This article presents an analysis of the spatiotemporal changes (1960–2100) in temperature and precipitation extremes of a typical arid zone (i.e., the Tarim River Basin) in Central Asia. The latest observations in the past five decades (1960–2009) and Coupled General Circulation Model (CGCM) multimodel ensemble projections (2010–2100) using the Bayesian Model Average (BMA) approach are employed for analysis in this study. Results indicate: (1) Most warm (cold) extreme temperature indices have shown significantly positive (negative) trends in the Tarim River Basin in past five decades, while only slight changes in precipitation extremes can be observed. From the spatial perspective, more significantly warm (cold) extremes are found in the desert zones than in upstream mountain zones (i.e., the Tian Shan Mountain and Kunlun Mountain systems which surround the basin). Whereas, there are no identical spatial patterns for the change in extreme precipitation; (2) Ensemble of five CGCM models in Phase 3 of the Coupled Model Intercomparison Project (CMIP3) based on the BMA method suggests that the increasing consecutive dry days (CDD), together with the decreasing frost day (FD) and increasing warm nights frequency (TN90) may lead to more frequent droughts in Tarim in future. Meanwhile, slight increase of annual count of days with precipitation of more than 10 mm (R10), maximum 5-day precipitation total (R5D), simple daily intensity index (SDII), and annual total precipitation with precipitation >95th percentile (R95) in projections indicate a probability of flood occurrence in summer together with frequent occurrence of droughts. The results can provide beneficial reference to water resource and eco-environment management strategies in arid zones for associated policymakers and stakeholders.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
 Substantial progress in both global and regional modeling at medium to high resolution allowed for an increasing number of studies on modeling climate changes. Recent modeling efforts have enabled us to characterize changes in climate extremes with closely relevance to impacts than the traditional climate model outputs of mean temperature and precipitation [e.g., Meehl et al., 2005; Intergovernmental Panel on Climate Change (IPCC), 2007]. In support of the IPCC Fourth Assessment Report [IPCC, 2007], over 20 modeling groups around the world conducted climate change simulation by different Coupled General Circulation Models (CGCM) [IPCC, 2007]. This constitutes Phase 3 of the Coupled Model Intercomparison Project (CMIP3) [Meehl et al., 2007] ensemble of simulations. These models provided extreme indices/indicators for the present and future climates, offering opportunities to conduct the multimodel ensemble analysis of the simulation and projection of climate extremes. For example, Tebaldi et al.  analyzed historical and future simulations of these indicators derived from an ensemble of nine CMIP3-CGCM models under a range of emission scenarios, and found that on global and continental scales, the simulated historical trends generally agree with previous observational studies, providing a sense of reliability for the simulations.
 Meanwhile, the Bayesian model averaging (BMA) approach is introduced to model evaluation and multimodel averaging with a systematic consideration of modern uncertainty in climate impact research [Min et al., 2004, 2005; Min and Hense, 2006], and its application to global mean surface air temperature (SAT) changes is shown from multi Atmosphere-Ocean General Circulation Model (AOGCM) [IPCC, 2007] ensembles of IPCC AR4 simulations. BMA provides a way to combine different models and is a rather promising method for calibrating ensemble in modeling and forecasts. BMA is also a method of combining forecasts from different sources into a consensus probability distribution function (PDF), an ensemble analog to consensus forecasting methods applied to deterministic forecasts from different sources [Krishnamurti et al., 1999]. BMA naturally applies ensemble systems to make a set of discrete models (such as the Canadian ensemble system). In BMA, the overall forecast PDF is a weighted average of individual forecast PDFs. The weights are the estimated posterior model probabilities and reflect the forecast skill of individual models in the training period. The weights can also provide a basis for selecting ensemble members: there is only little loss by removing the ensemble member with small weights [Raftery et al., 2005; Wilson et al., 2007]. This can be a useful strategy, given that the computational cost of running ensembles is more affordable nowadays. Due to pronounced advantages, increasing studies using various BMA methods for climate change detection and attribution [Min et al., 2004, 2005] as well as for future projections of climate changes [Tebaldi et al., 2006] were reported.
 However, even in the available literatures on climate impact research over the arid zones [e.g., Zhang et al., 2005; Klein Tank et al., 2006], in-depth studies regarding changes of climate extremes are still inadequate to understand the unique change of climate extremes in arid zones under the global warming conditions. Most of aforementioned efforts focused only on the detection of variability and trends in climate extremes. To the best of our knowledge, reports in constructing reliable scenarios of future climate extremes in arid zones are very limited so far, motivating our research conducted in this study. Our work strives to offer a comprehensive analysis of changes in temperature and precipitation extremes in a typical arid zone using the latest observations (1960–2009) and CMIP3-CGCM multimodel ensemble projections (2010–2100) through the Bayesian Model Averaging (BMA) approach. This article seeks to: (1) identify spatial and inter-annual changes in temperature and precipitation extremes in the Tarim River Basin (1960–2009) using the latest observations in the past five decades; and (2) construct scenarios of climate extremes using multimodel ensemble projections (2010–2100) in the basin provided by CMIP3 based on the BMA method.
2. Study Region
 The Tarim River Basin in Central Asia, one of the world's foremost endorheic drainage systems and the most densely populated and dominated by an arid inland climate [Chen et al., 2006], is selected in this study to demonstrate the regional response to global climate change in arid zones. The basin is well-known internationally for its unique and world's largest “Pulpous Euphratica” gene library in an extremely arid zone [Ministry of Water Resources (MWR), 1997–2000].
 Situated in the southern Xinjiang Autonomous Region of Northwest China, the Tarim River is 1,321 km long with a drainage area of 557,000 km2 (34°∼43°N, 73°∼93°E, Figure 1). Generally, the drainage system is composed of one mainstream (the Tarim River) and three major tributary rivers (i.e., the Aksu River, Yarkant River, and Khotan River) originated from the Tian Shan Mountain and Kunlun Mountain systems (the northern edge of the Tibetan Plateau). Among these tributary streams, the Aksu River is the most important tributary, accounting for 70–80 percents of total water volume. The basin is the driest region in Eurasia. Its predominant part is occupied by the Taklamakan Desert, whose sand area exceeds 270,000 km2. The basin is a relatively flat desert region with average annual precipitation of less than 50 mm and the total annual runoff of about 280 × 108 m3. The runoff principally comes from high mountain precipitation, and seasonal snow and glacier melting water [MWR, 1997–2000].
 Currently, water resources in the basin have been well developed and heavily committed for a variety of demands in drinking water supply, irrigation, flood control, and suppression of salinity intrusion for the past 30 years [MWR, 1997–2000]. The limited water resources severely conflict with different interests, particularly irrigation and ecological needs. With the negative impacts of global warming and rapidly growing water consumption by human society in the past, the oasis spreading along the downstream has been gradually degrading. These intensified extreme droughts have seriously threatened the sustainable socio-economic developmental efforts and triggered a series of serious issues of desertification in the lower part of Tarim River. Hence, the detection of historical changes and robust projection of future spatiotemporal changes in climate extremes over the basin is beneficial for formulating a sustainable regional water resources management strategy in the arid region.
3. Data and Methods
3.1. Data Sources of Observation
 Daily precipitation, maximum and minimum temperatures during 1960–2009 are provided by the National Climate Center, China Meteorological Administration. The density of distribution and the quality of observational data in China meet the World Meteorological Organization's standards at a total of 20 stations (Table 1) in the data sets. Most stations were established in the 1950s, but any data before 1960 was excluded due to quality reason.
Table 1. List of 20 Meteorological Gauges (1960–2009) in the Tarim River Basina
Altitude (m asl)
Source of data: National Center of Climate, China.
3.2. Definition of Extreme Indices
 A suite of climate change indices on extremes were developed by the joint CCl/CLIVAR/JCOMM Expert Team (ET) on Climate Change Detection and Indices (ETCCDI, http://ccma/seos.uvic.ca/ETCCDMI). In this effort, 27 indices based on daily temperature and precipitation data were defined and software packages were developed for end-users. In this study, 12 temperature and 6 precipitation indices (18 indices in total) are selected, many of which are commonly used to validate climate model simulations [Peterson et al., 2008]. Detailed descriptions are provided in Table 2.
Table 2. List of the 18 Climate Indices for the Tarim River Basina
All the indices are calculated by RClimDEX. Abbreviations are as follows: TX, daily maximum temperature; TN, daily minimum temperature; TG, daily mean temperature; RR, daily precipitation. A wet day is defined when RR ≥ 1 mm, and a dry day is defined when RR < 1 mm.
Annual count when TN(daily minimum) < 0°C
Annual count when TX(daily maximum) > 25°C
Growing season length
Annual count between first span of at least 6 days with TG > 5°C and first span after July 1 of 6 days with TG < 5°C
Annual maximum value of daily maximum temperature
Annual maximum value of daily minimum temperature
Annual minimum value of daily maximum temperature
Annual minimum value of daily minimum temperature
Cold nights frequency
Percentage of nights when TN < 10th percentile
Cold days frequency
Percentage of days when TX < 10th percentile
Warm nights frequency
Percentage of nights when TN > 90th percentile
Warm days frequency
Percentage of days when TX > 90th percentile
Diurnal temperature range
Annual mean difference between TX and TN
Max 1-day precipitation
Annual maximum 1-day precipitation
Max 5-day precipitation
Annual maximum consecutive 5-day precipitation
Simple daily precipitation intensity index
Average precipitation on wet days
Consecutive dry days
Maximum number of consecutive days with RR < 1 mm
Very wet day precipitation
Annual total precipitation when RR > 95th percentile
Annual total precipitation in wet days (PR ≥ 1 mm)
 The 18 indices were used primarily for the assessment of the many aspects of a changing global climate covering changes in intensity, frequency and duration of temperature and precipitation events [Alexander et al., 2006; You et al., 2010]. According to Alexander et al. , the indices are grouped into five different categories: (1) percentile-based indices, such as occurrence of cold nights (TN10), (2) absolute indices represent maximum or minimum values within a season or a year, such as maximum daily maximum temperature (TXx) and maximum 1-day precipitation amount (RX1day), (3) threshold indices defined as the number of days on which a temperature or precipitation value falls above or below a fixed threshold, such as the number of frost days (FD), (4) duration indices which define periods of excessive warmth, cold, wetness or dryness (or in the case of growing season length, periods of mildness), such as consecutive dry days (CDD), and (5) other indices, such as diurnal temperature range (DTR). Most indices have the same name and definition in previous studies [Aguilar et al., 2005; Alexander et al., 2006; Klein Tank et al., 2006; New et al., 2006; You et al., 2010], although their exact definitions may vary slightly.
3.3. Ensemble Projection of Future Climate Extreme
Table 3 lists the six IPCC AR4 global coupled climate models providing temperature and precipitation extremes. Three IPCC AR4 global coupled climate models were used in this study, including GFDL-CM2.0, CNRM-CM3 and MIROC3.2 (medres). These three models were selected because they showed relatively reasonable performances in simulating the temperature and precipitation extremes over Tarim.
Table 3. List of Six IPCC AR4 Global Coupled Climate Models Providing Temperature and Precipitation Extremesa
Correlation Coefficient for Daily Temperature
Correlation Coefficient for Daily Precipitation
Three models set in bold (i.e., GFDL-CM2.0, CNRM-CM3 and MIROC3.2-medres) are recommended and used for their relatively reasonable performances in simulating the temperature and precipitation extremes over Tarim.
Center National Weather Research, CNRM, METEO-FRANCE
Center for Climate System Research (The University of Tokyo), National Institute for Environmental Studies, and Frontier Research Center for Global Change (JAMSTEC)
Geophysical Fluid Dynamics Laboratory, NOAA
The National Center for Atmospheric Research, NCAR
Meteorological Research Institute, Japan Meteorological Agency, Japan
Institut Pierre Simon Laplace
 This set of scenarios spans almost the entire IPCC scenario range, with B1 being close to the low end of the range (CO2 concentration of about 550 ppm by 2100), A2 to the high end of the range (CO2 concentration of about 850 ppm by 2100) and A1B to the middle of the range (CO2 concentration of about 700 ppm by 2100). The models have different horizontal resolutions in the corresponding atmospheric components as shown in Table 3. To obtain the ensemble results of different models, the data were interpolated onto a common 1° × 1°grid. The data set was obtained from the PCMDI web site (www.pcmdi-llnl.gov) and more information about the participating models and data set can be found on this web site.
 Seven key indices (Table 4) suggested by Frich et al.  are chosen in constructing scenarios of future climate extremes in the basin. FD and TN90 represent the extreme temperature and the key indices for extreme precipitation are CDD, R10, R5D, SDII and R95. For the latter indices, higher values indicate more extreme precipitation. The CDD is the length of dry spell whereas R10, R5D, SDII, and R95 stand for the intensity or frequency of precipitation. All indices mentioned in this paper were calculated on an annual basis under three emission scenarios and 5 models ensemble means of the 20th century and the scenario simulations of 21st century, respectively.
Maximum number of consecutive dry days with RR < 1 mm
Maximum 5-day precipitation total
Simple daily intensity index: annual total precipitation divided by the number of wet days (defined as RR > =1.0 mm) in the year
Annual total precipitation when RR > 95th percentile
3.4. Trend Free Pre-whitening (TFPW) Approach
 Results of partial auto-correlation tests indicated the first-order auto-correlation exist in the climate series of some stations in Tarim. Hence, the station-based data were corrected to reduce the effects of serial correlation through the Trend Free Pre-Whitening (TFPW) approach developed by Yue et al. [2002, 2003]. The TFPW involves estimating a monotonic trend for the series, removing this trend prior to pre-whitening the series and finally adding the monotonic trend calculated in the first step to the pre-whitened data series. In essence, the TFPW approach attempts to separate the serial correlation that arises from a (linear) trend from the remaining serial correlation and then only removes the latter portion of the serial correlation. The TFPW-MK procedure of Yue et al.  is applied in the following manner to detect a significant trend in a serially correlated time series:
 1. The slope (β) of a trend in sample data is estimated using the approach proposed by Sen . The original sample data Xt were unitized by dividing each of their values with the sample mean E(Xt) prior to conducting the trend analysis [Yue et al., 2002]. By this treatment, the mean of each data set is equal to one and the properties of the original sample data remain unchanged. The trend is assumed to be linear, and the sample data are detrended by:
 2. The lag-1 serial correlation coefficient (r1) of the detrended series Yt is computed. If r1 is not significantly different from zero, the sample data are considered to be serially independent and the MK test is directly applied to the original sample data. Otherwise, it is considered to be serially correlated and pre-whitening is used to remove the AR(1) process from the detrended series as follows:
 The residual series after applying the TFPW procedure should be an independent series.
 3. The identified trend (Tt) and the residual Y′t are combined as:
 The blended series (Y″t) just includes a trend and a noise and is no longer influenced by serial correlation. Then the MK test is applied to the blended series to assess the significance of the trend.
3.5. Mann-Kendall Trend Analysis
 The Mann–Kendall test method [Mann, 1945; Kendall, 1975; Yang et al., 2009, 2010] was used to detect trends in regional series of annual climate extremes in this study, because serial correlation of regional series is not statistically significant (at the 5% level of significance) according to the results of partial auto-correlation test. Meanwhile, the Sen's slope method [Sen, 1968] was used to estimate the regression coefficients or trend magnitudes (slopes) based on Kendall's tau.
3.6. Bayesian Moving Average (BMA)
3.6.1. General Formulation
 Bayesian model averaging (BMA) has recently been proposed as a way of correcting under dispersion in ensemble forecasts [Raftery et al., 2005; Min and Hense, 2006]. BMA is a standard statistical procedure for combining predictive distributions from different sources and provides a way of combining statistical models and at the same time calibrating them using a training data set. The output of BMA is a probability density function (pdf), which is a weighted average of pdfs centered on the bias-corrected forecasts. The BMA weights reflect the relative contributions of the component models to the predictive skill over a training sample. The combined forecast pdf of a variable y is:
where p(y|Mk, yT) is the forecast pdf based on model Mk alone, estimated from the training data; k is the number of models being combined. p(Mk|yT) is the posterior probability of model Mk being correct given the training data. This term is computed with the aid of Bayes' theory:
 Considering the application of BMA to bias-corrected forecasts from the k models, equation (4) can be rewritten as:
where ωk = p(Mk|yT) is the BMA weight for model k computed from the training data set and reflects the relative performance of models k on the training period. The weights ωk add up to 1, the conditional probabilities pk[y|(fk, yT)] may be interpreted as the conditional pdf of y given fk (i.e., model k has been chosen) and training data yT. These conditional pdfs are assumed to be normally distributed as:
where the coefficients ak and bk are estimated from the bias-correction procedures described above. This means that the BMA predictive distribution becomes a weighted sum of normal distributions with equal variances but center at the bias-corrected forecast which can also be obtained from the BMA distribution using the conditional expectation of y given the forecasts:
 This forecast would be expected to be more skilful than either the ensemble mean or any one member, since it has been determined from an ensemble distribution that has had its first and second moments bias-corrected using recent verification data for all the ensemble members. It is essentially an “intelligent” consensus forecast, weighted by the recent performance results for the component models.
3.6.2. Model Weights
 The BMA weights and the variance σ2 are estimated using maximum likelihood [Raftery et al., 2005]. For given parameters to be estimated, the likelihood function is the probability of the training data and is viewed as a function of the parameters. The weights and variance are chosen so as to maximize this function (i.e., the parameter values for which the observation data were most likely to have been observed). The algorithm used to calculate the BMA weights and variance is called the expectation maximization (EM) algorithm [Dempster et al., 1977]. The method is iterative and converges to a local maximum of the likelihood. The detailed description of the BMA method is provided by Raftery et al. , and more complete details of the EM algorithm by McLachlan and Krishnan . The value of σ2 is related to the total pooled error variance over all the models in the training data set.
3.6.3. Length of Training Period
 In climate research, the longer the training period is, the better the BMA parameters are estimated. In this study, a 40-year period (1960–2000) was used to train BMA weights for five CGCM models under the 20C3M emission scenario. The rest period (2001–2100) was used in generating present and future scenarios of climate extremes in the Tarim River Basin. As three emission scenarios (A2, A1B and B1 scenario) are involved, the period (2001–2009) was not included in BMA training herein. Seven indices (i.e., FD, TN90, R10, CDD, R5D, SDII, and R95) produced by five CGCMs (Table 3) were used in constructing the present hindcast and future projection of climate extremes.
4.1. Observed Change of Climate Extremes in the Past Five Decades
 The analysis of temperature and precipitation reveals a variety of changes in extremes (1960–2009) in the Tarim River Basin. Spatial patterns of trends in temperature extremes have a much higher degree of coherence while precipitation in the region has more variability. The results for indices in climate extremes from the past observations are presented as below. As some station-based datum show autocorrelation (only first order), the TFPW approach is used to correct these datum. For the regional series of climate extremes, we also compared the TFPW results with the normal M-K results and found that they are quite similar (due to slight auto-correlation in the regional scale). Hence, the Mann-Kendall approach is used for regional series as they are not significantly autocorrelated. Meanwhile, the magnitudes of trend are estimated by Sen's slope estimator [Sen, 1968].
4.1.1. Cold Extremes (TX10, TN10, TXn, TNn, FD)
Figure 2 shows the spatial pattern of the trends of cold extremes for the 20 meteorological stations and Figure 3 demonstrates the regional annual series for indices of cold extremes in Tarim. The regional trends in indices of cold extremes are shown in Table 5. For cold nights (TN10, Figures 2a and 3a) and cold days (TX10, Figures 2b and 3b), about 80% and 75% of stations have decreasing trends which are statistically significant. Except for some dispersed small areas and the station Tashikuergan, significantly negative trends for TN10 and TX10 are observed in the rest region.
Table 5. Annual Trends, With 95% Confidence Intervals in Parentheses, for Regional Indices of Temperature and Precipitation Extremesa
Values for trends significant at the 5% level (t test) are set in bold face.
−0.29 (−0.68 to 1.66)
−0.15 (−0.41 to 1.20)
0.35 (−0.21 to 1.07)
0.24 (−0.36 to 1.40)
−0.12 (−0.12 to 0.23)
0.01 (−0.20 to 0.19)
0.03 (−0.20 to 0.19)
0.15 (−0.12 to 0.30)
0.18 (−0.22 to 0.44)
−0.40 (−1.81 to 1.36)
0.34 (−1.25 to 1.63)
0.41 (−2.56 to 2.70)
0.003 (−0.045 to 0.045)
0 (−0.02 to 0.01)
0 (−0.014 to 0.011)
0 (−0.035 to 0.018)
0.002 (−0.017 to 0.02)
2.59 (−10.03 to 22.17)
 Positive trends for the temperatures of coldest days in each year (TXn, Figures 2c and 3c) are observed over the whole Tarim except in station Tashikuergan. A similar pattern for the temperatures of coldest nights in each year (TNn, Figures 2d and 3d) is also found over Tarim. In addition to the above mentioned station, a region in central Tarim shows a negative trend for the TNn index. The TXn and TNn show increasing trends at approximately 90–95% of stations (Figures 2c and 3c). But only 26% and 71% of stations for these two indices have statistically significant trends due to the high variability. Positive trends in TNn at these stations are generally stronger than those at most stations.
 Strongly negative trends in the number of the frost days (FD, Figures 2e and 3e) are found in Tarim except in station Tashikuergan. About 78% of stations show statistically significant decreasing trends and stations with pronounced trend magnitudes are distributed in the western and southwestern basin.
4.1.2. Diurnal Temperature Range (DTR)
Figure 4 shows the spatial pattern of the trends of diurnal temperature range (DTR) for the 20 meteorological stations, and Figure 5 demonstrates the regional annual DTR series. Negative trends in diurnal temperature range (DTR, Figure 4) are found at 14 stations over Tarim. Positive trends in 6 stations are generally strong and significant. The strongest increasing trend at Qiemo is a linear trend of +0.026°C per decade. But negative trends are more obvious, approximately 60% of stations (Figure 4) show statistically significant decreasing trends and most stations are situated in the northern and western Tarim.
 The spatial patterns of observed trends and the regional annual series of warm extremes during 1961–2009 over Tarim are provided in Figures 6 and 7. The regional trends in indices of cold extremes are shown in Table 5. For the percentage of days exceeding the 90th percentiles (TX90, Figure 6a, and TN90, Figure 6b), statistically significant increasing trends are observed at 55% and 75% of stations. The patterns for the TX90 and TN90 indices are very similar, positive trends for most stations are statistically significant. Possible reasons for having such patterns might be due to global warming and urbanization. More than 75% of the basin showed positive trends for TXx (Figure 6c) including the northern Tarim, Kuluke Desert and Taklimakan Desert. The rest area where mainly located in the western Tarim shows negative trends. Tashikuergan is the only station which has significant negative trend for the index. Except for a small region in west Tarim, the rest parts show significant positive trends for TNx (Figure 6d). About 40% and 60% of stations have statistically significant increasing trends for TXx and TNx. Meanwhile, positive trends for growing season length (GSL, Figure 6e) are found over the whole Tarim excluding Tashikuergan. 75% of stations mainly located in the western and northwestern Tarim for this index have statistically significant positive trends. For the summer days indices (SU, Figure 6f), the spatial pattern is similar to the GSL index. In addition, Aheqi also shows negative trend, but it is very weak and not meaningful for this region.
4.1.4. Comparison of Warm and Cold Extremes
 In order to understand the relative changes of daily temperature, it is essential to compare trends in warm and cold indices, which are listed in Table 6.
Table 6. Number and Proportion of Individual Stations Where the Trend in One Index is of Greater Magnitude Than the Trend in a Second Indexa
Abbreviations are as follows: abs indicates that the absolute magnitudes of trends are compared; rel indicates that the signs of trends are retained during comparison.
TX90 > TX10
TN90 > TN10
TXx > TXn
TNx > TNn
TXx > TNx
TXn > TNn
Tx90 > TN90
TX90 > TN10
TX10 > TN10
TN90 > TX10
 About 45% of stations have higher trend magnitudes in TX90 than in TX10, and the absolute value of regional trend in TX90 is higher than that of TX10 (Table 6). For minimum temperature, the regional trend in TN90 (0.35 d/decade, Table 6) is higher than that of TN10 (−0.29 d/decade).While in individual stations, about 35% of stations have higher trend magnitudes in TN90 than in TN10. For TXx and TXn, regional trend in TXn (0.05°C/decade, Table 6) is higher than TXx (0.01°C/decade, Table 6), but only 15% of stations show larger trends in TXn. The magnitude of the regional trend in TNn is more than 3 times higher than that of TNx. At individual stations, about 70% of stations have greater trend magnitudes in TNn. Therefore, changes in TN90 and TX90 are more remarkable than changes in TN10 and TX10, while TNx and TXx seem to have smaller trend magnitudes than TNn and TXn.
4.1.5. Precipitation Extremes (SDII, RX1day, RX5day, R95, CDD, and PRCPTOT)
 Spatial distribution of temporal trends is shown in Figure 8 and regional annual series for precipitation indices are shown in Figure 9. In contrast to the temperature extremes, the significance of changes in precipitation extremes is low as suggested in Figures 8 and 9.
 Eight out of 20 stations show negative trends and the rest 12 stations show positive trends for the simply daily intensity index (SDII, Figure 8a). The strongest negative and positive trends for the index occurred in stations Tashikuergan and Shache respectively. The Shache station in the western Tarim is the only station that shows statistically significant trend.
 For the maximum 1-day and 5 day precipitation (RX1day and RX5day, Figures 8b and 8c), if Tarim is separated into two non-equal parts by a northeast-southwestward line, slightly negative trends can be observed in the southwest part and mixed negative and positive trends in northeast. This can be explained by that the south region surrounded by the Taklimagan and Kuleke Desert is extremely arid and has experienced serious droughts. About 55% and 50% of stations indicate negative trends for RX1day and RX5day, and the highest negative trends are detected at stations Kuerle and Kashi. Positive trends for the RX5day index are observed in a vast area located in the north and northwest Tarim.
 Positive trends of very wet days (R95, Figure 8d) are detected in the northern and western Tarim. The proportion of stations with positive trends for this index is 65%. Stations in the eastern Tarim have slightly positive or negative trends. The strongest positive and negative trends for R95 are found at stations Kashi and Kuerle, respectively.
 Positive trends of the consecutive dry days index (CDD, Figure 8e) are found in many parts of Tarim. In addition to the two stations (Tazhong and Roqiang) located in the Taklimakan Desert and Kuluke Desert, Kashi in the western Tarim, also has statistically significant increasing trend.
 In the northern and western Tarim, annual total precipitation (PRCPTOT, Figure 8f) shows positive trends and about 35% of stations have decreasing trends mostly occurring in the central and southern Tarim. No station has statistically significant trends. The strongest positive trend is found at the station Baicheng and the strongest negative trend is found at the station Kuerle. Both of these 2 stations are located in the northern Tarim.
4.2. Multimodel Ensemble Projections of Climate Extreme in the 21st Century
 In this section, multimodel ensemble projected future changes for the 7 temperature and precipitation-based indices (Table 4) over Tarim based on the BMA method are presented. To answer the growing public concerns on climate change of Tarim in the forthcoming decade of 21st century, we mainly demonstrated and addressed the spatial changes in the 2010s under A1B scenario for sake of brevity as well. However, the temporal change processes in climate extremes (1980–2100) under all the four scenarios (20C3M, A2, A1B and B1) are addressed to show all changes.
4.2.1. Temperature-Based Extremes
 A general increase of TN90 is observed in Tarim, indicating longer warm nights (Figure 10a) in the second decade of 21st century. The increase is more pronounced in the southern Tarim. The growth rate is about 4% to 5% over Tarim. As indicated by Figure 10b, consistent increases of TN90 can be commonly found in the 21st century under all three emission scenarios (A2, A1B and B1). However, the increases are not sensitive to the emission scenarios up till 2050. Since then, more pronounced increase of the TN90 index can be found in A2 and A1B scenario when compared with B1. In the end of 21st century, the increase of TN90 is around 63% (A2), 53% (A1B), and 39% (B1). There are some differences of the increase among the 3 scenarios and they are usually consistent with emission values.
 Changes of FD are presented in Figures 10c and 10d. Pronounced negative change is found in the western basin (Figure 10c). The decrease is around 9–20 days. Slight decrease (4–9 days) is found significant in the eastern Tarim. The temporal evolution of FD (Figure 10d) shows an obviously consistent decrease in the 21st century. The decrease is similar under A1B and A2 scenarios while relatively smaller decrease is found under B1 scenario.
4.2.2. Precipitation-Based Extremes
 Over the whole Tarim, CDD (consecutive dry days, Figure 11a) shows a negative change in the 2020s compared with 2000s. In the eastern Tarim, CDD tends to reduce notably to 129 days. CDD is decreasing (94 days/decade) at stations Aheqi and Keping, indicating a higher possibility of future humid in future ten years induced by impacts of global warming on glacier recession and snowmelt from the Tianshan Mountains. From north to south, these decreasing trends become slight. The regional mean CDD shows a slight increase in the temporal processes during the 21st century (Figure 11b). The trend magnitude is much smaller than TN90. Difference among the 3 emission scenarios is small and the curves overlap each other until the end of the century.
 Changes in R10 are presented in Figure 11c, which shows a negative changes. A decrease around 4–8 days is found in the south Tarim. Meanwhile, increasing R10 is also found in the southwest Tarim, indicating more floods. Temporal change of R10 shows similar features with CDD (Figure 11d). The non-significant change in the time series may be due to the incoherence of changes.
 The maximum 5d precipitation total (RX5day, Figure 12a) is an important index for flood events. As implied by Figure 12a, we can found increasing RX5day in East and decreasing RX5day in West. Increase of RX5day ranges from 0 to 14 mm, and decrease is found in 0–25 mm. However, increase of RX5day (0–14 mm) in the Taklimakan and Kuluke Desert cannot be transformed into effective runoff because of the strong potential evaporation (PET) in the desert. As showed in the Figure 12b, change of temporal processes of RX5day is different in three scenarios before 2060. After that, Difference across the 3 scenarios is small and the curves overlap each other until the end of the century.
 Changes in SDII (Figure 12c) show a general decreasing pattern over the Tarim. The decrease is more significant, especially in the Tashikuegan, Wuqiai, Kashi and Shache station located in the west Tarim (in the range of 3.4–4.8 mm, indicating a higher possibility of future drought there). The increase of R95 (Figure 12e), defined as the fraction of annual total precipitation from wetter than the 95th percentile of wet days (≧1 mm), is generally in the range of 10% in the east Tarim. Decreasing R95 is significant in west Tarim.
 Temporal changes of SDII (Figure 12d) and R95 (Figure 12e) show similar features, characterized by slight increases under all three scenarios in the 21st century. The increase is similar under A1B and A2, while they are not significant under the B1 scenario, particularly in the last 20 years of the 21st century. Compared to the temperature indices (TN90 and FD, Figures 10b and 10d), changes of the precipitation indices under the different scenarios are not distinct.
5. Conclusion and Discussions
 General findings and results in detecting changes of climate extremes using a wide range of statistical testing methods were presented by Easterling et al. [2000a, 2000b] and Alexander et al.  for the whole world, Choi et al.  for the Asia-Pacific Network (APN) region, Aguilar et al. [2005, 2009] for the Central America and northern South America, western central Africa, Zhai et al.  and You et al.  for China, Yang et al. [2009, 2010] for south China, and You et al.  for the Tibetan Plateau. However, reports in constructing reliable scenarios of future climate extremes in the Tarim River Basin of Northwest China are highly inadequate so far. Even in the study by Klein Tank et al.  on change of climate extremes in central and south Asia based on observations from 116 meteorological stations (1961–2000), only 1–2 stations were used for the basin. Hence, it is hard to help us in understanding the spatiotemporal change patterns of climate extremes in Tarim. Meanwhile, the observations in the second half of 20th century were out of date, major changes in the first decade in the 21st century need to be re-investigated using the latest datum. This work strives to present changes in temperature and precipitation extremes in Tarim using the latest observations (1960–2009) and CMIP3-CGCM multimodel ensemble projections (2010–2100). The major findings of the article are summarized and discussed as following:
 1. The 18 indices defined by the ETCCDI derived from stations over Tarim River basin during 1960–2009 were analyzed. In general, both warm and cold extremes (12 indices) have shown stronger trends in past five decades. Meanwhile, only slight changes in precipitation extremes can be observed, which means there were no overwhelming trends in precipitation extremes over Tarim in 1960–2009. From the spatial perspective, more significantly upward (downward) warm (cold) extremes are found around the desert zones than in upstream mountain zones (i.e., the Tian Shan and Kunlun Mountains systems which surround the basin). However, there are no identical governing spatial patterns for extreme precipitation change over Tarim, although some increases (decreases) are observed in lower Tarim River course (some upstream mountain zones).
 More specifically, negative trends of indices representing cold temperature extremes (i.e.TX10, TN10, FD) and conversely, positive trends for indices representing warm maximum and minimum temperature extremes (i.e., TX90, TN90, GSL, and SU25) are found in many regions of the basin. Meanwhile, the magnitudes of trends for cold/warm nights are higher than those for cold/warm days, thus trends in minimum temperature extremes are more significant than maximum temperature extremes, which is consistent with the observed decreases in DTR index [Easterling et al., 1997]. The decrease in Tarim is stronger like other parts of China than reported in all other regions [Alexander et al., 2006; You et al., 2010]. This may be associated with rapid urbanization, increased aerosol loading and other land use change. Same as other regions in China, Tarim has experienced rapid urbanization and dramatic economic growth since 1970s. Population of Tarim increased sharply from 1.7 million in 1970 to 4.2 million in 2000 [Tong, 2004], and rapid growth of urbanization and economy exerted effects on local climate. Many studies have shown these changes have strong effects on regional climate [e.g., You et al., 2008]. The warming climate causes the number of frost days (FD) to decrease significantly. This agrees with the findings in the World, APN region and China in past 50 years [Alexander et al., 2006; Choi et al., 2009; You et al., 2008]. Generally, increasing frequency and intensity of droughts are attributed to the warming climate in summer, normally drier months and ENSO events, particularly in the inland and arid zones like Tarim. Previous studies have confirmed this point in many parts of Asia [IPCC, 2007]. In addition, most temperature indices show spatially uniform patterns over the basin, even though the climate varies across the region.
 Although it is likely that there has been a statistically significant 2% to 4% increase in the frequency of heavy and extreme precipitation events when averaged across the middle and high latitudes [IPCC, 2007], our analyses indicated that the change of rainfall statistics through the second half of 20th century is dominated by variations on the inter-annual to inter-decadal time scale and that trend estimates are spatially incoherent. This result is consistent with a number of available reports [Alexander et al., 2006; You et al., 2008; Choi et al., 2009]. Besides, though IPCC found increasing trends for extreme precipitation for many locations throughout the world [IPCC, 2007], both slightly positive and negative trends are found over Tarim. However, more than half of Tarim exhibited a positive trend for the annual wet-day precipitation (PRCPTOT) index.
 2. By using the results from an ensemble of five CMIP3-CGCM models based on the BMA method [Raftery et al., 2005], the scenarios of future climate extremes (2010–2100) over Tarim are constructed and analyzed. Generally, trends of these multimodel ensemble projections are consistent with the observations in past five decades (1960–2009), which means the projections can reproduce the main features of climate extremes in past. This further confirms our findings in the spatiotemporal changes of climate extremes in the past five decades, namely, the most significant changes of extreme precipitation and temperature over Tarim are projected to show high variations in spatiotemporal scale which results in more frequent droughts with higher intensity and floods as well.
 In the climate projections for the 21st century, the extreme temperature indices (e.g., FD and TN90) show significant increases over Tarim in the end of the 21st century, indicating more frequent warm nights in the future. The increase under A2 and A1B scenarios is generally more pronounced than under B1, in correspondence to their greenhouse gas (GHG) concentration levels. Meanwhile, general increase of CDD is found over Tarim in the future. The increased CDD, together with the increases in the temperature indices may lead to more frequent droughts in the future. Slight increase of R10 in some areas, and increase of R5D, SDII, and R95 in the whole region indicate a probability of flood occurrence in Tarim along with the drought dominated in future. This finding is different from the tropical and sub-tropical monsoon dominated regions, where significant increase of precipitation extremes was projected. According to IPCC , some global models projected that during the warmer 21st century, precipitation will decrease in the subtropical regions and become more concentrated in intense rainfall events with a greater risk of droughts. In contrast, precipitation is projected to increase in the mid-high latitude regions (e.g., the Yangtze River basin in China [Xu et al., 2009]). From a regional and local perspective, there are many influences on climate in addition to broad global changes, including urbanization and terrain, and proximity to water bodies. For instance, the magnitudes of changes in temperature extremes over Tarim in continental regions are generally higher than island countries, e.g., Fiji and New Zealand [Choi et al., 2009]. This is partly because the moist atmosphere near the oceans may subdue the occurrences of extreme temperature events due to its high heat capacity compared with the drier inland atmosphere. While the magnitudes of changes in precipitation extremes over Tarim are smaller than island countries however, induced by the far proximity to water bodies.
 Although some preliminary results of changes in extreme indices over Tarim are obtained in the present work, a number of uncertainties still exist in assessing the changes of regional-scale extreme indices. More research work in the future, particularly the ensemble projections by higher resolution CGCMs or especially regional climate models (RCMs), as well as analyzing the uncertainties related to the model spread, are needed for a more profound understanding of the futures changes in climate extremes over the arid region. Meanwhile, for a large study region or a region with high heterogeneity in climate change, field significance using appropriate methods (e.g., the bootstrap resampling technique [Burn and Hag Elnur, 2002]) should be used in trend analysis in order to obtain better understanding of the spatial pattern of change.
 The work was jointly supported by grants from the National Natural Science Foundation of China (40901016, 40830639, 40830640), a grant from the State Key Laboratory of Hydrology-Water Resources and Hydraulic Engineering (2009586612, 2009585512), and grants of Special Public Sector Research Program of Ministry of Water Resources (201001066, 201001057), the National Basic Research Program of China “973 Program” (2010CB428405, 2010CB951101, 2010CB951003), and the Fundamental Research Funds for the Central Universities (2010B00714). The authors acknowledge the modeling groups, the Program for Climate Model Diagnosis and Intercomparison (PCMDI) and the WCRP's Working Group on Coupled Modeling (WGCM) for their roles in making the WCRP CMIP3 multimodel data set available. The IPCC Data Archive at Lawrence Livermore National Laboratory is supported by the Office of Science, U.S. Department of Energy. Finally, cordial thanks are also extended to the editor, Sara C. Pryor, and two referees for their valuable comments which greatly improved the quality of this paper.