This article was downloaded by: [Nkrintra Singhrattna]
On: 23 January 2012, At: 16:41
Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK
Hydrological Sciences Journal
Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/thsj20
Hydroclimate variability and long-lead forecasting
of rainfall over Thailand by large-scale atmospheric
variables
a
a
Nkrintra Singhrattna , Mukand S. Babel & Sylvain R. Perret
a
a b
Water Engineering and Management, Asian Institute of Technology, Pathumthani, Thailand
b
Centre de Coopération Internationale en Recherche Agronomique pour le Développement,
UMR G-Eau, Montpellier, France
Available online: 20 Jan 2012
To cite this article: Nkrintra Singhrattna, Mukand S. Babel & Sylvain R. Perret (2012): Hydroclimate variability and long-lead
forecasting of rainfall over Thailand by large-scale atmospheric variables, Hydrological Sciences Journal, 57:1, 26-41
To link to this article: http://dx.doi.org/10.1080/02626667.2011.633916
PLEASE SCROLL DOWN FOR ARTICLE
Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-conditions
This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to
anyone is expressly forbidden.
The publisher does not give any warranty express or implied or make any representation that the contents
will be complete or accurate or up to date. The accuracy of any instructions, formulae, and drug doses should
be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims,
proceedings, demand, or costs or damages whatsoever or howsoever caused arising directly or indirectly in
connection with or arising out of the use of this material.
26
Hydrological Sciences Journal – Journal des Sciences Hydrologiques, 57(1) 2012
Hydroclimate variability and long-lead forecasting of rainfall over
Thailand by large-scale atmospheric variables
Nkrintra Singhrattna1 , Mukand S. Babel1 and Sylvain R. Perret1,2
1
Water Engineering and Management, Asian Institute of Technology, Pathumthani, Thailand
[email protected], [email protected]
2
Centre de Coopération Internationale en Recherche Agronomique pour le Développement, UMR G-Eau, Montpellier, France
Downloaded by [Nkrintra Singhrattna] at 16:41 23 January 2012
Received 27 May 2010; accepted 25 March 2011; open for discussion until 1 July 2012
Editor Z.W. Kundzewicz
Citation Singhrattna, N., Babel, M.S. and Perret, S.R., 2012. Hydroclimate variability and long-lead forecasting of rainfall over
Thailand by large-scale atmospheric variables. Hydrological Sciences Journal, 57 (1), 26–41.
Abstract The development of statistical relationships between local hydroclimates and large-scale atmospheric
variables enhances the understanding of hydroclimate variability. The rainfall in the study basin (the Upper Chao
Phraya River Basin, Thailand) is influenced by the Indian Ocean and tropical Pacific Ocean atmospheric circulation. Using correlation analysis and cross-validated multiple regression, the large-scale atmospheric variables,
such as temperature, pressure and wind, over given regions are identified. The forecasting models using atmospheric predictors show the capability of long-lead forecasting. The modified k-nearest neighbour (k-nn) model,
which is developed using the identified predictors to forecast rainfall, and evaluated by likelihood function, shows
a long-lead forecast of monsoon rainfall at 7–9 months. The decreasing performance in forecasting dry-season
rainfall is found for both short and long lead times. The developed model also presents better performance in
forecasting pre-monsoon season rainfall in dry years compared to wet years, and vice versa for monsoon season
rainfall.
Key words rainfall; hydroclimate variability; ENSO; large-scale atmospheric variables; long-lead forecasting; statistical
approach; modified k-nn model; cross-validated multiple regression; Chao Phraya River Basin; Ping River Basin; Thailand
Variabilité hydroclimatique et prévision à long terme des précipitations en Thaïlande à l’aide de
variables atmosphériques de grande échelle
Résumé L’établissement de relations statistiques entre variables pluviométriques locales et variables atmosphériques de grande échelle permet une meilleure compréhension de la variabilité hydroclimatique. Les
précipitations dans le bassin d’étude (le bassin supérieur de la Rivière Chao Phraya, Thaïlande) sont influencées
par la circulation atmosphérique dans l’Océan Indien et dans l’Océan Pacifique tropical. Par l’analyse de corrélations et régressions multiples en validation croisée, les prédicteurs de variables atmosphériques de grande échelle,
à savoir température, pression et vent, sont identifiés sur des régions cibles, et montrent une bonne capacité de
prévision à long terme. Le modèle modifié des k-plus proches voisins (k-nn), développé par les prédicteurs identifiés pour prévoir les pluies, et évalué par fonction de probabilité, montre une prévisibilité à long terme (7–9 mois)
des pluies de mousson. La moindre performance dans la prévision des pluies de saison sèche est prouvée pour la
prévision à court et à long terme. Le modèle développé présente également une meilleure performance pour les
pluies de pré-mousson en année sèche, par rapport à une année humide, et inversement pour les pluies de mousson.
Mots clefs précipitations; variabilité hydroclimatique; ENSO; variables atmosphériques de grande échelle; prévision à long
terme; approche statistique; modèle k-nn modifié; régression multiple à validation croisée; bassin du Fleuve Chao Phraya;
bassin du Fleuve Ping; T haïlande
1 INTRODUCTION
The links between large-scale atmospheric variables and local hydroclimates are widely studied
and reported to diagnose the variability of local
ISSN 0262-6667 print/ISSN 2150-3435 online
© 2011 IAHS Press
http://dx.doi.org/10.1080/02626667.2011.633916
http://www.tandfonline.com
hydroclimates (Smith and O’Brien 2001). The
anomalies of the climate variables known as atmospheric indices, e.g. the anomalous sea-surface temperature (SST) index or El Niño-Southern Oscillation
Downloaded by [Nkrintra Singhrattna] at 16:41 23 January 2012
Hydroclimate variability and long-lead forecasting of rainfall over Thailand by large-scale atmospheric variables
(ENSO), can strengthen or weaken a monsoon
(Mason and Goddard 2001, Smith and O’Brien 2001).
Thus, the atmospheric anomalies are responsible for
the variability of local hydroclimates, with respect
to temperature (Gershunov 1998, Pavia et al. 2006),
precipitation (Masiokas et al. 2006) and streamflow
(Gutiérrez and Dracup 2001, Meko and Woodhouse
2005). Due to an anomalous phase of the ocean and
atmosphere, the responses of local hydroclimates may
not be experienced with the same level of variability. In some regions, storms develop more strongly
and frequently (Cañón et al. 2007, Kim et al. 2006),
whereas in other regions, dry conditions are experienced with droughts (Mendoza et al. 2005) during longer periods (Schöngart and Junk 2007). The
effects of anomalous oceanic–atmospheric circulation have changed climate patterns not only along
the coast, but also in regions located far from it
(Saravanan and Chang 2000). Moreover, their effects
are difficult to identify because the local climate may
be influenced by nearby and distant oceans through
the coupled atmospheric–oceanic circulation (Wu and
Kirtman 2004, Nagura and Konda 2007). The teleconnection influences could be determined by the
lag time of anomalous events up to several months
after a phenomenon occurs (Sourza Filho and Lall
2003, Grantz et al. 2005), which makes it difficult to
understand.
The annual variability of the Southeast Asian climate is influenced by the interactions between the
annual reversal of surface monsoonal winds from the
Indian Ocean to the equatorial western Pacific Ocean,
including the South China Sea, and the complex distribution of land, sea and terrain. For the seasonal
development of surface winds, the Asian summer
monsoon that develops due to the land–ocean temperature gradient is influenced by the sea-surface temperature (SST) anomalies over the equatorial eastern
Indian Ocean (Goswami et al. 2006) and the tropical Pacific Ocean (Fasullo and Webster 2002). The
non-convective anomalies over the equatorial Indian
Ocean tend to weaken the convection over the Bay of
Bengal, the eastern Indian Ocean, Southeast Asia and
the equatorial western Pacific Ocean (Krishnan et al.
2000). Thus, the anomalous oceanic–atmospheric circulation, e.g. ENSO, affects the Asian summer monsoon and subsequently the monsoon rainfall. A warm
phase of ENSO (El Niño) is associated with a weak
monsoon and below-normal rainfall, whereas it is
the opposite for a cold phase of ENSO, i.e. La
Niña (Shrestha 2000, Shrestha and Kostaschuk 2005).
However, the influences over a particular region
27
have to be carefully investigated due to the different responses which make it hard to define a certain
pattern of local hydroclimates (An et al. 2007).
The lagging relationships of local hydroclimates with anomalous atmospheric events are also
of interest when developing hydroclimatic forecasting models. Several models, such as artificial neural networks (ANNs) (Silverman and Dracup 2000),
multiple regression model (Schöngart and Junk 2007)
and principal component analysis (Eldaw et al. 2003),
have been developed, and show better performance
using the predictors of large-scale atmospheric information (Singhrattna et al. 2005a, Zehe et al. 2006).
The basic variables, e.g. SST, surface air temperature, atmospheric pressure, wind and ENSO standard indices, are applied in the forecasting models
(Gershunov 1998, Clark et al. 2001, Hamlet et al.
2002, McCabe and Dettinger 2002, Grantz et al.
2005) based on the significant relationships between
large-scale atmospheric variables and hydroclimates.
The forecasting models enable successful short-lead
projections to be made (Gutiérrez and Dracup 2001,
Schöngart and Junk 2007); however, the model performance of long-lead forecasts still needs to be
improved (Krishna Kumar et al. 1995, DelSole and
Shukla 2002). The decadal variability of large-scale
atmospheric variables used as the predictors of a
model could weaken the relationships with hydroclimates and subsequently decrease the model performance of long-lead forecasts.
The objectives of this research are to study the
climate characteristics and their variability in part
of the Upper Chao Phraya River Basin located in
Thailand (Southeast Asia), and to identify the predictors of large-scale atmospheric variables at long-lead
times for rainfall forecasting. Using the identified
variables as predictors, the rainfall forecasting model,
based on a statistical-stochastic approach, is ultimately developed to forecast the seasonal rainfall at
long lead times.
2 THE STUDY BASIN
Thailand, located between the Indian Ocean and the
Gulf of Thailand which is connected to the Pacific
Ocean (Fig. 1(a)), covers an area of 513 115 km2
with a population of 62.4 million (2005). Its climate is
influenced by the Indian Ocean and the Pacific Ocean
due to the land–ocean interrelated temperature gradients. The summer season, from mid-February to midMay, is responsible for developing the land–ocean
temperature gradient that strengthens the southwest
28
Downloaded by [Nkrintra Singhrattna] at 16:41 23 January 2012
(a)
Nkrintra Singhrattna et al.
(b)
(c)
Fig. 1 (a) Map of Thailand, (b) Chao Phraya River Basin, and (c) Ping River Basin and locations of rainfall and temperature
stations.
monsoon from the Indian Ocean in the following
rainy season. The rainy season is caused by the
Inter-Tropical Convergence Zone (ITCZ) and by the
southwest monsoon, and lasts from mid-May to midOctober. Due to the low intensity and uncertainty of
rainfall occurrence, the period from May to July is
called the pre-monsoon season, or transition period
from summer to monsoon season. In contrast, heavy
rainfall occurs from August to October. The average
annual rainfall in the country ranges from 1200 to
1600 mm. The winter season is characterized by dry
and cool winds brought by the northeast monsoon
from the mid-latitudes, and lasting from mid-October
to mid-February. In terms of streamflow, Thailand
generates about 289 × 109 m3 per year of total average runoff in 25 major river basins. However, only
38 × 109 m3 per year or 13.1% of the annual runoff
can be stored in reservoirs and supplied to various
sectors for use within the country (RID 2007).
The Chao Phraya River Basin (Fig. 1(b)) is the
largest among the 25 major basins, covering an area
of 178 000 km2 or 35% of the country’s land area.
Four major tributaries, the Ping, Wang, Yom and
Nan rivers, merge at Nakhon Sawan and form the
Chao Phraya River. The portion above the confluence
at Nakhon Sawan is called the Upper Chao Phraya
River Basin, whereas that below the confluence is
the Lower Chao Phraya River Basin. The upper basin
covers an area of 102 635 km2 or 58% of the Chao
Phraya River Basin.
The study basin, the Ping River Basin, is located
in northern Thailand, between 15◦ –19◦ N latitude
and 98◦ –100◦ E longitude, and covers an area of
33 899 km2 (Fig. 1(c)). The Ping River is 740 km long
and many storage dams are located along the river and
its tributaries. The most important is the Bhumipol
Dam, constructed for the purposes of hydropower,
agriculture, flood mitigation, fishery and transportation. It is 154 m high and 486 m long with a maximum
storage capacity of 13.462 × 109 m3 and installed
capacity of 779.2 MW of hydropower. In terms of
land use, 71.46% of basin area is covered with forests,
which are mostly located in the upstream or the origin of the Ping River. The remainder, along both sides
of the river and the plain downstream, is covered by
agricultural and residential areas and water bodies.
3 DATA
3.1 Hydroclimatic data
Two hydroclimate data sets—rainfall and
temperature—from the climatic stations located
in and around the Ping River Basin, are used in this
study. Both data are recorded on a daily basis. Out of
208 rainfall stations operated by the Royal Irrigation
Hydroclimate variability and long-lead forecasting of rainfall over Thailand by large-scale atmospheric variables
Department (RID), Thailand Meteorological Department (TMD) and Department of Water Resources
(DWR) of the Royal Thai Government, 50 stations
are selected based on the length of time series and
the occurrence of incomplete data. Daily rainfall data
observed at the selected stations are obtained for the
period 1950–2007, with incomplete data constituting
less than 5% of the recent 30 years’ data. Fisher’s
Transformation (z ) is used to test the hypothesis on
significance of a sample correlation coefficient by
(Haan 2002):
1+r
z = 0.5 ln
1−r
(1)
where r is the correlation coefficient between two
independent samples, and z is approximately normally distributed, and is used to compare with z of the
standard normal distribution to obtain the significance
level (p) of correlation.
Coefficients of correlation obtained for monthly
rainfall of all selected stations varied from 0.46 to
0.96 and are well correlated at the 95% significance
level by Fisher’s Transformation. Therefore, the average values are calculated over 50 selected stations.
For daily air temperature, 11 stations of the TMD
are selected with the length of data varying from 16 to
56 years. The average is estimated over 11 selected
stations which are statistically correlated at 95% confidence level. The locations of selected rainfall and
temperature stations are shown in Fig. 1.
the region of 2◦ N–2◦ S latitude and 70◦ –90◦ E longitude. The ION data set is provided by the Physical
Sciences Division of NOAA (PSD 2007b).
The SST over the South China Sea (SCS) is the
SST averaged over the region of 5◦ N–5◦ S latitude and
100◦ –115◦ E longitude. The SCS data set is provided
by the PSD (PSD 2007a).
3.3 Large-scale atmospheric data
The large-scale atmospheric variables used in this
study are obtained from re-analysis-derived data,
which are provided by the NCEP/NOAA (Kalnay
et al. 1996). Several variables, on both a daily and
a monthly basis, are available from 1948 to the
present, and cover a global grid of 2.5◦ latitude ×
2.5◦ longitude (PSD 2007a). This study uses monthly
data of four principal atmospheric variables from
1948–2007: the surface air temperature (SAT); sealevel pressure (SLP); surface zonal wind (SXW); and
surface meridian wind (SYW). These variables play
an important role in strengthening or weakening a
monsoon, and influence the convection over the study
basin. The significant ability of forecasting models
can be obtained using combinations of these variables
(Sahai et al. 2003).
4 CLIMATE DIAGNOSTIC
4.1 Inter-annual climate
The annual cycles of air temperature and rainfall in
the Ping River Basin are shown in Fig. 2. The summer
3.2 Standard sea-surface temperature (SST)
250
This study used monthly data of SST for 1950–2007
over four regions: NINO 3, NINO 4, ION and SCS, as
follows. The standard SSTs over the tropical Pacific
Ocean used in this study are:
These SSTs are provided by the Climate Prediction
Center (CPC 2007) of the US National Oceanic and
Atmospheric Administration (NOAA).
The standard SST over the Indian Ocean
(ION) is the SST obtained from two data sets:
the Comprehensive Ocean–Atmosphere Data Set
(COADS) (Woodruff et al. 1993) and real-time surface marine data from the National Centers for
Environmental Prediction (NCEP), and averaged over
32
Rainfall
Temperature
200
30
150
28
100
26
50
24
Temperature (°C)
– NINO 3: average SST over the area of 5◦ N–5◦ S
latitude and 90◦ –150◦ W longitude; and
– NINO 4: average SST over the area of 5◦ N–5◦ S
latitude and 150◦ W–160◦ E longitude.
Rainfall (mm)
Downloaded by [Nkrintra Singhrattna] at 16:41 23 January 2012
29
22
0
J
F
M
A
M
J
J
A
S
O
N
D
Fig. 2 Annual cycle of mean monthly air temperature from
11 selected stations for 1951–2006 and mean monthly
rainfall from 50 selected stations for 1950–2007.
season occurs during March-April-May (MAM). The
maximum temperature of 30.2◦ C is observed in April,
and the minimum of 22.9◦ C in December. Figure 2
shows that rainfall is bi-modal with two peaks: one in
May and another in August. Considering the primary
peak, the rainy or monsoon season in the basin occurs
during August-September-October (ASO). Moreover,
the secondary peak is observed during May-June-July
(MJJ), which is the pre-monsoon season. The secondary peak of the pre-monsoon is related to the
southwest monsoon and the ITCZ passing from the
Indian Ocean to Thailand in May and to the South
China Sea and central China in mid-June. However,
the primary peak is associated with the ITCZ as
it moves back over Thailand during ASO. From
58 years of rainfall data, the maximum MJJ rainfall of 651.8 mm and the maximum ASO rainfall
of 948.3 mm are found in 1950. The minimum MJJ
rainfall of 254.0 mm is in 1997 and the minimum
ASO rainfall of 387.7 mm in 2004. The total annual
rainfall varies from 843.0 to 1605.6 mm. The MJJ
and ASO rainfall is about 88% of the total annual
rainfall. The remaining rainfall is the dry-season rainfall which falls in November–April and ranges from
49.4 to 295.2 mm over 58 years.
4.2 Inter-seasonal climate
MAM Temperature (°C)
The air temperature is related to rainfall in terms of
developing the temperature gradient between land and
sea that subsequently strengthens the monsoon. In this
study, the summer season temperature is averaged
for MAM. Figure 3 shows the inverse relationship
between the dry-season rainfall and MAM air temperature. As expected, if more rainfall is received
during November–April, it would tend to cool the land
and atmosphere, thus decreasing the air temperature
during MAM and vice versa.
31
30
29
28
T = -0.0011R + 29.47
27
50
100
150
200
250
Dry Season Rainfall (mm)
300
Fig. 3 Scatter plots between dry-season (November–April)
rainfall and March-April-May (MAM) air temperature
(T: temperature; R: rainfall).
MJJ Rainfall (mm)
Nkrintra Singhrattna et al.
800
(a)
R = -16.103T + 914.42
600
400
800
200
(b)
600
400
R = 18.276T + 23.72
200
27.5 28.0 28.5 29.0 29.5 30.0 30.5 31.0
MAM Temperature (°C)
ASO Rainfall(mm)
Downloaded by [Nkrintra Singhrattna] at 16:41 23 January 2012
30
Fig. 4 Scatter plots between March-April-May (MAM)
air temperature and (a) May-June-July (MJJ) rainfall, and
(b) August-September-October (ASO) rainfall (T: temperature; R: rainfall).
The MAM temperature subsequently plays a role
in driving pre-monsoon and monsoon season rainfall.
Figure 4 shows the relationships between the MAM
air temperature and both MJJ rainfall and ASO rainfall. During the pre-monsoon season or the transition
period, the land–ocean temperature gradient is not
fully developed. Hence, inverse relationships between
MAM temperature and MJJ rainfall are found. During
the monsoon season, as the positive relationships indicate, the higher MAM air temperature increases the
ASO rainfall, whereas the lower MAM air temperature decreases the ASO rainfall.
To further enhance the understanding of the
land–sea temperature gradient that influences the seasonal rainfall, the relationships between the MAM
surface temperature of nearby oceans and MJJ and
ASO rainfall are investigated (Fig. 5). The positive relationships between MAM air temperature over
the study basin and ASO rainfall are stronger than
those of MJJ rainfall, which is consistent with the
relationships in Fig. 4.
For the ASO rainfall (Fig. 5(b)), the positive correlations at 95% significance level (i.e. upper and
lower bounds of +0.26 and –0.26, respectively) are
associated with the MAM surface temperature especially over the study basin; negative correlations at
99% significance level (between +0.34 and–0.34) are
associated with MAM surface temperature over the
South China Sea. This indicates the developing land–
sea temperature gradient.
To confirm the influence of land–sea temperature
gradient on the monsoon rainfall, the differences in
MAM temperature over the study basin and over the
Hydroclimate variability and long-lead forecasting of rainfall over Thailand by large-scale atmospheric variables
31
(a)
Downloaded by [Nkrintra Singhrattna] at 16:41 23 January 2012
(b)
Fig. 5 Correlation maps for 1950–2007 between March-April-May (MAM) surface air temperature and (a) May-June-July
(MJJ) rainfall, and (b) August-September-October (ASO) rainfall. The 95% and 99% significance levels correspond to
±0.26 and ±0.34, respectively.
regions of nearby oceans such as the Pacific Ocean
(NINO 3 and NINO 4), the Indian Ocean (ION)
and the South China Sea (SCS) are estimated and
correlated to the MJJ and ASO rainfall (Table 1).
With the 90% confidence level, the significant
Table 1 Cross-correlations between seasonal rainfall during May-June-July (MJJ) and August-September-October
(ASO) and temperature differences during March-AprilMay (MAM) over the study basin and four different
regions over the oceans.
Season
Correlation coefficient:
NINO 3
MJJ rainfall
ASO rainfall
relationships between MJJ rainfall and MAM temperature differences are found to correspond to the
NINO 3 region in the Pacific Ocean. The ASO rainfall
develops stronger relationships than MJJ rainfall with
the MAM temperature differences corresponding to
the regions of NINO 4, ION and SCS. Thus, the
land–sea temperature gradients developed from different sources have an influence on the seasonal rainfall. However, more investigation and detailed study
are required to distinguish the influences on local
hydroclimates.
∗
0.23
0.18
NINO 4
ION
SCS
0.04
0.25∗
−0.07
0.25∗
−0.12
0.35∗∗
NINO 3: a region over the central equatorial Pacific Ocean
between 5◦ N–5◦ S latitude and 90–150◦ W longitude; NINO 4:
a region over the western Pacific Ocean between 5◦ N–5◦ S latitude and 150◦ W–160◦ E longitude; ION: a region over the Indian
Ocean between 2◦ N–2◦ S latitude and 70–90◦ E longitude; SCS:
a region over the South China Sea between 5◦ N–5◦ S latitude
and 100–115◦ E longitude. Figures with ∗ and ∗∗ represent the
correlations being significant at 90% and 99% confidence level,
respectively.
4.3 Trend of seasonal hydroclimates
To support the influences of temperature on the seasonal rainfall and to investigate the variability of
seasonal hydroclimates over the decades, trend analysis of seasonal rainfall and temperature is carried
out. The time series of seasonal rainfall, i.e. dry
(November–April), pre-monsoon (MJJ) and monsoon
(ASO) season, are presented in Fig. 6, and the time
series of MAM temperature are shown in Fig. 7.
Based on the linear trend, the dry-season rainfall
Rainfall (mm)
32
Nkrintra Singhrattna et al.
1000
(a)
Time series
Linear regression fit
800
600
P = 0.2858t - 424
400
200
1000
0
(b)
800
600
400
(c)
800
600
400
200
P = -1.8239t + 4167
2005
2010
2000
1990
1995
1985
1975
1980
1970
1960
1965
1950
0
1955
Rainfall (mm)
0
1000
Year (t)
Fig. 6 Time series of rainfall during: (a) the dry season
(November–April), (b) the pre-monsoon season (MJJ), and
(c) the monsoon season (ASO) (P: rainfall).
31
T = -0.0107t + 50.379
30
29
5 PREDICTOR IDENTIFICATION
28
5.1 Correlations with large-scale atmospheric
variables
Time series
Linear regression fit
2010
2005
2000
1995
1990
1985
1980
1975
1970
1965
1960
1955
27
1950
MAM Temperature (°C)
Downloaded by [Nkrintra Singhrattna] at 16:41 23 January 2012
200
Rainfall(mm)
P = -0.6109t + 1649
where β̂ is the slope of a fitting regression (yi = α +
β̂xi + εi ), β 0 is the specific value for testing, i.e. 0 in
this case, and SE is the standard error (ε i ) of leastsquares of the estimates.
The decreasing trend of MAM temperature subsequently causes the decreasing trends in the MJJ
and ASO rainfall, as shown in Fig. 6, due to the
weakening of development of land–sea temperature
gradient. For the period 1950–2007, the MJJ rainfall shows a decreasing trend of 34.2 mm, and the
ASO rainfall shows a decreasing trend of 102.1 mm.
It is also important to note that the years reported
as the warm phase of ENSO or El Niño (CPC 2006,
COAPS 2006) coincided with below-normal MJJ and
ASO rainfall, and vice versa during the La Niña years.
This is corroborated by Singhrattna et al. (2005b)
and Chen and Yoon (2000). Moreover, the decreasing trend of rainfall is associated with that reported
by the Intergovernmental Panel on Climate Change
(Trenberth et al. 2007). Based on historical observations (1900–2005), they found that precipitation over
the areas between 10◦ N and 30◦ N tended to decline
since 1970 due to global warming and climate change.
However, the precipitation variability is dependent
upon region and season.
Year (t)
Fig. 7 Time series of MAM air temperature (T).
shows a slightly increasing trend of 16.0 mm over
57 years. Based on the inverse correlations between
dry-season rainfall and MAM temperature (Fig. 3),
the increasing trends of dry-season rainfall tend to
decrease the MAM temperature, as shown in Fig. 7.
There is a decrease of 0.6◦ C in the MAM temperature
over 56 years of data. This trend is statistically significant at 95% confidence level by the standard t-test
(Haan 2002), which is applied to test a null hypothesis of the slope of a regression line or linear trend
being different from a specific value. The standard
t-test equation is as follows:
t=
β̂ − β0
SE
(2)
The climate diagnostic shows that the increasing
land–sea temperature gradient strengthens the monsoons and influences the rainfall. Since an ultimate
objective of this study is to develop a model to forecast rainfall at long lead times using the large-scale
atmospheric variables, this section aims to develop
the relationships between rainfall in the study basin
and other large-scale atmospheric variables such as
SAT, SLP, SXW and SYW at various lead times
to identify the predictors of rainfall. The relationships are developed using correlation maps which
are the interactive plots and analysis provided by
the Earth System Research Laboratory (ESRL 2008).
Although the physical long lead–lag relationship is
not plausible, Sahai et al. (2003) found that the correlation patterns between Indian summer monsoon
rainfall (ISMR) and SST show a slow and consistent temporal evolution which addresses the significant lead time relationships of SST even four years
prior to the monsoon season. Clark et al. (2000)
also used various Indian Ocean indices at long lead
Downloaded by [Nkrintra Singhrattna] at 16:41 23 January 2012
Hydroclimate variability and long-lead forecasting of rainfall over Thailand by large-scale atmospheric variables
times to develop the relationships for rainfall predictions. Nicholls (1983) found relationships between
ISMR and 16-month leading SST near Indonesia that
can show the long lead time relationships and the
influences of large-scale atmospheric circulation at
distance from the source (Saravanan and Chang 2000,
Harshburger et al. 2002, Tereshchenko et al. 2002).
Therefore, in this study, the correlation maps are
adopted using the large-scale atmospheric variables,
e.g. SAT, SLP and surface wind with varying lead
times from 4 to 15 months prior to season starting.
The correlation maps also cover the oceans, e.g. the
Indian Ocean, the Pacific Ocean and the South China
Sea that indicate influences on rainfall in the study
basin. The predictors are selected based on longleading relationships and over regions giving high
correlations at 95% significance level. Then, the optimal combination cases of predictors will be identified
by cross-validated multiple regression and used in the
forecasting model.
33
estimation at the dropped point obtained from the
fitting regression, n is the total number of observed
data, and m is the number of independent variables or
predictors applied to the fitting multiple regression.
From the 2k – 1 combination cases, the best combination of predictors is selected due to the minimum
GCV which is associated with the smallest error from
the fitting regression using the minimum number of
mutually exclusive variables.
6 FORECASTING MODEL
6.1 Description of modified k-nearest neighbour
(k-nn) model
The nonparametric regression is an alternative of
the fitting data model. The regression equation is as
follows:
y = f (x1 , x2 , x3 , . . . , xk ) + e
(4)
5.2 Combination cases of the predictors
To avoid redundancy in the case of several identified predictors, the optimal sub-set of predictors that
is composed of the minimum number of mutually
exclusive variables is selected by model selection
methods. In this study, a cross-validated multiple
regression, namely the generalized cross-validation
(GCV) with the leave-one-out technique is adopted
to select an optimal combination of predictors. For k
multiple independent variables, there are 2k – 1 crosscombination cases of variables. Each combination
case is evaluated using the leave-one-out technique,
which is applied to all points of data (n) by dropping
one point out of the dependent (yi ) and independent
(xi ) data set. Then, the regression is developed to fit
the remaining points of data (n – 1). At the dropped
point, the estimation of dependent variable (yi ) is
done with the developed regression using the dropped
independent variables (xi ). The residual (yi – yi ) at
this point is also calculated. The procedure is repeatedly applied to all points of data (n), and the GCV is
calculated by:
GCV =
2
n yi − y i
n
i=1
(1 − m/n)2
(3)
where yi is observed data at the dropped point
(xi ) based on the leave-one-out technique, yi is the
where f is the fitting function of the independent variables x1 , x2 , x3 , . . ., xk (in this case, the identified
large-scale atmospheric variables); y is the dependent
variable (in this case, seasonal rainfall); and e is the
error or residual assumed to be normally distributed
with mean = 0 and variance = σ .
Unlike the parametric approach, the prior
assumption of relationships between dependent and
independent variables such as linear relationships,
which is one of the drawbacks of parametric models, is not required for the nonparametric approach.
The fitting functions (f ) capture relationships locally
or by a small set of neighbours (k) at a given point
(xi ). Because the functions are flexible and able to
fit any arbitrary data, for example, bivariate data
and multivariate data, the nonparametric approach
can determine relationships better than the parametric regression. Moreover, a drawback of global fitting
in the parametric approach, where all points of data
may have a large influence on an individual point of a
curve, can be solved by the local fitting functions.
There are several approaches of nonparametric
regression. One that grants the discontinuities of the
derivative curve is the spline approach. Another that
applies a regression locally to a given point (xi )
and the neighbours around xi is referred to as the
local polynomials approach. This approach includes
locally-weighted polynomials (Loader 1999) and knearest neighbour (k-nn) local polynomials (Owosina
34
Nkrintra Singhrattna et al.
Downloaded by [Nkrintra Singhrattna] at 16:41 23 January 2012
1992, Rajagopalan and Lall 1999). Two parameters—
the size of neighbourhood (k), referred to as bandwidth, and the order of polynomials (p)—are required
to develop the fitting regression. Both parameters
can be estimated by various criteria and objective
functions such as GCV and likelihood.
In terms of forecast, the modified k-nn model
is adopted to forecast seasonal rainfall using the
identified large-scale atmospheric predictors as the
independent variables (xi ). The fitting of a regression,
rainfall forecast and resample data to achieve rainfall
ensembles are described in the following steps:
1. For the fitting process, the size of neighbours (k)
and the order of polynomial (p), normally 1 or
2, will be selected as the combination of k and
p obtaining the minimum GCV score by:
GCV(k, p) =
2.
3.
4.
5.
n e2
i
n
i=1
(1 − m/n)2
(5)
where ei is the error, n is the number of data points
(xi ), and m is the number of parameters.
The regression is fitted locally with k and p
obtained from Step 1.
The rainfall according to the developed fitting
regressions is then estimated and called the mean
estimations (ȳ1 , ȳ2 , ȳ3 , . . . , ȳn ). Then, the residuals (e1 , e2 , e3 , . . . , en ) are computed.
At a new point of independent variable (xnew ), the
forecast of rainfall is required. The mean estimation (ȳnew ) is obtained from the developed fitting
regression.
An ensemble forecast is then obtained by adding
the residual (ei ) to ȳnew . The residual ei corresponds to one of the k-nearest neighbours (k-nn)
of xnew which is randomly selected by using a
weight function as follows:
W (j) =
1/j
k
(6)
(1/i)
i=1
where W (j) is the weight value of a neighbour of
xnew whose distance from xnew falls in the jth rank,
and k is the size of the neighbours, which does
not have to
√ be the same as for fitting polynomial.
The term n − 1 is used, in practice, to estimate k
where n is the total number of xi . The weight function gives more weight to the nearest neighbour
and less weight to the farthest neighbour. From
the weight equation, the distances between xnew
and all points of xi have to be estimated. There
are several methods to calculate distance between
two points of data, such as the Euclidean distance and Mahalanobis distance. The Euclidean
distance is applied in this study and the equations
of the Euclidean distance for univariate and multivariate data are shown in equations (7a) and (7b),
respectively:
n
di = (xnew − xi )2
(7a)
i=1
2
m di = xnew,j − xi,j
(7b)
j=1
where i = 1, 2, 3, . . . , n, and m is the number of
predictors.
6. Repeat Step 5 as many times as required to
achieve the forecasting ensembles of rainfall, or a
number of simulations (N); in this case N = 300.
7. Repeat steps 4–6 for each forecasting point.
The N ensembles of forecasting rainfall can
develop a box-plot or a probability density function (pdf). It can provide an exceedence and a nonexceedence probability of the extreme events, e.g. wet
or dry. The exceedence and non-exceedence probabilities can support water resources managers to
implement a strategic planning of water allocation
and management and agricultural practices.
6.2 Model evaluation
The likelihood function (lhf) is applied to evaluate the modified k-nn model on capturing the pdf
which is an essential tool of decision making for
water resources management. The initial step of lhf
calculation is to divide the climatological data of
1950–2007 into three categories: below-normal (dry),
normal and above-normal (wet). In this case, three
categories are determined separately for each season
of rainfall by dividing the seasonal rainfall data at the
33rd percentile and the 67th percentile: the rainfall
data below the 33rd percentile fall into the dry category, whereas those above the 67th percentile are
in the wet category; the remaining data are in the
normal category. Next, the categorical probabilities
of climatology that are the proportion of historical
data in each category are calculated. Subsequently,
for a given year, rainfall ensembles in each category
Hydroclimate variability and long-lead forecasting of rainfall over Thailand by large-scale atmospheric variables
(out of 300 ensemble members obtained from the
modified k-nn model) are also determined and the
categorical probabilities are estimated. The categorical probabilities of model ensembles for a given year
are calculated by the proportion of ensemble members falling in each category—in this case, they are
computed separately for each season, year and forecasting lead time. Since the historical rainfall data in
this case are divided at the 33rd and 67th percentiles,
the climatological categorical probability of each category is one third. Then, the lhf for a given year is
estimated as:
Downloaded by [Nkrintra Singhrattna] at 16:41 23 January 2012
n
lhft =
t=1
n
P̂j,t
(8)
Pcj,t
t=1
where n is the number of years; j is the category
of the observed value in year t; P̂j,t is the probability of rainfall ensembles in category j in year
t; P̂j,t = (P̂1,t , P̂2,t , P̂3,t , . . . , P̂k,t ), in which k is the
number of categories; and Pcj,t is the climatological
probability of category j in year t, which in this case
indicates an equal value for all three categories, i.e.
1/3. For example, for 1972, the observed MJJ rainfall
falls in the dry category (i.e. <410 mm, or less than
the 33rd percentile of 1950–2007 MJJ rainfall). Out
of 300 ensemble members obtained from the modified k-nn model, there are 212 ensembles having less
than 410 mm. Therefore, P̂dry,1972 = 212/300, and
lhf1972 = (212/300)/(1/3) = 2.12.
The lhf score varies from 0.0 to number of categories, i.e. 3.0 in this study. A score of 0.0 indicates
lack of model performance in capturing the pdf of
climatology, while a score greater than 1.0 indicates
better performance than the climatology. A score of
3.0 indicates perfect performance of the model.
7 RESULTS AND DISCUSSION
7.1 Identifying predictors and combination cases
To develop a long-lead forecasting model, the statistical relationships between rainfall during MJJ,
ASO, NDJ and FMA and the large-scale atmospheric variables are developed by correlation maps
with varying lead times of the atmospheric variables. Figure 8 shows examples of developed relationships between pre-monsoon season (MJJ) rainfall
and SAT for 12 different lead times varying from
35
4 to 15 months before the beginning of the premonsoon season in May. The solid box represents the
selected region of predictor which has high correlation at the 95% significance level (upper and lower
bounds of +0.26 and –0.26, respectively). Statistical
relationships from correlation maps were also developed for the variables SLP, SXW and SYW and ASO,
NDJ and FMA rainfall data (not shown). The significant relationships of monsoon rainfall (i.e. MJJ and
ASO rainfall) show development with the large-scale
atmospheric variables over the study basin and the
nearby seas and oceans, e.g. South China Sea, the
Indian Ocean and the Pacific Ocean. The developed
relationships are found to be associated with long
lead times of atmospheric variables of 8–15 months
prior to the start of the season. This is corroborated by Sahai et al. (2003) and Nicholls (1983),
who presented long-lead relationships between Indian
summer monsoon rainfall and atmospheric variables,
e.g. SST. The significant relationships between monsoon (i.e. MJJ and ASO) rainfall and large-scale
atmospheric variables suggest the ability of the rainfall forecasting models to make long-lead forecasts
using the identified atmospheric predictors, which is
hardly found in any study of Thailand rainfall forecasting (Hung et al. 2009, Singhrattna et al. 2005a,
Weesakul and Lowanichchai, 2005). However, the
dry-season (i.e. NDJ and FMA) rainfall gave shorter
lead-time relationships with large-scale atmospheric
variables of 7–10 months. Furthermore, significant
links are found over distant regions, e.g. Sumatra
and Java (Indonesia), and northeast India, which
shows the remote influence of atmospheric circulation
(Harshburger et al. 2002, Tereshchenko et al. 2002).
Table 2 summarizes the identified predictors of largescale atmospheric variables. It is also noted that the
significant relationship between SLP and NDJ rainfall
is hardly found over the region, e.g. the study basin,
the Pacific Ocean and the South China Sea. Moreover,
the observed rainfall data (RAIN) are included in the
identified predictors because, from the lag autocorrelation analysis (Fig. 9), the negative correlations at
95% significance level are associated with a 6-month
lag (and the positive correlations – with a 12-month
lag), which indicates the ability of forecasting models
to give long-lead forecasts using RAIN as the predictor. Hence, there are five identified predictors each of
the MJJ, ASO and FMA rainfall, and four predictors
for the NDJ rainfall.
Subsequently, an optimal combination of predictors is selected by cross-validated multiple regression. Based on k independent predictors, the 2k – 1
Downloaded by [Nkrintra Singhrattna] at 16:41 23 January 2012
36
Nkrintra Singhrattna et al.
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
(i)
(j)
(k)
(l)
Fig. 8 Correlation maps between MJJ rainfall and surface air temperature (SAT) at 15- to 4-month lead times, respectively:
(a) FMA, (b) MAM, (c) AMJ, (d) MJJ, (e) JJA, (f) JAS, (g) ASO, (h) SON, (i) OND, (j) NDJ, (k) DJF and (l) JFM. The 95%
significance levels correspond to ±0.26.
cross-combination cases can be set. In this study, the
optimal sub-sets are identified separately for each season of rainfall and each lead period of predictors that
can provide forecasts varying from 1 to 12 months
prior to the season starting. For example, for MJJ
rainfall, the 15-month lead-time predictors, which are
averaged during February-March-April (FMA) in the
current year, are used to forecast the MJJ rainfall
of the following year. The forecast can be issued on
1 May, i.e. 12 months in advance. From the GCV,
an optimal sub-set of predictors is selected from the
2k – 1 cross-combination cases corresponding to the
minimum GCV. Table 3 summarizes the best combination cases of predictors for rainfall during MJJ,
ASO, NDJ and FMA. The selected combinations consist of 1–2 mutually exclusive predictors, except for
Hydroclimate variability and long-lead forecasting of rainfall over Thailand by large-scale atmospheric variables
Table 2 Summary of the large-scale atmospheric predictors identified from the correlation maps for seasonal
rainfall during May-June-July (MJJ), August-SeptemberOctober (ASO), November-December-January (NDJ) and
February-March-April (FMA).
Season of
rainfall
MJJ
NDJ
FMA
20.0◦ N
7.5–10.0◦ N
0◦
0–2.5◦ N
15.30–19.36◦ N
2.5–5.0◦ N
17.5–20.0◦ N
10.0◦ N
10.0◦ N
15.30–19.36◦ N
2.5–7.5◦ S
2.5–5.0◦ S
20.0–22.5◦ N
15.30–19.36◦ N
7.5◦ S
15.0◦ N
17.5◦ N
2.5◦ S–2.5◦ N
15.30–19.36◦ N
SAT
SLP
SXW
SYW
RAIN
SAT
SLP
SXW
SYW
RAIN
SAT
SXW
SYW
RAIN
SAT
SLP
SXW
SYW
RAIN
Longitude
97.5–102.5◦ E
102.5–107.5◦ E
82.5–87.5◦ E
172.5–175.0◦ E
98.05–100.70◦ E
107.5–110.0◦ E
97.5–100.0◦ E
100.0–102.5◦ E
95.0–97.5◦ E
98.05–100.70◦ E
97.5–102.5◦ E
62.5–67.5◦ E
85.0◦ E
98.05–100.70◦ E
110.0–112.5◦ E
187.5–192.5◦ W
140.0–150.0◦ E
95.0–97.5◦ E
98.05–100.70◦ E
ACF
0.0
0.5
1.0
SAT: surface air temperature; SLP: sea-level pressure; SXW: surface zonal wind; SYW: surface meridian wind; RAIN: observed
rainfall over study basin.
–0.5
Downloaded by [Nkrintra Singhrattna] at 16:41 23 January 2012
ASO
Atmospheric Spatial coverage
variables
Latitude
0
5
10 15
Lag
20
Fig. 9 Correlogram or lag autocorrelation (ACF) of
monthly rainfall for 1950–2007. The dashed lines represent
the correlations being significant at the 95% confidence
level.
two combination cases for ASO rainfall corresponding to the 6- and 10-month forecasting lead times
that address three predictors of SAT, SXW and SYW,
and SAT, SYW and RAIN, respectively. The identified sub-sets indicate that the univariate regression of
large-scale atmospheric variables (e.g. SXW or SYW)
and the multivariate regression (e.g. SYW and RAIN
37
or SLP and SXW) are more efficient than the univariate regression of observed rainfall (i.e. RAIN). This
suggests the better performance of a fitting regression and forecasting model with the incorporation
of large-scale atmospheric predictors, e.g. SAT and
SLP (Basson and Rooyen, 2001, Trenberth et al.
2006, Sadhuram and Ramana Murthy, 2008) compared to developing a regression from only lagging
relationships of hydroclimatic variables, e.g. RAIN.
7.2 Forecasting evaluation
The lhf scores used to evaluate the modified k-nn
model with the leave-one-out technique are computed separately for each season of rainfall, for each
year of 1950–2007 and each forecasting lead time.
Figure 10 shows plots of median lhf for seasonal rainfall during MJJ, ASO, NDJ and FMA. A score of
0.0 indicates lack of model performance in capturing the pdf of climatology, whereas a score greater
than 1.0 indicates better model performance than climatology. The darker shading represents higher lhf
scores or better model performance. From the median
lhf of 1950–2007 rainfall ensembles (Fig. 10(a)), high
scores can be found corresponding to the forecast
of MJJ and ASO rainfall at long lead times (e.g.
7–9 months and 11 months). The successful longlead forecasting of the modified k-nn model indicating higher skills than climatology (i.e. lhf scores
greater than 1.0) was hardly found in previous studies (Kulkarni, 2000, DelSole and Shukla, 2002). The
modified k-nn model also performed better than the
climatology of MJJ and ASO rainfall at short forecasting lead times, e.g. 1–2 months (Singhrattna et al.
2005a, Grantz et al. 2005). However, associated with
lhf scores of less than 1.0, the forecasting performance of the model with dry-season (i.e. NDJ and
FMA) rainfall decreases considerably even at short
lead times because, in terms of physical mechanism,
the dry-season rainfall over the study basin is influenced by unstable local conditions, e.g. higher surface
temperature and humidity, at finer time scales, e.g.
hourly and daily (TMD 2007). Hence, the modified
k-nn model, which is developed based on the relationships between rainfall and large-scale atmospheric
variables at seasonal time scales, performs better in
monsoon rainfall forecasting than dry-season rainfall
forecasting.
To compare the prediction ability of the modified k-nn model in extreme years, Fig. 10(b) and (c)
present plots of median lhf for dry rainfall years (rainfall below the 33rd percentile) and wet rainfall years
38
Nkrintra Singhrattna et al.
Table 3 Summary of the optimal sub-sets of predictors (•) identified from the cross-validated multiple regression for seasonal rainfall during (a) May-June-July (MJJ), (b) August-September-October (ASO), (c) November-December-January
(NDJ) and (d) February-March-April (FMA). (SAT: surface air temperature; SLP: sea-level pressure; SXW: surface zonal
wind; SYW: surface meridian wind; RAIN: observed rainfall over study basin.)
Predictor
SAT
SLP
Season of predictors (forecasting lead time in month)
SXW
SYW
RAIN
•
•
•
•
•
•
Downloaded by [Nkrintra Singhrattna] at 16:41 23 January 2012
(a) MJJ Rainfall
FMA (12)
MAM (11)
AMJ (10)
MJJ (9)
JJA (8)
JAS (7)
ASO (6)
SON (5)
OND (4)
NDJ (3)
DJF (2)
JFM (1)
(b) ASO Rainfall
MJJ (12)
JJA (11)
JAS (10)
ASO (9)
SON (8)
OND (7)
NDJ (6)
DJF (5)
JFM (4)
FMA (3)
MAM (2)
AMJ (1)
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
(c) NDJ Rainfall
ASO (12)
SON (11)
OND (10)
NDJ (9)
DJF (8)
JFM (7)
FMA (6)
MAM (5)
AMJ (4)
MJJ (3)
JJA (2)
JAS (1)
•
•
•
•
•
•
•
•
•
•
•
•
•
•
(d) FMA Rainfall
NDJ (12)
DJF (11)
JFM (10)
FMA (9)
MAM (8)
AMJ (7)
MJJ (6)
JJA (5)
JAS (4)
ASO (3)
SON (2)
OND (1)
•
•
•
•
•
•
•
•
•
•
•
•
Hydroclimate variability and long-lead forecasting of rainfall over Thailand by large-scale atmospheric variables
2
8
6
1
4
0
FMA
2
NDJ
2
3
10
ASO
4
12
MJJ
FMA
NDJ
MJJ
2
Forecasting Lead Time (month)
4
6
NDJ
6
8
FMA
8
10
ASO
10
12
MJJ
Forecasting Lead Time (month)
12
ASO
Forecasting Lead Time (month)
Downloaded by [Nkrintra Singhrattna] at 16:41 23 January 2012
(c)
(b)
(a)
39
Fig. 10 Median likelihood (lhf) scores for seasonal rainfall during May-June-July (MJJ), August-September-October (ASO),
November-December-January (NDJ) and February-March-April (FMA) for: (a) 1950–2007, (b) dry years, and (c) wet years.
(rainfall above the 67th percentile), respectively. For
MJJ rainfall, the lhf scores of dry years are higher than
those of wet years, especially at long lead times, e.g.
9–12 months. In contrast, the modified k-nn model
shows the greater predictability of ASO rainfall in
wet years than in dry years. However, the highest
lhf scores (i.e. 2.0–2.5) for ASO rainfall forecast in
both dry and wet years can be found associated with
the 6-month forecasting lead time, which shows an
impressive ability to forecast extreme events with a
long lead period. It is also noted that the better performance than climatology (i.e. lhf greater than 1.0) of
wet ASO rainfall can be observed for all 12 forecasting lead times. In addition, the good performance of
long-lead rainfall forecasts is obtained for dry NDJ
(and short-lead forecasts for wet NDJ). For FMA
rainfall, the forecast of dry years shows better performance than wet years, in particular, at the forecasting
lead time of 5–6 months.
Therefore, the large-scale atmospheric variables
are identified and used as predictors of the modified k-nn model to forecast seasonal rainfall in the
Ping River Basin for 12 forecasting lead periods.
The optimal combination sets of predictors that are
determined by GCV consist of 1–2 predictors. This
can confirm the comparative performance of a fitting
regression and forecasting model using the largescale atmospheric predictors. Since the irrigated area
of the study basin covers 2332 km2 , and the agriculture depends on both the rainfall during the monsoon season and the irrigated water (also based on
rainfall), the ability of the modified k-nn model to
make long-lead forecasts for monsoon rainfall, e.g.
7–9 months prior to the season starting, is very useful
for water resources planning and management, agricultural practices and cropping strategies. Moreover,
the operation planning of several storage dams along
the Ping River including the Bhumipol Dam can be
done in advance to deal with the extreme events, i.e.
dry and wet episodes, and to efficiently serve the
water demands of the Ping and Chao Phraya river
basins, consisting of 6127 and 11 000 × 106 m3
per year, respectively. The non-exceedence and exceedence probabilities of extreme events obtained from
the pdf of rainfall forecasting ensembles serve as a
tool for decision making in reservoir operation, and
the modified k-nn model shows better performance in
capturing the pdf than the climatology.
8 CONCLUSION
This study enhances the understanding of local hydroclimate variability and develops statistical relationships between rainfall and large-scale atmospheric
variables. The land–sea temperature gradient plays a
key role in strengthening or weakening the monsoon
over the Upper Chao Phraya River Basin in Thailand.
The amount of dry-season (November–April) rainfall
determines the MAM temperature which is responsible for setting up the land–sea temperature gradient.
Higher MAM temperature tends to increase monsoon
(i.e. ASO) rainfall. Due to the anomalous atmospheric
variables, the below-normal rainfall corresponds to
the El Niño phase of ENSO and, conversely, the
above-normal rainfall corresponds to the La Niña
phase.
Using the correlation maps, the large-scale atmospheric variables are identified as the predictors
Downloaded by [Nkrintra Singhrattna] at 16:41 23 January 2012
40
Nkrintra Singhrattna et al.
of seasonal rainfall. The significant relationships
between rainfall and large-scale atmospheric variables indicate the long-lead forecasting ability of
forecasting models. There are five identified predictors for each of MJJ, ASO and FMA rainfall, and
four identified predictors for NDJ rainfall. Based on
the cross-validated multiple regression, optimal combinations of predictors are selected separately for
each season of rainfall and forecasting lead period.
The selected combinations composed of one or two
mutually exclusive predictors present the comparative performance of a regression and forecasting
model by incorporation of the large-scale atmospheric
variables.
The modified k-nn model is ultimately developed
to forecast rainfall at long lead times using the optimal
sets of predictors. Based on lhf scores, the modified
k-nn model shows the ability to give long-lead forecasts of MJJ and ASO rainfall at 7–9 months lead
time. However, decreasing performance is found in
dry-season rainfall forecasts at both short and long
lead times. The ability to forecast rainfall in dry MJJ
is more significant than in wet MJJ, especially at lead
times of 9–12 months. The best performance of ASO
rainfall forecast in extreme years, i.e. dry and wet, is
observed at a 6-month lead time. The long-lead forecasting ability of the modified k-nn model for rainfall
in the Ping River Basin is an impressive result that can
be applied to manage water resources and develop the
long-term plans for reservoirs.
Acknowledgements The authors would like to
thank Dr Sutat Weesakul and Dr Kiyoshi Honda
for their valuable comments. The authors also
thank the Royal Irrigation Department of Thailand,
the Thailand Meteorological Department and
the Department of Water Resources for providing the
data for the study. The first author is grateful to the
Asian Institute of Technology, Thailand for the graduate fellowships, to the Department of Public Works
and Town & Country Planning for its kind support
and to Piyawat Wuttichaikitcharoen for providing
some data used in this research. The research funding
from CIRAD is gratefully acknowledged. Finally, the
authors would like to thank the anonymous reviewers
whose comments improved the manuscript.
REFERENCES
An, S.I., Kug, J.S., Timmermann, A., Kang, I.S. and Timm, O., 2007.
The influence of ENSO on the generation of decadal variability
in the North Pacific. Journal of Climate, 20 (4), 667–680.
Basson, M.S. and Rooyen, J.A., 2001. Practical application of probabilistic approaches to the management of water resource
systems. Journal of Hydrology, 241 (1–2), 53–61.
Cañón, J., González, J. and Valdés, J., 2007. Precipitation in the
Colorado River basin and its low frequency associated with
PDO and ENSO signal. Journal of Hydrology, 333 (2–4),
252–264.
Chen, T.-C. and Yoon, J.-H. (2000) Interannual variation in Indochina
summer monsoon rainfall: possible mechanism. Journal of
Climate, 13 (11), 1979–1986.
Clark, C.O., Cole, J.E. and Webster, P.J., 2000. Indian Ocean SST
and Indian summer rainfall: predictive relationships and their
decadal variablitiy. Journal of Climate, 13, 2503–2519.
Clark, M.P., Serreze, M.C. and McCabe, G.J., 2001. Historical effects
of El Niño and La Niña events on the seasonal evolution of the
montane snowpack in the Columbia and Colorado River basins.
Water Resources Research, 37 (3), 741–757.
COAPS, 2006. ENSO index according to JMA SSTA (1868 – present).
http://coaps.fsu.edu/jma.shtml.
CPC (Climate Prediction Center), 2006. Cold and warm
episodes by season. http://www.cpc.ncep.noaa.gov/products/
analysis_monitoring/ensostuff/ensoyears.shtml.
CPC (Climate Prediction Center), 2007. The tropical Pacific Ocean
SST indices. http://www.cpc.noaa.gov/data/indices.
DelSole, T. and Shukla J., 2002. Linear prediction of Indian monsoon
rainfall. Journal of Climate, 15, 3645–3658.
Eldaw, A.K., Salas, J.D. and Garcia, L.A., 2003. Long-range forecasting of the Nile River flows using climatic forcing. Journal
of Applied Meteorology, 42 (7), 890–904.
ESRL (Earth System Research Laboratory), 2008. Interactive plotting and analysis: linear monthly/seasonal correlations. http://
www.esrl.noaa.gov/psd/data/correlation.
Fasullo, J. and Webster, P.J., 2002. Hydrological signatures relating
the Asian summer monsoon and ENSO. Journal of Climate,
15(21), 3082–3095.
Gershunov, A., 1998. ENSO influence on intraseasonal extreme
rainfall and temperature frequencies in the contiguous United
States: implications for long-range predictability. Journal of
Climate, 11 (12), 3192–3203.
Goswami, B.N., Wu, G. and Yasunari, T., 2006. The annual cycle,
intraseasonal oscillation, and roadblock to seasonal predictability of the Asian summer monsoon. Journal of Climate, 19 (20),
5078–5099.
Grantz, K., Rajagopalan, B., Clark, M. and Zagona, E., 2005.
A technique for incorporating large-scale climate information
in basin-scale ensemble streamflow forecasts. Water Resources
Research, 41, W10410, doi:10.1029/2004WR003467.
Gutiérrez, F. and Dracup, J.A., 2001. An analysis of the feasibility of long-range streamflow forecasting for Colombia using
El Niño-Southern Oscillation indicators. Journal of Hydrology,
246 (1–4), 181–196.
Haan, C.T., 2002. Statistical methods in hydrology, second edition.
Iowa: Iowa State Press.
Hamlet, A.F., Huppert, D. and Lettenmaier, D.P., 2002. Economic
value of long-lead streamflow forecasts for Columbia River
hydropower. Journal of Water Resource Planning – ASCE
128 (2), 91–101.
Harshburger, B., Ye, H. and Dzialoski, J., 2002. Observation evidence
of the influence of Pacific SSTs on winter precipitation and
spring stream discharge in Idaho. Journal of Hydrology, 246
(1–4), 157–169.
Hung, N.Q., Babel, M.S., Weesakul, S. and Tripathi, N.K., 2009.
An artificial neural network model for rainfall forecasting in
Bangkok, Thailand. Hydrology and Earth System Sciences, 13,
1413–1425.
Kalnay et al., 1996. The NCEP/NCAR 40-year reanalysis project.
Bulletin of the American Meteorological Society, 77, 437–470.
Downloaded by [Nkrintra Singhrattna] at 16:41 23 January 2012
Hydroclimate variability and long-lead forecasting of rainfall over Thailand by large-scale atmospheric variables
Kim, T.W., Valdés J.B.., Nijssen, B. and Roncayolo, D., 2006.
Quantification of linkages between large-scale climatic patterns and precipitation in the Colorado River basin. Journal of
Hydrology, 321 (1–4), 173–186.
Krishna Kumar, K., Soman, M.K., and Rupakumar, K. (1995)
Seasonal forecasting of Indian summer monsoon rainfall:
a review. Weather, 50, 449–467.
Krishnan, R., Zhang, C. and Sugi, M., 2000. Dynamics of breaks in
the Indian summer monsoon. Journal of Atmospheric Science,
57 (9), 1354–1372.
Kulkarni, A., 2000. A note on the performance of IMD prediction model for ISMR. Annual Monsoon Workshop of Indian
Meteorological Society, Pune Chapter, 20 December 2000.
Loader, C. (1999) Local regression and likelihood. New York:
Springer.
Masiokas, M.H., Villalba, R., Luckman, B.H., Quesne, C.L. and
Aravene, J.C., 2006. Snowpack variations in the central Andes
of Argentina and Chile, 1951–2005: large-scale atmospheric
influences and implications for water resources in the region.
Journal of Climate, 19 (24), 6334–6352.
Mason, S.J. and Goddard, L., 2001. Probabilistic precipitation
anomalies associated with ENSO. Bulletin of the American
Meteorological Society, 82, 619–638.
McCabe, G.J. and Dettinger, M.D., 2002. Primary modes and predictability of year-to-year snowpack variation in the western
United States from teleconnections with Pacific Ocean climate.
Journal of Hydrometeorology, 3 (1), 13–25.
Meko, D.M. and Woodhouse, C.A., 2005. Tree-ring footprint of
joint hydrologic drought in Sacramento and upper Colorado
river basins, western USA. Journal of Hydrology, 308 (1–4),
196–213.
Mendoza, B. et al., 2005. Historical droughts in central Mexico and
their relation with El Niño. Journal of Applied Meteorology,
44 (5), 709–716.
Nagura, M. and Konda, M., 2007. The seasonal development of an
SST anomaly in the Indian Ocean and its relationship to ENSO.
Journal of Climate, 20 (1), 38–52.
Nicholls, N., 1983. Predicting Indian monsoon rainfall from sea surface temperature in the Indonesia–north Australia area. Nature,
306, 576–577.
Owosina, A., 1992. Methods for assessing the space and time
variability of ground water data. Thesis (MSc), Utah State
University, Utah, USA.
Pavia, E.G., Graef, F. and Reyes, J., 2006. PDO-ENSO effects
in the climate of Mexico. Journal of Climate, 19 (24),
6433–6438.
PSD (Physical Sciences Division, NOAA), 2007a. Create a
monthly/seasonal mean time series from the NCEP reanalysis
data set. http://www.esrl.noaa.gov/psd/cgi-bin/data/timeseries/
timeseries1.pl.
PSD (Physical Sciences Division, NOAA), 2007b. The Indian Ocean
SST index. http://www.esrl.noaa.gov/psd/forecasts/sstlim/
timeseries/index.html.
Rajagopalan, B. and Lall, U., 1999. A k-nearest-neighbor simulation for daily precipitation and other weather variables. Water
Resources Research, 35 (10), 3089–3101.
RID (Royal Irrigation Department), 2007. The 25 major river basins
in Thailand. http://www.rid.go.th.
Sadhuram, Y. and Ramana Murthy, T.V., 2008. Simple multiple
regression model for long range forecasting of Indian summer
41
monsoon rainfall. Meteorology and Atmospheric Physics,
99 (1–2), 17–24.
Sahai, A.K., Grimm, A.M., Satyan, V. and Pant, G.B., 2003.
Long-lead prediction of Indian summer monsoon rainfall
from global SST evolution. Climate Dynamics, 20,
855–863.
Saravanan, R. and Chang, P., 2000. Interaction between tropical
Atlantic variability and El Niño-Southern Oscillation. Journal
of Climate, 13 (13), 2177–2194.
Schöngart, J. and Junk, W.J., 2007. Forecasting the flood-pulse in
Central Amazonia by ENSO-indices. Journal of Hydrology, 335
(1–2), 124–132.
Shrestha, A. and Kostaschuk, R., 2005. El Niño/Southern Oscillation
(ENSO)-related variability in mean-monthly streamflow in
Nepal. Journal of Hydrology, 308 (1–4), 33–49.
Shrestha, M.L., 2000. Interannual variation of summer monsoon rainfall over Nepal and its relation to Southern Oscillation index.
Meteorolical and Atmospheric Physics, 75 (1–2), 21–28.
Silverman, D. and Dracup, J.P., 2000. Artificial neural networks and
long-range precipitation prediction in California. Journal of
Applied Meteorology, 39 (1), 57–66.
Singhrattna, N., Rajagopalan, B., Clark, M. and Kumar, K.K., 2005a.
Forecasting Thailand summer monsoon rainfall. International
Journal of Climatology, 25 (5), 649–664.
Singhrattna, N., Rajagopalan, B., Kumar, K.K. and Clark, M., 2005b.
Interannual and interdecadal variability of Thailand summer
monsoon season. Journal of Climate, 18 (11), 1697–1708.
Smith, S.R. and O’Brien J.J., 2001. Regional snowfall distributions
associated with ENSO: implications for seasonal forecasting.
Bulletin of the American Meteorological Society, 82 (6), 1179–
1191.
Sourza Filho, F.A. and Lall, U., 2003. Seasonal to interannual
ensemble streamflow forecasts for Cearra, Brazil: applications
of a multivariate semiparametric algorithm. Water Resources
Research, 39 (11), doi:10.1029/2002WR001373.
Tereshchenko, I., Filonov, A., Gallesgos, A., Monzón C. and
Rodríguez, R., 2002. El Niño 997–98 and the hydrometeorological variability of Chapala, a shallow tropical lake in Mexico.
Journal of Hydrology, 264 (1–4), 133–146.
TMD (Thailand Meteorological Department), 2007. The natural disasters in Thailand (in Thai). http://www.tmd.go.th/info/risk.
pdf.
Trenberth, K.E. et al., 2007. Observations: surface and atmospheric
climate change. New York: Cambridge University Press.
Trenberth, K.E., Moore, B., Karl, T.R. and Carlos, N., 2006.
Monitoring and prediction of earth’s climate: a future perspective. Journal of Climate, 19 (20), 5001–5008.
Weesakul, U. and Lowanichchai, S., 2005. Rainfall forecast for agricultural water allocation planning in Thailand. Thammasat
International Journal of Science and Technology, 10 (3), 18–27.
Woodruff, S.D. et al., 1993. Comprehensive Ocean – Atmosphere
Data Set (COADS) Release 1a: 1980–92. Earth System
Monitor, 4 (1), 1–8.
Wu, R. and Kirtman, B.P., 2004. Understanding the impacts of the
Indian Ocean on ENSO variability in a coupled GCM. Journal
of Climate, 17 (20), 4019–4031.
Zehe, E., Singh, A.K. and Bárdossy, A., 2006. Modelling of monsoon rainfall for a mesoscale catchment in North-West India
II: stochastic rainfall simulations. Hydrology and Earth System
Sciences 10, 807–815.