Analysis of changes in extreme climate events is important due to the potentially high social, economic and ecological impact of such events. In the past, the limited availability of long records of daily climate data in some parts of the world hampered efforts to analyse the impacts of climate change and variability on the frequency and severity of climate extremes around the globe (Folland et al., 2001). Since that time, international collaboration has significantly improved the situation, culminating in an analysis (Alexander et al., 2006) that provided a near-global perspective on changing climate extremes for the Intergovernmental Panel on Climate Change (IPCC) Fourth Assessment Report (IPCC, 2007). This global assessment greatly benefited from the contributions from a series of workshops (Peterson and Manton, 2008) coordinated by the Expert Team on Climate Change Detection and Indices (ETCCDI) which is jointly sponsored by the World Meteorological Organization (WMO) Commission for Climatology, the Joint Commission for Oceanography and Marine Meteorology (JCOMM) and the Research Programme on Climate Variability and Predictability (CLIVAR). The ETCCDI workshops seek to bring participants together from countries within a data sparse region to fill in data gaps and to provide capacity building.
The availability of daily observations is steadily improving and has led to the development of gridded regional (Haylock et al., 2008) and global datasets (Caesar et al., 2006). For some areas of the world, daily data availability is still relatively poor and therefore workshops continue to be organised to aid with the collection of climate extremes indices in data sparse regions. This paper describes the outcomes of a workshop held in Hanoi, Vietnam, attended by representatives from meteorological services around the Indo-Pacific region from 3rd to 7th December 2007 with the objective of improving and updating the assessment of climate extremes over the broad Southeast Asian region. The workshop followed an established pattern developed from previous workshops (Peterson and Manton, 2008) and representatives from 12 countries (Australia, Bhutan, Cambodia, Fiji, Laos, Maldives, Myanmar, Nepal, Sri Lanka, Thailand, Timor Leste and Vietnam) attended the event and were asked to bring a selection of their best quality digitised daily temperature and precipitation series. The workshop included a number of countries which have not previously been included in global or regional analyses of climate extremes such as a previous Southeast Asia workshop described in Manton et al. (2001). Some of these countries have short and/or fragmented climate records, often as a result of armed conflict at various times in the last 50 years. In a few cases, the available records are currently too short to be usable for the adequate calculation of climate trends, but these records are still valuable in providing a baseline for future analyses, as well as monitoring inter-annual climate variability. Freely available software was provided to enable participants to carry out quality control and calculate indices (Peterson et al., 2002) and also to test the homogeneity of their series. As with the preceding workshop in Brazzaville, Congo (Aguilar et al., 2009), this climate extremes indices workshop followed on from a 3-week climate data management system training programme. Data analyses continued after the workshop stage as additional stations were made available, and this allowed more time to properly assess the quality control, homogeneity checks and trend analyses.
In addition to identifying trends in extreme events, it is also useful to begin to investigate potential causal mechanisms underlying any observed trends or variability. It is well known that the effects of the El Niño-Southern Oscillation (ENSO) are strongly felt in the climate variability of countries surrounding the Pacific basin (McBride and Nicholls, 1983; Ropelewski and Halpert, 1987; Halpert and Ropelewski, 1992; Nicholls et al., 2005). Kenyon and Hegerl (2008) found that temperature extremes globally are influenced by large-scale circulation patterns, such as ENSO, and these effects are seen most clearly around the Pacific Rim and in North America. We therefore also investigated the association between variations in the observed extremes indices and in SST patterns in the Indo-Pacific region.
Other large-scale features can influence the occurrence of climate extremes over parts of this region. During late 2006 and early 2007, the worst flooding in a century over southern Peninsular Malaysia was caused by heavy precipitation associated with strong northeasterly winds over the South China Sea (Tangang et al., 2008). Other important factors in this event were the absence of the synoptic Borneo vortex and the influence of the Madden–Julian Oscillation (MJO), which interact to influence the variability of deep convection across the region (Chang et al., 2005). Although we do not explicitly consider these mechanisms in this study, there are clear links to SST variability and ENSO (Salahuddin and Curtis, 2009).
2. Data and methods
2.1. Station data and quality control
The workshop participants brought daily time series of temperature (maximum and minimum) and precipitation for a total of 67 stations from the 12 countries, with some of the data being supplied following the workshop (Table I). The stations covered a substantial part of southern and south-eastern Asia, along with Fiji, Timor Leste and northern Australia. An additional station was subsequently contributed by Brunei Darussalam, which was not represented at the workshop.
Table I. List of stations
Victoria River Downs
Prachuap Khiri Khan
Buon me Thuot
Data were processed using freely available software packages: RClimDex, which performs data quality control and calculates indices, and RHtest, which performs homogeneity tests. These packages can be downloaded from the ETCCDI website (http://cccma.seos.uvic.ca/ETCCDI/). We made use of an updated version of RClimDex that takes account of the precision of the input data (Zhang et al., 2009).
The station data underwent homogeneity testing at the workshop using the RHtest software package, which can help to identify step changes in a time series by comparing the goodness of fit of a two-phase regression model with that of a linear trend for the entire series (Wang, 2003, 2008a, 2008b). RHtest is used to help identify series break points for further investigation by the workshop participants. We do not attempt to adjust the data as a result of the homogeneity tests and take a conservative approach of excluding sections of data that appear to be potentially suspect. It should be noted that well-established statistical methods for testing the homogeneity of daily resolution series are lacking (Wijngaard et al., 2003), and there are a range of alternative tests which could be used to assess data homogeneity (Wijngaard et al., 2003; Della-Marta and Wanner, 2006; Reeves et al., 2007). A review of methods used for the statistical detection of inhomogeneities (Reeves et al., 2007) found that two-phase regression methods, as implemented in the RHtest software, had a comparable level of performance to methods such as the standard normal homogeneity test (Alexandersson, 1986), with the optimal choice depending on the priorities of the user (e.g. accurately detecting the date of a changepoint or minimising the number of false alarms). Reeves et al. also found that developments in the two-phase regression method since 1995, reflected in recent versions of RHtest, had substantially improved its performance.
It is generally recommended to apply homogeneity tests relatively, that is, testing a station with respect to its nearest neighbours (Peterson et al., 1998), but this is often not possible due to the sparseness of the station network. In this study, we considered stations individually without using a reference series, as suitable reference series were unavailable in some of the countries; this has the effect of limiting the detection of smaller inhomogeneities. Many stations displayed potential break points around the time of the 1997–1998 El Niño, or other ENSO events, suggesting that a genuine climatic process was being identified as an inhomogeneity; this is a potential consequence of not using reference series. Comprehensive metadata were also not available for many stations, so we used a combination of the RHtest results and their graphical output to identify serious inhomogeneities that did not coincide with known ENSO events. Based upon this, we excluded some time series, either in whole or in part, from further analysis and no attempt was made to adjust time series to remove inhomogeneities. Following the workshop, we also carried out further homogeneity tests as used by the European Climate Assessment and Dataset (ECAD; Wijngaard et al., 2003). These produced results largely comparable with the results of RHtest, particularly, in terms of identification of years during which particularly large break points occurred.
We selected 56 stations for studying changes during the period 1971–2005. Selection was based on series length and completeness, quality control and homogeneity. Some indices, such as TN90p, require a base period over which to calculate, in this case, the 90th percentile of daily Tmin. Variations in TN90p (and other percentile-based indices) are therefore calculated relative to the defined base period. This is essentially similar to the definition of a 30-year climatological normal. In this study, we used a base period of 1971–2000 where the station data covered that period. In some cases, we used a shorter base period; for example, the data supplied for Bhutan began in 1985, so we used a base period of 1985–2000.
A total of 27 indices, based upon recommendations of the ETCCDI, are calculated using RClimDex. Many of the indices use locally defined thresholds, making it easier to compare results over a wide region. The ETCCDI website lists the complete set of indices and their definitions, and Table II lists the indices presented in this paper.
Table II. List of the ETCCDI indices used in this study
Simple daily intensity index (average rainfall on days with ≥ 1 mm rain)
Heavy precipitation days (10 mm or more)
Very heavy precipitation days (20 mm or more)
Consecutive dry days
Consecutive wet days
Very wet days (above the 95th percentile of days with ≥ 1 mm rain)
Extremely wet days (above the 99th percentile of days with ≥ 1 mm rain)
Annual total wet-day precipitation
2.2. Trend calculation and regional series
Trends for stations and regional series are calculated using the Kendall's slope estimator. This method involves two steps: one is to estimate a linear trend between two observations by dividing the difference of the two values by the time period; the second is to rank the trends computed from all possible data pairs and to take the median values of the trends as the final result (Sen, 1968; Zhang et al., 2000; Wang and Swail, 2001). This method has been used for trend analysis of results from previous ETCCDI workshops (Zhang et al., 2005) because it is more suitable for dealing with outliers and non-normal distributions. The confidence limits of the trend are derived from the Kendall test but with a modification to account for lag-1 autocorrelation in the time series residuals using the technique described in Zhang et al. (2005). This is because the results of the Mann-Kendall test for trend depend strongly on the autocorrelation of the series: a significant trend may be falsely detected due to an autocorrelation in the series (Kulkarni and von Storch, 1995). Trends are significant at the 5% level when results are ± 1.96 standard deviations from the median trend. We require at least 70% of annual data to be non-missing to calculate a trend and refer to trends as being significant if they are determined to be statistically significant at the 5% level. As discussed in the introduction, a number of station records, notably those from Cambodia and Timor Leste, were too short to adequately calculate trends (Table I).
We combined all available station indices to produce a regional mean time series for each index. As the number of stations with indices varies over time (e.g. the stations for Bhutan start in 1985), we have adjusted the mean series to account for changes in the variance that result from the differing number of data points each year, taking account of the correlation between stations (Osborn et al., 1997; Brunet et al., 2007). As the stations cover a fairly expansive geographical area, we also produce four sub-regional time series. The stations are divided into the sub-regions as indicated in Figure 1. These are referred to as Indian Ocean (Maldives and Sri Lanka), Himalayas (Nepal and Bhutan), South China Sea (Vietnam, Cambodia, Laos, Thailand, Myanmar and Brunei) and South Pacific (Fiji, Timor Leste and Australia).
2.3. Sea surface temperature
To investigate the relationship between the indices and SST patterns, we used the Hadley Centre sea ice and sea surface temperature data set (HadISST), which is a combination of monthly globally complete fields of SST and sea ice concentration on a 1° latitude–longitude grid (Rayner et al., 2003). HadISST is available from the Met Office Hadley Centre (http://www.metoffice.gov.uk/hadobs/). We removed a linear trend from the regional indices time series and correlated them with similarly detrended annual SST grid point time series to produce spatial correlation maps.
3.1. Trends in temperature indices
Individual stations show most spatial coherence in the TN90p index, that is, frequency of nights warmer than the 90th percentile (Figure 2). Nearly half of the available stations indicate a significant increase in this index over the 1971–2005 period.
The regional series for the percentile-based temperature indices are shown in Figure 3. The frequencies of warm days and nights, relative to the 1971–2000 base period, increased strongly between 1971 and 2005, with a large increase in the number of nights per year exceeding the 90th percentile threshold. There were also large reductions in the frequency of cold nights and cold days over the 35 years.
As the frequency of warm nights (TN90p) shows the most consistent pattern of trends of all the indices we considered, we also produced time series for the sub-regions identified in Figure 1 (Figure 4). These indicate an increase in all regions, although a relatively small one in the Himalayas. The Indian Ocean region shows the largest increase in TN90p.
Table III lists the regional and sub-regional trends for all the temperature indices examined in this paper and also compares them with the global indices calculated by Alexander et al. (2006). For the percentile-based indices, we see general agreement in the sign and significance of trends in most regions compared with the global results. There are some differences for the Himalayas, where many of the records are quite short, but the trends are non-significant where there is disagreement with the global results.
Table III. Regional and global trends in temperature indices for the period 1971–2005
South China Sea
Global trends were calculated from Alexander et al. (2006) data and referred to the period 1971 to 2003. Trends significant at the 5% level are shown in bold.
Percentage of days in a year per decade
In general over the entire region, the frequency of warm days and warm nights has increased, and the frequency of cold days and cold nights has decreased. This agrees with the results from other studies which have analysed these trends across different parts of the Asia-Pacific region (Griffiths et al., 2005; Klein Tank et al., 2006; Choi et al., 2009). However, the results for the absolute temperature indices (TXx, TNx, etc.) defined for the entire region are sensitive to the large variability in these indices across the region. This can be seen in Table III where the trends for the Indo-Pacific region are generally small and non-significant which is in contrast to some of the sub-regional results. The percentile indices (e.g. TN90p) are more robust across large regions because they account for the influence of local climate effects. There has been a significant increase in the absolute annual maximum of both daily maximum and minimum temperatures, again in common with the global picture. Increases in the temperature of the hottest day of the year have been high in the Himalayas at over 1 °C per decade but also strong over the Indian Ocean (0.78 °C per decade) where the trend is significant.
3.2. Trends in precipitation indices
The station trends map for RX5day (Figure 5) is shown as an example of the relative lack of spatial coherence seen in the precipitation indices. The number of stations with significant trends in either direction is low.
There is an increasing trend in the Indo-Pacific regional R95p index (Figure 6(a)) indicating that the annual amount of precipitation contributed on days exceeding the long-term 95th percentile has increased from about 400 mm to around 450 mm, but this change is non-significant (Table IV). The annual maximum 5-day rainfall (RX5day) index (Figure 6(b)) does not indicate much change over the 35-year period.
Table IV. Regional and global trends in precipitation indices for the period 1971–2005
South China Sea
Note that global trends were calculated from Alexander et al. (2006) data and referred to the period 1971 to 2003. Trends significant at the 5% level are shown in bold.
The sub-regional time series for R95p (Figure 7) emphasise the low spatial coherence of this index. There are increases in R95p over the Himalayas and Indian Ocean, a marginal increase over the South China Sea region, and a decrease over the South Pacific region; of these, only the Himalayan change is statistically significant. A similar pattern is seen in total annual precipitation (not shown).
Table IV lists the regional trends for the precipitation indices and also the global trends. The same problems exist with defining some of the precipitation indices across the whole region that applied to the absolute temperature indices, and indices defined relative to a local climatology (e.g. percentile based) are preferable for comparing across such a large region. Compared with the temperature indices, there are fewer significant trends in the precipitation indices. In contrast to the other sub-regions, the South Pacific region has decreasing trends in all precipitation indices, apart from the consecutive dry day index, suggesting a consistent change towards drier conditions. However, it must be emphasised that these trends are non-significant. Over the region as a whole, the precipitation trends are mixed. This does not parallel the global results of Alexander et al. (2006) indicating consistent trends towards wetter conditions across nearly all of the indices, although it should be noted that analysis was for a different time period (1951–2003) and had only limited coverage of the tropics.
3.3. Relationship with SSTs
The regional and sub-regional series of TN90p (Figures 3(d) and 4) display some annual peaks coincident with El Niño events; in particular, 1998 stands out and, as mentioned in Section 2.1, was identified by the homogeneity testing software as an inhomogeneity for some of the stations. These features are more noticeable for the three maritime regions than they are for the Himalayas.
The correlation pattern between SST and TN90p (Figure 8(a)) shows positive correlations in the South China Sea and Bay of Bengal, suggesting a relationship between the frequency of warm nights and the local SSTs. Correlations with temperatures in the equatorial Pacific are low, though Figure 4 shows peaks in the sub-regional time series of TN90p during some of the more intense ENSO events, in particular, 1997–1998.
The correlation patterns between SST and some indices related to heavy rainfall (R10mm, R20mm) have an ENSO-like pattern (Figure 8(b)). Negative SST anomalies in the equatorial Pacific (i.e. La Niña conditions) are correlated with increased heavy precipitation days. One possible explanation is that La Niña is known to be related to an increase in the formation of tropical cyclones in the western Pacific, and changes in the track of the storms mean that more traverse the South China Sea and make landfall over Vietnam and neighbouring countries (Elsner and Liu, 2003). PRCPTOT (not shown) has a similar pattern of correlation with SST to that of R20mm in Figure 8(b), but precipitation intensity (SDII) does not appear related to ENSO-like SST patterns (not shown).
4. Discussion and conclusions
We have compiled and analysed a set of daily station observations from countries in the Indo-Pacific region to enable an assessment of changes in climate extremes over the region. For some countries in the region, this is the first time that their data have been included in such an analysis.
Our finding that the increasing trends for warm nights were the most spatially coherent index is consistent with the results of other regional workshops (Klein Tank et al., 2006, Choi et al., 2009) and the global analysis of Alexander et al. (2006). We see less spatial coherence in trends in precipitation indices across the region and fewer trends that are locally significant when compared with the temperature indices. In the few cases where statistically significant trends in precipitation indices are identified for regions and sub-regions, there is generally a trend towards wetter conditions in common with the global results of Alexander et al. (2006).
A preliminary investigation of the relationship between the extremes indices and SST indicates that the inter-annual variability of temperature extremes may be related to local SSTs. However, the inter-annual variability of the regional series also seems to indicate that the peaks in the frequency of ‘warm extremes’ may coincide with large El Niño events. The homogeneity tests also indicated that statistical break points in the station series often occurred during the 1997–1998 El Niño period. Some of the precipitation indices, notably total annual precipitation and the number of wet days above 10 and 20 mm, appear to be positively correlated with a La Niña-like SST pattern. Clearly, causes other than large-scale SST variability are likely to underlie the observed changes in extreme events, but our results indicate that further investigation of the ENSO link would be worthwhile.
The Vietnam workshop was supported by the UK Met Office through WMO Voluntary Co-operation Programme funds and coordinated by the WMO World Climate Data and Monitoring Programme (WCDMP). The workshop participants would like to thank the Hydrometeorological Service of Vietnam for their assistance and hospitality. We would also like to acknowledge the support provided by Omar Baddour and Hamma Kontongomde of the WMO WCDMP. We are grateful to Xuebin Zhang, Xiaolan Wang and Feng Yang of the Meteorological Service of Canada for developing the RClimDex and RHtest software and providing continuing support. John Caesar was supported by the Joint DECC, Defra and MoD Integrated Climate Programme—DECC/Defra (GA01101), MoD (CBC/2B/0417_Annex C5).