occurs during the southwest monsoon from mid-May to September. Its dry
season occurs during the northeast monsoon, from mid-October to mid-March.

5. RESULTS

5.1. Annual precipitation analysis

The root-mean-square error (RMSE) for the annual precipitation in the year 2001 is analysed for the global
area (Ross and Krishnamurti, submitted) and Southeast Asia. The smoothed RMSE annual precipitation
between the forecast models and observed rainfall acquired by SSM/I–TRMM was found to be the least
for the superensemble and the ensemble models, compared with the other models, for the day-1 to day-4
forecast over the Southeast Asia region (Figure 1). The values of the forecast predicted by the other 10
member models exhibit errors that are higher than that over the larger global area that is confined within the
domain of 55 ° N to 55 ° S. The RMSE of less than 2 mm/day is maintained by the superensemble forecast until
the day-4 forecast, whereas the rest of the ensemble and member models reveal larger errors, at more than
8 mm/day, for all the days investigated. The notations of FSU C, FSU O, FSU J, FSU T and FSU G represent
the different rain-rate algorithms that utilize the Ferraro (1977, Ferraro et al., 1998) algorithm and SSM/I
rain rates; the Olson et al. (1990) algorithm and SSM/I rain rates; the blended geostationary and microwave
rain-rate (GEO) algorithm (Turk et al., 2001); the combined TRMM, SSM/I and geostationary satellite-based
algorithm (Turk et al., 2001); and the combined TRMM (2A12) and SSM/I (GPROF) algorithm (Kummerow
et al., 2000) respectively.
Generally, the first day’s forecast exhibited the lowest RMSE compared with the forecasts for subsequent
days. The JMA model is considered the best operational prediction model over Southeast Asia. Comparatively,
the RMSE magnitudes of the multianalysis models (FSU C, FSU O, FSU J, FSU T and FSU G) that utilized
the different algorithms and rain rates were smaller than the six individual operational models. The relatively
moderate value of RMSE for the mean ensemble forecast reflects the performance skill of the combined
poorer and better member models. The ensemble mean forecast is unable to retain the value of the low
RMSE obtained for the first day’s forecast. The subsequent forecasts for days 2 and 3 exhibit errors that are
even larger than the best model.
These results demonstrate that the magnitude of the regional superensemble RMSE over Southeast Asia is
smaller than the superensemble over the global area from 55 ° N to 55 ° S (Ross and Krishnamurti, submitted).
Krishnamurti et al. (2001) also found that the superensemble skills are consistently high in every part of the
regional domain that they studied for each day of forecast. The superensemble skill was noted to be higher
over the Asian domain of 15 ° S to 45 ° N and 50 to 120 ° E, in contrast to their best model.
The precipitation correlation for the superensemble forecast over Southeast Asia illustrates a high correlation
of 0.7 on the day-1 forecast compared with the other models (Table I). The superensemble correlation between
the observed and forecast precipitation for Southeast Asia was slightly higher than the global region, which
ranged from 0.7 to 0.6 for day-1 to day-3 forecasts. This finding is also similar to a study by Krishnamurti et al.
(2000a), where the forecast skill of the superensemble technique for up to 3-day forecasts for regional areas
Copyright  2004 Royal Meteorological Society Int. J. Climatol. 24: 1963–1972 (2004)
 SUPERENSEMBLE FORECASTING OVER SOUTHEAST ASIA 1967

(a) Annual 2001 Precipitation RMSE over Southeast Asia

14
RPN
12 BMR
FSU
10
RMSE (mm/day)

FSU C

8 FSU G
FSU J
6
FSU O

4 FSU T
NCEP
2 JMA
NRL
0
Day 1 Day 2 Day 3 Day 4 Ensemble
Forecast Day Superensemble

(b) Annual 2001 Precipitation RMSE over Globe

12
RPN
BMR
10
FSU

8 FSU C
RMSE (mm/day)

FSU G
6 FSU J
FSU O
4 FSU T
NCEP
2
JMA
NRL
0
Day 1 Day 2 Day 3 Day 4 Ensemble

Forecast Day Superensemble

Figure 1. The smoothed RMSE for the annual precipitation in 2001 over (a) Southeast Asia and (b) globe

such as Africa and the tropical Americas was found to be more effective than over the global tropics, where
correlation coefficients as high as 0.7 were found, signifying the larger climatological rainfall components
over this region.
The precipitation equitable threat score for the day-1 forecasts was found to exhibit the highest value of the
superensemble technique at 0.86 for the precipitation threshold of 0.2 mm/day (not shown). Equitable threat
score accounts for the probability of the areas of overlap between the forecast and observed precipitation.
Ideally, a value of unity is preferred, where this represents an exact placement of the forecast rainfall area. A
value of less than one indicates underforecasting, whereas a value of more than one indicates overforecasting
of the rainfall areas compared with observation. The observed best rainfall analysis utilized for this study is
obtained from the TRMM and SSM/I dataset. For the sake of comparison with the global area, the 2.0 mm/day
precipitation threshold was selected. It was found that the superensemble forecast for this light precipitation
presented the highest equitable threat score, at a value of 0.38, over the globe. Nevertheless, some member
models are found to forecast higher equitable threat score values compared with others over the Southeast
Asian region. The superensemble forecast over the Southeast Asian region is higher in value at 0.49, but this
Copyright  2004 Royal Meteorological Society Int. J. Climatol. 24: 1963–1972 (2004)
1968 M. MAHMUD

Table I. The results of the precipitation correlation, equitable threat score and bias over the globe and the Southeast
Asian region

Forecast Member Model

RPN BMRC FSU FSU C FSU G FSU J FSU O FSU T NCEP JMA NRL Ensemble Super-
mean ensemble

Correlation coefficient
Global (55 ° N to 55 ° S)
Day 1 0.45 0.43 0.37 0.40 0.40 0.37 0.37 0.45 0.53 0.49 0.43 0.54 0.61
Day 2 0.38 0.40 0.27 0.31 0.28 0.26 0.28 0.31 0.47 0.45 0.38 0.44 0.53
Day 3 0.33 0.35 0.24 0.26 0.24 0.23 0.24 0.27 0.42 0.41 0.34 0.40 0.50
Southeast Asia
Day 1 0.36 0.38 0.24 0.29 0.32 0.25 0.28 0.38 0.50 0.48 0.43 0.52 0.70
Day 2 0.30 0.35 0.16 0.20 0.18 0.16 0.18 0.18 0.45 0.41 0.40 0.38 0.64
Day 3 0.26 0.31 0.16 0.18 0.16 0.14 0.17 0.16 0.39 0.36 0.35 0.34 0.61
Equitable threat score
Global (55 ° N to 55 ° S)
Day 1 0.33 0.24 0.22 0.20 0.24 0.24 0.21 0.26 0.34 0.30 0.21 0.34 0.38
Day 2 0.28 0.22 0.15 0.15 0.17 0.16 0.15 0.16 0.30 0.28 0.18 0.25 0.33
Day 3 0.20 0.17 0.13 0.14 0.11 0.11 0.12 0.11 0.21 0.21 0.16 0.25 0.33
Southeast Asia
Day 1 0.38 0.45 0.09 0.21 0.15 0.12 0.11 0.26 0.59 0.46 0.29 0.38 0.49
Day 2 0.33 0.32 0.05 0.07 0.07 0.08 0.07 0.08 0.55 0.35 0.34 0.31 0.41
Day 3 0.04 0.24 0.05 0.18 0.08 0.11 0.04 0.08 0.30 0.12 0.28 0.12 0.38
Bias
Global (55 ° N to 55 ° S)
Day 1 1.13 1.54 1.62 2.37 1.66 1.64 1.72 1.67 1.42 1.41 0.53 1.64 1.08
Day 2 1.17 1.60 2.21 2.50 2.23 2.22 2.34 2.25 1.53 1.41 0.58 2.07 1.2
Day 3 1.85 2.35 3.2 2.22 3.05 3.05 3.3 3.09 2.24 2.18 0.82 2.01 1.22
Southeast Asia
Day 1 1.28 1.30 1.32 1.73 1.41 1.40 1.41 1.47 1.19 1.58 0.36 1.52 2.00
Day 2 1.51 1.65 1.44 1.63 1.24 1.17 1.53 1.42 1.33 1.82 0.65 1.75 2.32
Day 3 2.11 2.70 3.12 2.09 3.07 2.97 3.02 3.19 2.23 2.29 1.10 1.70 2.43

is still slightly lower than the JMA member model, considered the best model for the 2.0 mm/day threshold
chosen. However, the superensemble is still better than the ensemble mean forecast for the region.
The equitable threat scores for the individual multianalysis models are lower than the other member models
over Southeast Asia, except for the FSU global spectral model. This may be due to the poor forecast presented
by the FSU model itself. The various rain-rate algorithms, represented by the notations FSU C, FSU G, FSU
J, FSU O and FSU T, have slightly improved the forecasts of these areas compared with the initial FSU
forecast that did not utilize the rain-rate initializations. In the contrary, the performances of the FSU model
and the multianalysis models are better over the globe, where the magnitudes are comparable to the other
operational member models. This indicates that the FSU model performed fairly poorly over the small domain
of the equatorial Southeast Asian region.
The bias score represents a contrast between the relative amounts of the number of model grids of the
forecast precipitation and the observed precipitation. It determines the exactness of the areal coverage of a
specific threshold quantity, regardless of the accuracy in location. A negative value indicates an underforecast
of precipitation, whereas a positive value indicates an overforecast. Precipitation bias scores for the first
day’s forecast show that the superensemble slightly overforecasts the area coverage of precipitation by
0.05, compared with the slight underforecast presented by the ensemble model and all the other member
models for the precipitation threshold of 0.2 mm/day. There is a tendency for overforecasting the areal
Copyright  2004 Royal Meteorological Society Int. J. Climatol. 24: 1963–1972 (2004)
 SUPERENSEMBLE FORECASTING OVER SOUTHEAST ASIA 1969

precipitation of thresholds 2.0, 5.0, 10.0 and 50.0 mm/day. Some member models exhibit the tendency to
underforecast the areal precipitation of the different thresholds of rainfall, whilst others have a tendency to
overforecast the higher intensities of rainfall of more than 25 mm/day. The superensemble still retains a high
skill in illustrating the correct forecast of low threshold of 0.2 mm/day, whereas the higher precipitation
thresholds of 2.0 mm/day or more are overforecast by the third day. Other member models are consistently
underforecasting the light rainfall threshold. This illustrates that the multianalysis–multimodel superensemble
is able to present superior bias skill for the first day’s forecast for very light precipitation of 0.2 mm/day, but
it has the tendency to overforecast the higher threshold of precipitation over the Southeast Asian region. This
is due to the contribution of overprediction by some of the member models that make up the superensemble.
The performance of the bias skill for the superensemble is better over the global belt. The value of the bias
is the least compared with the other member models or the ensemble mean.

5.2. The wind field analysis

As approximately half of the Southeast Asian region is covered with water, assessment of the models’
abilities in producing strong winds over the South China Sea where there is a paucity of meteorological
stations is invaluable. The RMSE for the multimodel superensemble forecast of the 850 hPa vector winds for
the Southeast Asian region is comparable to the superensemble forecast of the global area, at a magnitude
of less than 2 m/s for the first day’s forecast, which is nevertheless smaller than the rest of the individual
member models (Figure 2). A few of the member models exhibit RMSEs that are comparatively larger than
the rest of the models, and these errors intensify with subsequent forecast days. The performance of the
superensemble forecast for total winds by the third day is still small, at 2.5 m/s, yet lower than the rest of
the member models for the first 2 days of their forecasts.
The JMA model presents the lowest RMSE over the Southeast Asian region and is considered the best
individual operational model. The ensemble mean still exhibits an RMSE larger than even the JMA. This
shows that the contribution from the poorer member models does have an effect in degrading the forecast
skill. The value of the superensemble RMSE for the global region is also the lowest compared with the other
multimodels and the ensemble mean. This is in contrast to the regional Southeast Asia results, where the
ensemble mean presents the next lowest RMSE after the superensemble.
Higher values of the RMSE are found for the 200 hPa vector winds in contrast to the magnitude of the
errors at 850 hPa (Figure 3). Comparatively, the magnitudes of the annual 200 hPa wind vector RMSE over
the Southeast Asian region are lower than over the globe. At this level, the values of the RMSE for the
superensemble are the lowest compared with the member models over the global belt and the Southeast
Asian region. The ensemble forecast presents the next lowest RMSE compared with the other individual
member models.

6. CONCLUDING REMARKS

The forecast skill for precipitation over the Southeast Asian region appears to be promising and currently is
one of the better predictors of rainfall compared with many operational forecasts, particularly for the skill for
the 1 to 3 days’ forecast. This is also verified by Krishnamurti et al. (2000), who showed similar findings for
selected regions, such as Africa and tropical America.
The RMSE for the precipitation of the superensemble forecast is approximately 80% lower than that of the
ensemble or individual operational models over this region. We have verified this with the results from the
global superensemble forecast and found that, over the selected equatorial Southeast Asia, the RMSEs are as
small as 2 mm/day.
We found that the annual, seasonal and monthly statistics (not shown), i.e. the RMSE, bias and equitable
threat score, were mostly better than the statistics for the daily event. The best available rainfall estimates
of the combined TRMM–SSM/I algorithm are used as the benchmark training set that greatly improved the
forecast of the superensemble, which appears to be the better estimate of the forecast rainfall in terms of the
correct placement but still requires improvement in forecasting the higher intensities of rainfall.
Copyright  2004 Royal Meteorological Society Int. J. Climatol. 24: 1963–1972 (2004)
1970 M. MAHMUD

(a) Southeast Asia 850 hPa Annual Wind Vector

6

4 RPN
RMSE (m/s)

BMRC
FSU
3
NCEP
JMA
2
NRL
Ensemble Mean
1
Superensemble

0
Day 1 Day 2 Day 3 Day 4 Day 5
Forecast Day

(b) Global 850 hPa Annual Wind Vector

6

RPN
4
RMSE (m/s)

BMRC
FSU
3
NCEP
JMA
2 NRL
Ensemble Mean
1 Superensemble

0
Day 1 Day 2 Day 3 Day 4 Day 5
Forecast Day

Figure 2. The RMSE for the annual vector winds at 850 hPa over (a) Southeast Asia and (b) the globe

We have illustrated that the superensemble technique has the potential to provide better precipitation
forecasts over equatorial Southeast Asia, as shown by the RMSE, correlation, equitable threat scores and
bias. There is a link to the Website at http://lexxy.met.fsu.edu/rtnwp/forecasts/forecast.html for the updated
superensemble forecasts for the Southeast Asian region.
The wind field analysis also shows that the superensemble forecast presents better skill in predicting winds
at the lower and higher levels of the atmosphere, as indicated by the lowest RMSE values presented compared
with other operational member models and the ensemble mean.

ACKNOWLEDGEMENTS

I am indebted to Professor T. N. Krishnamurti, the originator of the superensemble, for his encouragement and
the opportunity to utilize the FSU superensemble data. Dr Robert Ross is also thanked for his comments and
provision of the global data. Appreciation is extended to Brian Mackey, Nihat Cubukcu, V. Sanjay and Darlene
Oosterhof for their assistance on the computing technicalities, graphics and administration respectively.
The modelling centres RPN, NRL, BMRC, NCEP, JMA, FSU, TRMM and SSM/I are acknowledged for
making their datasets available, without which this work would not be possible. My support by the Fullbright
Copyright  2004 Royal Meteorological Society Int. J. Climatol. 24: 1963–1972 (2004)
 SUPERENSEMBLE FORECASTING OVER SOUTHEAST ASIA 1971

(a) Southeast Asian 200 hPa Annual Wind Vector

8

6
RPN
5
RMSE (m/s)

BMRC
FSU
4
NCEP
3 JMA
NRL
2
Ensemble Mean
1 Superensemble

0
Day 1 Day 2 Day 3 Day 4 Day 5
Forecast Day

(b) Global 200 hPa Annual Wind Vector

10
9
8
7 RPN
RMSE (m/s)

6 BMRC
FSU
5
NCEP
4 JMA
3 NRL
2 Ensemble Mean
Superensemble
1
0
Day 1 Day 2 Day 3 Day 4 Day 5
Forecast Day

Figure 3. The RMSE for the annual vector winds at 200 hPa over (a) Southeast Asia and (b) the globe

Programme is greatly appreciated. The research report was supported by the FSU Research Foundation Center
of Excellence Award Grant 1368-760-45, FSU Research Foundation Grant 1338-831-45, NSF ATM-0108741,
NASA NAG5-9662, NOAA NA06GPO512 and NOAA NA16GP1365.

APPENDIX A


1
N
RMSE =  (fn − on )2
N n=1


N
(fn − f )(on − o)
n=1
Correlation coefficient = 

1
N
N
 (fn − f )2 (on − o)2
N n=1 n=1

Copyright  2004 Royal Meteorological Society Int. J. Climatol. 24: 1963–1972 (2004)
1972 M. MAHMUD

 
O
H− F
N
Equitable threat score =  
O
F +O −H − F
N
Nf
Bias =
No

where N is the number of grid points, fn is the forecast value at grid point n, on is the observed value at grid
point n, f is the mean area of the forecast values, o is the mean area of the observed values, F is the forecast
area of a specified threshold amount, O is the observed area of a specified threshold amount, H is the area
of overlap between O and F , Nf is the number of grid points for F and No is the number of grids for O.

REFERENCES
Brooks HE, Doswell CA. 1999. New technology and numerical weather prediction. Weather 48: 173–177.
Ferraro RR. 1997. Special Sensor Microwave Imager derived global rainfall estimates for climatological applications. Journal
Geophysical Research 102: 16 715–16 735.
Ferraro RR, Grody N, Weng F, Moore D. 1998. Microwave surface and precipitation products from the AMSU. In 14th International
Conference on Interactive Information and Processing System (IIPS) for Meteorology, Oceanography, and Hydrology; 132–133.
Gel Y, Raftery A, Gneiting T. 2002. Uncertainty in numerical weather forecasting via ensembles of initializations. In Twelfth PSU/NCAR
Mesoscale Model Users’ Workshop, NCAR, 24–25 June.
Krishnamurti TN, Xue J, Bedi HS, Ingles K, Oosterhof D. 1991. Physical initialization for numerical weather prediction over the tropics.
Tellus A 43(4): 53–81.
Krishnamurti TN, Kishtawal CM, Larow TE, Bachiochi DR, Zhang Z, Willowford CE, Gadgil S, Surendran S. 1999. Improved weather
and seasonal climate forecasts from multimodel superensemble. Science 285: 1548–1550.
Krishnamurti TN, Kishtawal CM, Shin DW, Williford CE. 2000a. Improving tropical precipitation forecasts from a multianalysis
superensemble. Journal of Climate 13: 4217–4227.
Krishnamurti TN, Kishtawal CM, Zhang Z, Larow TE, Bachiochi DR, Willowford CE. 2000b. Multimodel ensemble forecasts for
weather and seasonal climate. Journal of Climate 13: 4196–4216.
Krishnamurti TN, Surendran S, Shin DW, Correa-Torres RJ, Kumar TSVV, Willowford CE, Kummerow C, Adler RF, Simpson J,
Kakar R, Olson WS, Turk FJ. 2001. Real-time multianalysis–multimodel superensemble forecasts of precipitation using TRMM
and SSM/I products. Monthly Weather Review 129: 2861–2883.
Krishnamurti TN, Kumar TSVV, Yun WT, Chakraborty A, Stefanova L, Cubukcu N, Williford E, Mackey B, Sanjay J. 2002.
Superensemble forecasts for weather and climate. In ECMWF Seminar of Predictability of Weather and Climate, 9–13 September.
http://lexxy.met.fsu.edu/rtnwp/pubs/ecmwf krish superensemble.pdf [Accessed July 2004].
Krishnamurti TN, Surendran S, Rajendran K, Vijaya Kumar TSV, Lord S, Toth Z, Zou X, Cocke S, Ahlquist JE, Navon IM. 2003.
Improved skill for the anomaly correlation of geopotential heights at 500 hPa. Monthly Weather Review 131: 1082–1102.
Kummerow C, Simpson J, Thiele O, Barnes W, Chang ATC, Stocker E, Adler RF, Hou A, Kakar R, Wentz F, Ashcroft P, Kozu T,
Hong Y, Okamoto K, Iguchi T, Kuroiwa H, Im E, Haddad Z, Huffman G, Ferrier B, Olson WS, Zipser E, Smith EA, Wilheit TT,
North G, Krishnamurti T, Nakamura K. 2000. The status of the Tropical Rainfall Measuring Mission (TRMM) after two years in
orbit. Journal of Applied Meteorology 39: 1965–1982.
Olson WS, LaFontaine FJ, Smith WL, Achtor TH. 1990. Recommended algorithms for the retrieval of rainfall rates in the tropics using
the SSM/I (DMSP-8). University of Wisconsin-Madison.
Ross RS, Krishnamurti TN. Submitted. Reduction of forecast error for global numerical weather prediction by the Florida State
University (FSU) superensemble. Journal of Meteorology and Atmospheric Physics.
Turk J, Liou C-S, Qiu S, Scofield R, Ba M, Gruber A. 2001. Capabilities and characteristics of rainfall estimates from geostationary
and geostationary + microwave-based satellite techniques. In Preprints, Symposium on Precipitation Extremes: Prediction, Impacts,
and Responses. American Meteorological Society: Albuquerque, NM; 191–194.
Vijaya Kumar TSV, Krishnamurti TN, Fiorino M, Nagata M. 2003. Multimodel superensemble forecasting of tropical cyclones in the
Pacific. Monthly Weather Review 131: 574–583.

Copyright  2004 Royal Meteorological Society Int. J. Climatol. 24: 1963–1972 (2004)