Predictability of Persistent Thailand Rainfall during the Mature Monsoon Season in 2011 Using Statistical Downscaling of CGCM Seasonal Prediction

ABSTRACT

Predictability of above-normal rainfall over Thailand during the rainy season of 2011 was investigated with a one-tier seasonal prediction system based on an atmosphere-ocean coupled general circulation model (CGCM) combined with a statistical downscaling method. The statistical relationship was derived using singular value decomposition analysis (SVDA) between observed regional rainfall and the hindcast of tropical sea surface temperature (SST) from the seasonal prediction system, which has an ability to forecast oceanic variability for lead times up to several months. The downscaled product of 2011 local rainfall was obtained by combining rainfall patterns derived from significant modes of SVDA. This method has the advantage in terms of flexibility that phenomenon-based statistical relationships, such as teleconnections associated with El Niño-Southern Oscillation (ENSO), Indian Ocean dipole (IOD), or the newly recognized central Pacific El Niño, are considered separately in each SVDA mode. The downscaled prediction initialized from 1 August 2011 reproduced the anomalously intense precipitation pattern over Indochina including northern Thailand during the latter half of the rainy season, even though the direct hindcast from the CGCM failed to predict the local rainfall distribution and intensity. Further analysis revealed that this method is applicable to the other recent events such as heavy rainfall during the rainy seasons of 2002 and 2008 in Indochina.

(ProQuest: ... denotes formulae omitted.)

1. Introduction

The Thailand floods during the 2011 monsoon season and the following autumn are known to be the worst flooding on record in terms of both the volume of water and the number of people affected. Large parts of the country were submerged including Ayutthaya, the industrial zones, and the vicinity of Bangkok. The disaster brought immense human and social damage. There were 813 flood casualties and approximately 9.5 million people were affected. Economic losses were estimated to be $40 billion (U.S. dollars; from the International Disaster Database, http://www.emdat.be/). Several causes have been reported as the trigger and aggravation factors: persistent and pronounced monsoon rainfall from May (Komori et al. 2012), landfall of the Tropical Storm Nock-ten at the end of July, and active convection over the tropical western Pacific associated with La Niña since August. Moreover, nonmeteorological factors such as land-use change from agricultural to vulnerable industrial areas, changes in the hydrography of the Chao Phraya River, and dam operation policies (Peterson et al. 2012) have also been considered to be potential drivers. Some of these factors are intertwined.

Seasonal-scale predictions conducted by atmosphere- ocean coupled general circulation models (CGCMs) are expected to be useful in minimizing damage induced by such seasonal-scale climate events such as La Niña. Current CGCMs have the ability to predict large-scale tropical variables for up to six months or longer (Jin et al. 2008; Wang et al. 2009). El Niño-Southern Oscillation (ENSO) is one of the major interannual climate phenomena controlled by air-sea interactions and can be a useful predictand to evaluate a seasonal prediction system. On the other hand, most CGCMs lack the resolution to reproduce regional rainfall, which is affected by small-scale atmospheric eddies and topography. Furthermore, many CGCMs even have difficulty in reproducing large-scale phenomena such as the monsoon rainfall extending over the Indochina region.

Statistical downscaling is one approach that may help overcome the drawbacks inherent in predictions made with CGCMs. Its common strategy is to establish an empirical statistical relationship between regional-scale climate and large-scale variables. The CGCM output is used as a predictor of a large-scale field and is converted to a regional-scale predictand by a function describing a statistical relationship. Several statistical downscaling schemes are based on regression analysis or similar linear methods (Wetterhall et al. 2005; Feddersen and Andersen 2005; Kang et al. 2007). However, a single statistical function may be not sufficient, because it is likely to be built upon a mixture of several phenomena and may not necessarily be based on the physical mechanism of the individual phenomenon. From this perspective, Chu et al. (2008) derived an empirical statistical relationship between local precipitation over northern Taiwan and large-scale atmospheric fields from the Seasonal Prediction Model Intercomparison Project (SMIP)-type hindcasts (Kang et al. 2004)by applying singular value decomposition analysis (SVDA). They used the leading six modes as the statistical functions and successfully reconstructed the regional rainfall of the target month from the predictor of the GCM hindcasts. The function based on SVDA is most effective where the target predictand is influenced by multiple large-scale phenomena and has independent statistical relationships with them. Therefore, this method would be well suited to our target, Thailand, and the surrounding Indochina region, as its climate is affected by multiple phenomena from both the Pacific and the Indian Oceans with distinct underlying physics.

In terms of Thailand rainfall, Kang et al. (2007) downscaled the SMIP-type hindcast outputs of large-scale atmospheric variables (geopotential height at 500 hPa and sea level pressure) and assessed the predictability of regional rainfall for a one-month lead hindcast from 1983 to 2003. In their work, however, only the leading mode of SVDA was adopted as a downscaling function and seasonality was not considered.

Many hydrologists, meteorologists, social scientists, and policymakers around the world are focusing on the Thailand disaster of 2011 and have discussed its climatic and anthropogenic causes, damage assessment, predictability, and possible risk management for the future. This study is the first assessment of the predictability of the climatic aspects of this disaster. Exploiting the stateof-the-art seasonal ensemble hindcast generated from our CGCM, we investigate the predictability of anomalously heavy Thailand rainfall during the mature rainy season of 2011. We set the target predictand to the accumulated rainfall anomaly from August to October, the latter half of the monsoon season, because the total rainfall through this season is a key factor for water resource management in Thailand (Komori et al. 2012). If authorities had advance notice of the tendency of the continuous seasonal rainfall at the beginning of August 2011, the flood damage could have been reduced. We based the statistical function on SVDA following the case of Taiwan in Chu et al. (2008), but chose the predictor after taking full account of the climatic properties of Thailand and the skill of the CGCM forecast. The novelty of our forecasting system is that it is based on a fully coupled model including an ocean assimilation system (one-tier forecasting system), although most SMIP-type systems used in Chu et al. (2008) and Kang et al. (2007) are two-tier forecast systems, where sea surface temperature (SST) is given as the boundary forcing. The one-tier coupled forecast allows local air-sea interactions, which is not the case with the two-tier system. Thus, we could choose SST as a predictor, while Chu et al. (2008) and Kang et al. (2007) used atmospheric variables instead.

In the following section, we describe the observational data used in this paper, our seasonal prediction system, how the predictor was chosen, and the downscaling method. Statistical relationships isolated by SVDA are analyzed in section 3a and the prediction results of Thailand rainfall are examined in section 3b. Discussions about factors controlling the predictability are presented in section 4. The final section contains a summary and conclusions.

2. Data and methodology

The entire statistical downscaling process in this paper is summarized in Fig. 1. The details are as described in sections 2a-2c.

a. Observational data and CGCM hindcast products

The observed regional precipitation dataset used in this study is the Asian Precipitation-Highly Resolved Observational Data Integration Toward Evaluation (APHRODITE) version APHRO-V1003R1 (Yatagai et al. 2009), which has 0.258 horizontal resolution and covers the period from 1964 to 2007. APHRODITE is used both for identifying the statistical relationship and for validation. For validation after 2008, we adapted the Tropical Rainfall Measuring Mission (TRMM) 3B43 product with spatial resolution of 0.258, which is provided as a quasi-real-time product from 1998 (Huffman et al. 1995). For validation of an oceanic field, SST from the ProjD reanalysis dataset (Ishii and Kimoto 2009) was used. ERA-Interim (Dee et al. 2011) by the European Centre for Medium-Range Weather Forecasts (ECMWF) and the National Oceanic and Atmospheric Administration (NOAA) interpolated outgoing longwave radiation [OLR; Liebmann and Smith (2006)] were also used to verify atmospheric fields.

The CGCM hindcast products were taken from outputs of the seasonal prediction system built on the Model for Interdisciplinary Research on Climate, version 5 (MIROC; Watanabe et al. 2010). The model consists of the atmosphere, land, river, sea ice, and ocean components with no flux adjustments. The resolution of the atmospheric component is triangular spectral truncation at total horizontal wavenumber 85 (T85) with 40 vertical layers. The oceanic component has a horizontal resolution of 1.48 in longitude and 0.98 in latitude (0.58 near the equator), and 44 vertical levels spanning 5 m in the top level and 250 m in the bottom level. In the initialization process, the ProjD reanalysis dataset (Ishii and Kimoto 2009) was assimilated from 1950 under the twentieth century and RCP4.5 climate forcing of solar radiation, volcanic forcing, greenhouse gases, ozone, aerosol, and land-use change (Tatebe et al. 2012). The hindcast experiments were performed in accordance with the Climate-system Historical Forecast Project (CHFP) protocol. Initial states were taken from the above ocean data assimilation system, while atmospheric variables were initiated from the National Centers for Environmental Prediction (NCEP) reanalysis in February, May, August, and November of every year from 1979 to the present. As of this writing, an eight-member ensemble is available, which is formed by the lagged average forecast (LAF) method at 12-h intervals. Each ensemble member was run for 12 months. Imada et al. (2015, manuscript submitted to Mon. Wea. Rev.) describe the specific performance of this forecasting system for predicting extensive large-scale climate phenomena. In this study, we used August-September- October (ASO)-averaged hindcast products started from August of each year (three-month lead predictions).

b. Choice of predictor

To perform the best downscaled prediction, an optimal predictor should be chosen from outputs of the CGCM hindcast simulations. There are two requirements in choosing a predictor. First, it has to be well simulated by the GCM. Second, there should be a stable relationship between the predictor and the predictand (Wilby et al. 1999; Kang et al. 2007; Chu et al. 2008).

To meet the first requirement, we compared the raw hindcast outputs of precipitation, sea level pressure (SLP), geopotential height at 500 hPa (Z500), and SST with observations. Figure 2 shows the horizontal distributions of ASO time correlation coefficients (TCC) between the raw hindcast outputs and the corresponding observations on each grid. In the TCC for precipitation, significant predictability exists only in the central tropical Pacific Ocean (Fig. 2a), which indicates the difficulty the CGCM has in representing observed rainfall patterns. This emphasizes the necessity for an appropriate downscaling method. The TCC for SLP and Z500 show significant predictability in low latitudes (Figs. 2c,d). However, atmospheric variables score less than the SST. That is because seasonal predictability arises from longer oceanic memories, whereas atmospheric fields act as noise in the prediction. Therefore, in this study, we chose SST as the predictor. We targeted the domain between 308Sand 308N in the Pacific and Indian Oceans.

It is well known that the weather systems over Indochina (Thailand inclusive) usually originate from the Pacific and Indian Oceans. To meet the second requirement for choice of a predictor, we examined the relationship between our selected predictor (tropical SST) and the Indochina rainfall focusing on three major interannual variability: the canonical ENSO, Indian Ocean dipole (IOD), and the recently identified type of El Niño referred to as El Niño Modoki (Ashok et al. 2007), date line El Niño (Larkin and Harrison 2005), and central Pacific (CP) El Niño (Kao and Yu 2009), which has warm signals in the central Pacific warm pool region. To investigate the potential impacts of each phenomenon, the Niño-3 index, dipole mode index [DMI; Saji et al. (1999)], and El Niño Modoki index [EMI; Ashok et al. (2007)] are used in this study, respectively. The Niño-3 index is defined by the SST anomaly (SSTA) averaged over 58S-58N, 908-1508W. DMI is defined as the SSTA difference between the western (108S-108N, 608-808E) and eastern (108S-equator, 908-1108E) Indian Ocean. EMI is defined as SSTA (108S-108N, 1658E- 1408W) minus 0.5 SSTA (158S-58N, 1108-708W) minus 0.5 SSTA (108S-208N, 1258-1458E) (Ashok et al. 2007).

Figures 3a, 3c, and 3e show the observed regressions of ASO SSTA with unnormalized ASO Niño-3, EMI, and DMI, respectively. Here, as EMI and DMI are influenced in part by the Niño-3 index, the ENSO effect is first removed using the linear regression with respect to the Niño-3 index and then the regression is obtained. The canonical El Niño features a warming in the eastern equatorial Pacific(Fig. 3a), whereas the CP El Niño is characterized by a warming over the central Pacific Ocean with cold SSTA to the west and east of the warming center (Fig. 3c). The positive IOD event is characterized by cooler-than-normal water in the tropical eastern Indian Ocean, and warmer-than-normal water in the tropical western Indian Ocean (Fig. 3e). Figures 3b, 3d, and 3f show corresponding regressions of rainfall and surface wind, respectively. As a response to the cold SSTA in the western North Pacific (WNP) during El Niño, an anticyclonic circulation appears over the WNP region [Fig. 3b; Wang et al. (2000)]. The northern segment of this anticyclonic response weakens the northeasterly monsoon in early autumn over Indochina, and brings less-than-normal orographic rainfall along the eastern coast of Vietnam (Yen et al. 2011). The anticyclonic anomaly also brings drier condition in the southern Indochina Peninsula.

On the other hand, the CP El Niño is not associated with the anticyclonic response in WNP (Fig. 3d). Instead, an anomalous cyclone exists around the Philippines as a response to the SST warming in the central Pacific (Zhang et al. 2011). This circulation pattern brings a trough in northeastern Indochina and a ridge in southwestern Indochina and determines anomalous precipitation patterns. Previous studies have also reported similar differences in the teleconnection patterns between the two types of El Niño (Weng et al. 2007, 2009; Yuan and Yang 2012; Zhang et al. 2011; Yang and Jiang 2014). During a positive IOD event (Fig. 3f), convection over the Bay of Bengal is activated because of the weakened Hadley cell associated with the colder SST in the eastern equatorial Indian Ocean (Saji and Yamagata 2003). This stronger convection brings more rainfall over the northern inland areas of the Indochina Peninsula.

To evaluate the three major phenomena simulated by our CGCM, we show in Fig. 4 the results from the regression analysis of the ASO hindcasts. The SST anomalies of the canonical ENSO and CP ENSO projected by the CGCM (Figs. 4a,c) extend too far west compared to the patterns extracted from the observation (Figs. 3a,c). This is a systematic bias of MIROC5 originated with too strong SST-wind coupling. Despite the existence of such a systematic bias, the simulated phase variations of the Niño-3 index and EMI are in good agreement with the observed variations. A temporal correlation coefficient is 0.83 for the Niño-3 index and 0.70 for EMI. On the other hand, the simulated IOD is not in agreement with the observation. Although the simulated pattern is similar to the observed pattern (Fig. 4e), our model underestimates the amplitude of IOD and the variation is not in good agreement with the observation (Fig. 4f). Generally, IOD prediction is more difficult than ENSO prediction. Differences of predictability among the phenomena potentially influence results of this study.

The concept of predictor choice mentioned above has also been emphasized in previous studies (e.g., Kang et al. 2007; Chu et al. 2008). However, the prediction skill of the selected variables in GCMs for each year and month has never been fully discussed. Figure 2 clearly shows that atmospheric variables, which were used as predictors in previous studies, do not necessarily have sufficient predictability over the western North Pacific, and that oceanic variables are sometimes more appropriate as a predictor.

c. Statistical downscaling method

We expanded the method of Chu et al. (2008) to downscale the regional rainfall in Indochina. We used SVDA to obtain statistical relationships between large-scale tropical SST and local rainfall in Indochina. It gives the coupled patterns and time series as follows:

..., (1)

..., (2)

where Pi (x) and Qi(y) indicate the ith mode of the singular vector of the large-scale SST from an CGCM hindcast (predictor) and of regional rainfall (predictand), respectively. The terms Ai (t) and Bi(t) denote time expansion coefficients of the ith mode for the predictor and the predictand, respectively, and m is the total number of SVD modes. Given ttarget as the target prediction month, SST(ttarget, x) can be obtained from outputs of a CGCM seasonal prediction. The future time expansion coefficient for SST is calculated as

... (3)

Because the time coefficient for target precipitation, PRCP(ttarget, y), is unknown, Ai (ttarget) is used as a substitute. Then the downscaling transform function can be given as

..., (4)

where N is the total number of SVD modes.

Here, we carried out cross validation to identify time stable modes. In cross validation, the target month for the prediction is excluded from the training period of the statistical analysis of SVDA to prevent the signals of forecast months from being included in the statistical function. This procedure is repeated 33 times, and yields rainfall predictions for 33 years. During the cross validation, each relationship between the rainfall and SST in the transfer functions should be maintained. On the basis of this criterion, the leading six modes are identified as the time stable modes (not shown) and retained in our downscaling scheme.

3. Results

a. Statistical relationship from SVDA

Here we analyze the statistical relationship derived by SVDA [Eqs. (1) and (2)]. We examine spatial patterns of SST and regional precipitation through values regressed onto expansion coefficients Ai (t) and Bi (t), instead of singular vectors Pi (x) and Qi (y), because a singular vector is normalized and thus unitless.

The leading statistical relationships extracted by the SVDA between predicted SST and observed Indochina rainfall are shown in Figs. 5a-c. Although an ENSO-like SST anomaly pattern is visible in the regressed SST field (Fig. 5b), the time coefficients of the leading mode indicate decadal-scale variations rather than interannual variability, and show a continuous La Niña-like phase over the past two decades. In addition, SST anomalies over the entire Indian Ocean show an opposite sign to the ENSO-like anomalies in the eastern tropical Pacific. These notable signals include a La Niña-like tendency and warmer conditions in the Indian Ocean over the past few decades, which is consistent with the hiatus hypothesis discussed in Luo et al. (2012). Therefore, the first SVD mode seems to mainly reflect the recent decadal tendency. In the same manner, Figs. 5d-f show the results of the second mode. This mode shows a typical ENSO-like pattern in the tropical Pacific with dominant interannual variability. The difference from the first mode is found in the Indian Ocean; the IOD pattern is visible. It is also observed that some IOD events have occurred with ENSO events. Impacts of both phenomena are visible in the Indochina regional rainfall. The results of the third mode shown in Figs. 5g-i reflect the features of the CP El Niño as shown in Figs. 3c and 3d.

As seen above, the statistical relationship between the Indochina regional rainfall and the remote oceans depends on which phenomenon develops, and the SVD analysis can cover each relationship according to each phenomenon. Here, the contribution ratio of the leading mode of SVDA is small at less than 15% in both the observed and the simulated fields; nevertheless, typical linear analyses (e.g., an empirical orthogonal function) show that ENSO is a distinctly dominant phenomenon among the tropical upper oceans. This discrepancy arises because the local rainfall in Indochina is influenced by not only ENSO but also by other modes of climate variability, such as IOD and CP El Niño. It means that different phenomena have different physical relationships. That is why we adopt not only the leading mode but also all significant modes.

Note that the correlation coefficient between the time expansion coefficients of SST and local precipitation is 0.80 in Fig. 5a, 0.75 in Fig. 5d, and 0.72 in Fig. 5g. This is an unavoidable bottleneck for the approach of this study, because we assume that these two time series are identical in the process represented in Eq. (4). With this in mind, the skill of our downscaling approach based on SVDA is examined in the next section.

b. Prediction of Thailand rainfall in the mature rainy season in 2011

The results of downscaling obtained from Eqs. (3) and (4) are compared with observations over the latter half of the 2011 monsoon season. Figure 6c shows the precipitation anomaly obtained from the TRMM/3B43 data. More-than-normal rainfall is found not only around Thailand but also over most of Indochina except for some coastal areas. A similar tendency is also predicted in the product downscaled from the CGCM hindcast (Fig. 6d), although some extreme peaks are underestimated in the prediction because the atmospheric noise components are cut in the statistical approach. We can confirm the efficacy of the downscaling method used in this study when the results are compared with the predicted rainfall anomalies directly output from the CGCM (Fig. 6e), in which intensity is underestimated and drier conditions prevail over the southwestern peninsula.

In the background mean state (Fig. 6a), the decaying La Niña signal dominates in the PacificOceanwith peak anomalies in the vicinity of the equator and in the eastern subtropics. As a measure of the atmospheric anomalies, contours of OLR anomalies are also shown in Fig. 6a. Convection was active and brings much rainfall over Indochina. In the Indian Ocean, the IOD structure is observed with warmer SST to the west and colder SST to the east. The eastern cooling around Indonesia enhances a local descending flow, and induces extra convection over Indochina through the change in the Hadley cell (Saji and Yamagata 2003). Furthermore, intraseasonal disturbances, that are generated in the western tropical Pacific and hit the Indochina region, increase during the La Niña phase due to the persistent warm SST anomalies in the warm pool and the enhanced Walker circulation (not shown). These underlying SST conditions that bring more precipitation over Indochina are well captured in the SST hindcast by MIROC except that cold anomalies along the equator in the central to eastern Pacific are underestimated (Fig. 6b). On the other hand, the model hindcast fails to predict the negative OLR (active convection) over Indochina, which is why it is difficult to use atmospheric variables as a predictor.

So far, the results shown in Fig. 6 suggest that our downscaling approach with the CGCM prediction has the ability to predict Thailand rainfall in the 2011 rainy season. Needless to say, however, total predictability over a longer period is necessary for a reliable prediction. Figure 7a shows downscaled rainfall anomalies after 1999 averaged around the Chao Phraya basin (138- 178N, 1008-1018E) relative to the observations. A correlation coefficient between the downscaled hindcasts and TRMM/3B43 is 0.74, which exceeds the 95% significance level of the t test, indicating that our downscaling method is applicable to recent rainfall events in Thailand. To see the applicability to rainfall in other areas, the map of TCC at each grid point between the TRMM/3B43 dataset and the downscaled prediction for recent years (1999-2011) is shown in Fig. 7b.A positive value implies good predictive skill, with the same sign of anomalies between observation and prediction. In Fig. 7b, significantly high predictive skills are found not only over central Thailand but also in the Mekong Delta, and in a few parts of Laos.

4. Discussion

Figure 7b exposes that this downscaling prediction is not always possible anytime and anywhere. It is natural to think that there are both years when prediction is easy and years when prediction is difficult. As seen in Fig. 4, the skill of seasonal prediction differs depending on the dominating phenomenon such as ENSO, IOD, the new type of El Niño, and so forth (Luo et al. 2005; Kim et al. 2009; Hendon et al. 2009). Furthermore, even for the same ENSO, a developing phase shows higher predictability than a decaying phase (Jin et al. 2008). All of these factors are combined in the resulting skill shown in Fig. 7. Therefore, we checked another particular rainfall event in recent years. Figure 8 shows anomalies in ASO 2002 and ASO 2008. During the monsoon season of 2002, a CP El Niño was developing in the Pacific Ocean (Fig. 8a). The associated atmospheric circulation brought wetter-than-normal conditions in northern Indochina, and drier-than-normal conditions in southwestern Indochina (Fig. 8b), which is in good agreement with the relationship shown in Figs. 3c and 3d. This event caused intermissive floods in the Mekong area and hundreds of casualties in Thailand and Vietnam from August to October. The intense rainfall anomalies in these areas are also predicted in the product downscaled from the CGCM hindcast (Fig. 8c).

The background state of 2008 was quite similar to the condition of 2011: the La Niña tendency in the Pacific Ocean and the positive IOD tendency in the Indian Ocean (Fig. 8d). Intense rainfall hit northern Vietnam (Fig. 8e) and caused serious damages and casualties. The heavy rainfall along the east coast of Indochina is well predicted in the downscaled product (Fig. 8f). Overall, the downscaled prediction tends to be successful when robust climate variability exists in the tropical Pacific Ocean.

There are two potential causes for skill loss: one is prediction errors in CGCM hindcast and the other is the imperfect statistical relationship extracted by the SVD analysis. Although the time-mean skill has a critical upper limit due to the unavoidable imperfect statistical relationship, reliability of the relationship depends on SVD modes. The significance of the statistical relationship in each mode can be known in advance by analyzing the degree of coincidence between two expansion coefficients derived from SVDA (e.g., black and red lines in Figs. 5a,d,g). Meanwhile, we can predict in advance which mode is likely to develop in a coming season by referring to Ai (ttarget) in Eq. (4), which gives a future coefficient of a target predictand. Combining these two estimations, the reliability of each forecast can also be predicted. With the addition of such reliability information, forecasted products would surely become useful.

5. Summary and conclusions

This paper describes a methodology to forecast the prolonged and anomalously heavy Thailand rainfall in the mature rainy season in 2011, which is notorious for the disasters caused by a succession of severe floods near Bangkok. It is socially important for the purpose of water resource management to know in advance the tendency of the continuous seasonal rainfall at the beginning of August. Our seasonal prediction system is based on a fully coupled CGCM MIROC (one-tier forecasting system) and shows significant predictive skill in the tropical SST field for up to several months. We introduced downscaling method with SVDA to transform SST hindcast into the Thailand regional rainfall from August to October. The SVDA approach separates the climatic fields into several dominant phenomena and provides each phenomenon-based statistical relationship between the observed station-based rainfall and the large-scale variable hindcasted by the CGCM. Downscaled Indochina rainfall products in the last decades showed consistently significant predictability both in the central Thailand and Mekong Delta areas and some inland regions. Our downscaling approach successfully forecasted the anomalously heavy seasonal rainfall in 2011 in the Thailand area.

These results tell us that regional predictability fluctuates depending on the dominating phenomenon in the background large-scale field, which might be easy or difficult to predict. Therefore, if we have advance knowledge of the tendency of the large-scale circulation, we can estimate whether the following forecast would be reliable. In this regard, the SVDA approach is a powerful tool. Each mode revealed by SVDA has the potential to identify each physical mechanism controlling the relationship between a large-scale and regional-scale phenomenon, and thus helps us to understand the factor that controls the predictability. The modes derived from our SVDA analysis captured each growing or decaying phase of ENSO, IOD, and the new type of CP El Niño. Mode-by-mode analysis will tell us not only the physical processes behind but also the strengths and weaknesses of a CGCM simulation, and how to reduce biases in each mode.

This paper highlights the potential of the one-tier seasonal forecasting system to predict cases like the Thailand rainfall of 2011 and contributes to flood risk assessment. Although the predictor used in this paper was based on a single CGCM hindcast, some precedent studies of seasonal prediction incorporated multimodel one-tier hindcast outputs and showed that the multimodel analysis improves seasonal prediction skill (e.g., Jin et al. 2008). The quasi-real-time seasonal prediction products based on the one-tier coupled systems from institutions around the world will become accessible in the near future. Our next step is to evaluate the approach of this study based on the multimodel outputs.

Moreover, monsoon rainfall in the late rainy season of Southeast Asia could be affected by land surface conditions (e.g., Kanae et al. 2001). Predictability analysis focusing on land surface conditions (e.g., Kanae et al. 2006) is also an important objective and will be one of our next targets.

Acknowledgments. The authors are grateful to J. Luo and anonymous reviewers for their encouraging comments. This work was supported by the Program for Generation of Climate Change Risk Information (SOUSEI project) of the Japanese Ministry of Education, Culture, Sports, Science and Technology, Japan Science and Technology Agency (JST), Japan International Cooperation Agency (JICA), Core Research for Evolutionary Science and Technology (CREST), and Data Integration and Analysis System (DIAS).

REFERENCES

Ashok, K., S. K. Behera, S. A. Rao, H. Weng, and T. Yamagata, 2007: El Niño Modoki and its possible teleconnection. J. Geophys. Res., 112, C11007, doi:10.1029/2006JC003798.

Chu, J.-L., H. Kang, C.-Y. Tam, C.-K. Park, and C.-T. Chen, 2008: Seasonal forecast for local precipitation over northern Taiwan using statistical downscaling. J. Geophys. Res., 113, D12118, doi:10.1029/2007JD009424.

Dee, D. P., and Coauthors, 2011: The ERA-Interim reanalysis: Configuration and performance of the data assimilation system. Quart. J. Roy. Meteor. Soc., 137, 553-597, doi:10.1002/qj.828.

Feddersen, H., and U. Andersen, 2005: A method for statistical downscaling of seasonal ensemble predictions. Tellus, 57A, 398-408, doi:10.1111/j.1600-0870.2005.00102.x.

Hendon, H. H., E. Lim, G. Wang, O. Alves, and D. Hudson, 2009: Prospects for predicting two flavors of El Niño. Geophys. Res. Lett., 36, L19713, doi:10.1029/2009GL040100.

Huffman, G. J., R. F. Adler, B. Rudolph, U. Schneider, and P. Keehn, 1995: Global precipitation estimates based on a technique for combining satellite-based estimates, rain gauge analysis, and NWP model precipitation information. J. Climate, 8, 1284-1295, doi:10.1175/1520-0442(1995)008,1284: GPEBOA.2.0.CO;2.

Ishii, M., and M. Kimoto, 2009: Reevaluation of historical ocean heat content variations with time-varying XBT and MBT depth bias corrections. J. Oceanogr., 65, 287-299, doi:10.1007/ s10872-009-0027-7.

Jin, E. K., and Coauthors, 2008: Current status of ENSO prediction skill in coupled ocean-atmosphere models. Climate Dyn., 31, 647-664, doi:10.1007/s00382-008-0397-3.

Kanae, S., T. Oki, and K. Musiake, 2001: Impact of deforestation on regional precipitation over the Indochina Peninsula. J. Hydrometeor., 2, 51-70, doi:10.1175/1525-7541(2001)002,0051: IODORP.2.0.CO;2.

_____, Y. Hirabayashi, T. Yamada, and T. Oki, 2006: Influence of ''realistic'' land surface wetness on predictability of seasonal precipitation in boreal summer. J. Climate, 19, 1450-1460, doi:10.1175/JCLI3686.1.

Kang, H., K.-H. An, C.-K. Park, A. L. S. Solis, and K. Stitthichivapak, 2007: Multimodel output statistical downscaling prediction of precipitation in the Philippines and Thailand. Geophys. Res. Lett., 34, L15710, doi:10.1029/2007GL030730.

Kang, I.-S., J.-Y. Lee, and C.-K. Park, 2004: Potential predictability of summer mean precipitation in a dynamical seasonal prediction system with systematic error correction. J. Climate, 17, 834-844, doi:10.1175/1520-0442(2004)017,0834: PPOSMP.2.0.CO;2.

Kao, H. Y., and J. Y. Yu, 2009: Contrasting eastern Pacific and central Pacific types of ENSO. J. Climate, 22, 615-632, doi:10.1175/2008JCLI2309.1.

Kim, H.-M., P. J. Webster, and J. A. Curry, 2009: Impact of shifting patterns of Pacific Ocean warming on North Atlantic tropical cyclones. Science, 325, 77-80, doi:10.1126/science.1174062.

Komori, D., and Coauthors, 2012: Characteristics of the 2011 Chao Phraya River flood in central Thailand. Hydrol. Res. Lett., 6, 41-46, doi:10.3178/hrl.6.41.

Larkin, N. K., and D. Harrison, 2005: On the definition of El Niño and associated seasonal average U.S. weather anomalies. Geophys. Res. Lett., 32, L13705, doi:10.1029/2005GL022738.

Liebmann, B., and C. A. Smith, 2006: Description of a complete (interpolated) outgoing longwave radiation dataset. Bull. Amer. Meteor. Soc., 77, 1275-1277.

Luo, J.-J., S. Masson, S. Behera, S. Shingu, and T. Yamagata, 2005: Seasonal climate predictability in a coupled OAGCM using a different approach for ensemble forecasts. J. Climate, 18, 4474-4497, doi:10.1175/JCLI3526.1.

_____, W. Sasaki, and Y. Masumoto, 2012: Indian Ocean warming modulates Pacific climate change. Proc. Natl. Acad. Sci. USA, 109, 18701-18706, doi:10.1073/pnas.1210239109.

Peterson,T.C.,P. A. Stott,andS.Herring, 2012: Explaining extreme events of 2011 from a climate perspective. Bull. Amer. Meteor. Soc., 93, 1041-1067, doi:10.1175/BAMS-D-12-00021.1.

Saji, N. H., and T. Yamagata, 2003: Possible impacts of Indian Ocean dipole mode events on global climate. Climate Res., 25, 151-169, doi:10.3354/cr025151.

_____, B. N. Goswami, P. N. Vinayachandran, and T. Yamagata, 1999: A dipole mode in the tropical Indian Ocean. Nature, 401, 360-363.

Tatebe, H., and Coauthors, 2012: Initialization of the MIROC climate models with hydographic data assimilation for decadal prediction. J. Meteor. Soc. Japan, 90A, 275-294, doi:10.2151/ jmsj.2012-A14.

Wang, B., R. Wu, and X. Fu, 2000: Pacific-East Asian teleconnection: How does ENSO affect East Asian climate? J. Climate, 13, 1517-1536, doi:10.1175/1520-0442(2000)013,1517: PEATHD.2.0.CO;2.

_____, and Coauthors, 2009: Advance and prospectus of seasonal prediction: Assessment of the APCC/CLiPAS 14-model ensemble retrospective seasonal prediction (1980-2004). Climate Dyn., 33, 93-117, doi:10.1007/s00382-008-0460-0.

Watanabe, M., and Coauthors, 2010: Improved climate simulation by MIROC5: Mean states, variability, and climate sensitivity. J. Climate, 23, 6312-6335, doi:10.1175/2010JCLI3679.1.

Weng, H., K. Ashok, S. K. Behera, S. A. Rao, and T. Yamagata, 2007: Impacts of recent El Niño Modoki on dry/wet conditions in the Pacific rim duringboreal summer. Climate Dyn., 29, 113- 129, doi:10.1007/s00382-007-0234-0.

_____, S. K. Behera, and T. Yamagata, 2009: Anomalous winter climate conditions in the Pacific rim during recent El Niño Modoki and El Niño events. Climate Dyn., 32, 663-674, doi:10.1007/s00382-008-0394-6.

Wetterhall, F., S. Halldin, and C.-Y. Xu, 2005: Statistical precipitation downscaling in central Sweden with the analogue method. J. Hydrol., 306, 174-190, doi:10.1016/ j.jhydrol.2004.09.008.

Wilby, R. L., L. E. Hay, and G. H. Leavesley, 1999: A comparison of downscaled and raw GCM output: Implications for climate change scenarios in the San Juan River basin, Colorado. J. Hydrol., 225, 67-91, doi:10.1016/ S0022-1694(99)00136-5.

Yang, S., and X. Jiang, 2014: Prediction of eastern and central Pacific ENSO events and their impacts on East Asian climate by the NCEP climate forecast system. J. Climate, 27, 4451- 4472, doi:10.1175/JCLI-D-13-00471.1.

Yatagai, A., O. Arakawa, K. Kamiguchi, H. Kawamoto, M. I. Nodzu, and A. Hamada, 2009: A 44-year daily gridded precipitation dataset for Asia based on a dense network of rain gauges. SOLA, 5, 137-140, doi:10.2151/sola.2009-035.

Yen,M.-C.,T.-C. Chen,H.-L. Hu,R.-Y. Tzeng,D. T.Dinh,T. T.T. Nguyen,andC.J. Wong,2011:Interannualvariation ofthefall rainfall in central Vietnam. J. Meteor. Soc. Japan, 89A, 259- 270, doi:10.2151/jmsj.2011-A16.

Yuan, Y., and S. Yang, 2012: Impacts of different types of El Niño on East Asian climate: Focus on ENSO cycles. J. Climate, 25, 7702-7722, doi:10.1175/JCLI-D-11-00576.1.

Zhang, W., F.-F. Jin, J. Li, and H.-L. Ren, 2011: Constrasting impacts of two-type El Niño over the western North Pacific during boreal autumn. J. Meteor. Soc. Japan, 89, 563-569, doi:10.2151/jmsj.2011-510.

YUKIKO IMADA

Meteorological Research Institute, Japan Meteorological Agency, Tsukuba, Ibaraki, Japan

SHINJIRO KANAE

Tokyo Institute of Technology, Meguro, Tokyo, Japan

MASAHIDE KIMOTO AND MASAHIRO WATANABE

Atmosphere and Ocean Research Institute, The University of Tokyo, Kashiwa, Chiba, Japan

MASAYOSHI ISHII

Meteorological Research Institute, Japan Meteorological Agency, Tsukuba, Ibaraki, Japan

(Manuscript received 11 July 2014, in final form 9 December 2014)

Corresponding author address: Yukiko Imada, Meteorological Research Institute, Japan Meteorological Agency, 1-1, Nagamine, Tsukuba, Ibaraki, 305-0052, Japan.

E-mail: [email protected]