A preliminary study of a point-by-point spatial precipitation reconstruction for northwestern (NW) China is explored, based on a tree-ring network of 132 chronologies. Precipitation variations during the past ~200–400 yr (the common reconstruction period is from 1802 to 1990) are reconstructed for 26 stations in NW China from a nationwide 160-station dataset. The authors introduce a “search spatial correlation contour” method to locate candidate tree-ring predictors for the reconstruction data of a given climate station. Calibration and verification results indicate that most precipitation reconstruction models are acceptable, except for a few reconstructions (stations Hetian, Hami, Jiuquan, and Wuwei) with degraded quality. Additionally, the authors compare four spatial precipitation factors in the instrumental records and reconstructions derived from a rotated principal component analysis (RPCA). The northern and southern Xinjiang factors from the instrumental and reconstructed data agree well with each other. However, differences in spatial patterns between the instrumentation and reconstruction data are also found for the other two factors, which probably result from the relatively poor quality of a few stations. Major drought events documented in previous studies—for example, from the 1920s through the 1930s for the eastern part of NW China—are reconstructed in this study.
At present, water resources are crucial in dry northwestern (NW) China for both the civic and communal water supplies, as well as for agriculture activities. The severe drought in more than eight provinces in the semihumid-to-arid region of northern China in early 2009 is one such case. During that period, 4.6 million people were short of drinking water and nearly 30 million people suffered from water shortages for agriculture activities, forcing the central government to enact its top-level emergency drought response for the first time. This 2009 drought ranks as the most extreme during the period of instrumental monitoring for many areas (China Daily, 17 February 2009), which regrettably extends to the 1950s for most stations. The brevity of the instrumental record can limit our ability to evaluate current drought severity relative to long-term variability inherent in climate systems. We therefore draw upon the paleoclimate records available to us to place the current drought regimes into a longer context of the natural variability of climate in our study region in NW China (the area north of 35°N and west of 110°E). Several tree-ring based climate reconstructions have been published for a geographic point or small areas of dozens of square kilometers in this region (e.g., Fang et al. 2009, 2010b; Gou et al. 2008; Li et al. 2006; Liu et al. 2004; Shao et al. 2005; Sheppard et al. 2004; Yuan et al. 2003; Zhang et al. 2003). In this paper, we aim to establish the drought footprint for the entire region and to analyze its broad spatiotemporal features.
The wide distribution and annual resolution of tree rings make it possible to use quantitative methods to not only reconstruct local climate variability but also reconstruct large-scale spatial climate patterns (Cook et al. 1999). Previous dendroclimatic studies have used a tree-ring network to reconstruct the two leading drought fields of central High Asia (Fang et al. 2010a), including large parts of NW China. This research first delimited the study region into the western and eastern portions because of their distinct drought variability and reconstructed the first principle component of drought indices for both portions. The field reconstruction, however, has a limited ability to investigate drought history over space because it contains a priori spatial patterns of instrumental data in its design. For example, temporal variations in the spatial boundaries of the specified drought modes cannot be identified with this method. A point-by-point regression (PPR) spatial climate reconstruction method can ameliorate this situation by reconstructing each climate point independently, thereby allowing for “local control” over the reconstruction at each point (Cook et al. 1999). One of the most notable examples of these methods in the recent literature is the development of the North America Drought Atlas (NADA), the methodology of which was first described by Cook et al. (1999) and later fully published in Cook et al. (2004). Spatial reconstruction of all climate points in a region also facilitates an examination of underlying climate forcing via a comparison of the reconstruction with results from various general circulation models (GCMs). Several subsequent papers utilized this set of gridded reconstructions for detailed analyses of the causes of drought over a large portion of North America (e.g., Seager 2007).
The PPR-based spatial reconstructions have been conducted for many areas, such as continental North America (Cook et al. 1999, 2004; Meko et al. 1993). Recently, Cook et al. (2010) published the first ensemble-based PPR reconstruction for East Asia by calibrating 327 tree-ring chronologies with 534 grid points of the Palmer drought severity index (PDSI) for monsoonal Asia. We introduce herein a precipitation reconstruction using modified PPR methods. Our study region in the arid northwest has perhaps the densest tree-ring network in China, because of the availability of old-growth and climate-sensitive forests, though with far less spatial coverage than the previously noted studies of Cook et al. (1999, 2004) and Meko et al. (1993) over North America. The target data for reconstruction are derived from a nationwide 160-station precipitation dataset (see climate data section) that contains 26 stations within NW China. We also selected this climate dataset as predictands because it is a basic and frequently used dataset for China (e.g., Gemmer et al. 2004; Lau and Weng 2001). One difference from the study by Cook et al. (2010) is that we aim to reconstruct a 26-station precipitation dataset for NW China from a tree-ring network of 132 chronologies. Another difference is that we used a modified PPR method for reconstruction based on a “search spatial correlation contours” (see section 2d for details).
Since regional precipitation variability for much of NW China is directly related to monsoon dynamics (Fang et al. 2010a; Li et al. 2009; Shi et al. 2007), a reconstructed PPR precipitation dataset can be used to investigate long-term precipitation history and its relations with the Asian monsoon. Spatial precipitation reconstruction would be more challenging relative to a PDSI reconstruction, because tree rings in this area are often highly correlated with PDSI (Fang et al. 2009, 2010b; Li et al. 2006). We describe the tree-ring network, climate data, and regression and other analytical methods in section 2. Special attention is paid to the regression methods because they differ slightly from previous studies. For each station, the regression models are validated by independent data to test for temporal stability. The regression models are further examined according to their ability to model the temporal variability and recover the spatial patterns. These results and a discussion are presented in section 3, and section 4 summarizes the results of this study.
2. Data and methods
a. Study region
Located in east-central Asia (Fig. 1), NW China features a dry continental climate with large temperature gradients between winter and summer. According to the precipitation records used herein, it is relatively humid (more than ~200 mm yr−1) for the northwestern and southeastern portions of NW China, while it is very dry (less than 60 mm yr−1) in between, such as in southern Xinjiang, western Gansu, and western Neimenggu (Fig. 2). Peak rainfall occurs in June for western NW China (Xinjiang), where the May–July seasonal precipitation accounts for 36% of the annual precipitation. For the eastern region (areas other than Xinjiang), peak rainfall occurs in August, and July–September precipitation accounts for 64% of the yearly total precipitation. Previous studies documented differences in climate variability between western and eastern NW China (Chen et al. 2008; Li et al. 2009; Qian and Qin 2008; Shi et al. 2007; Zou et al. 2005). For example, in recent decades an increase in precipitation has been observed for western NW China, while the opposite is true for the eastern portion (Shi et al. 2007).
b. Tree-ring data
Tree-ring samples were collected from the old-growth forests in scattered mountainous ranges distributed within the treeless, dry lowlands. The patchiness of high mountains and old-growth forests leads to the uneven distribution of tree-ring sites, as shown in Fig. 1. As listed in Table 1, our tree-ring network in NW China and vicinity are derived from previously published tree-ring chronologies (e.g., Fang et al. 2009, 2010b; Li et al. 2006; Liu et al. 2004; Shao et al. 2005; Sheppard et al. 2004; Yuan et al. 2003; Zhang et al. 2003), the International Tree Ring Data Bank (ITRDB; http://www.ngdc.noaa.gov/paleo/ftp-treering.html), our contributed tree-ring chronologies, and the Chinese Tree Ring Data Center (CTRDC; http://ctrdb.ibcas.ac.cn/index.asp). All of the samples were taken from coniferous species (Pinaceae and Cupressaceae, as shown in Table 1).
Most raw tree-ring measurements were standardized to remove age-related biological trends by fitting straight lines or negative exponential curves. These fitted “conservative” curves preserve low-frequency climate signals that can often be removed by more data-adaptive methods (Briffa and Jones 1990). We used a rigid cubic-spline curve, with a 50% cutoff equal to two-thirds of each series’ length, to treat some tree-ring series with strong disturbances that were not well fitted by more conservative curves. Tree-ring indices were calculated as ratios between the actual and fitted values. Detrended tree-ring indices were averaged to generate chronologies based on robust mean methodology (Cook 1985). In this study, we use the standard and autoregressive standardization (ARSTAN) chronologies but not the residual chronologies, for which low-frequency variations are largely eliminated (Cook 1985). The following analyses are only based on the reliable portion of tree-ring chronologies with more than six individual series (Meko et al. 1993). A tree-ring network with 132 chronologies was developed, and the mean length of the reliable portions of all chronologies is 391 yr (Table 1). Most tree-ring chronologies were developed from total ring-width data, and a few of them were established from early-wood width, late-wood width, early-wood density, minimum density, late-wood density, or maximum density.
c. Climate data
The target climate records for reconstruction are the monthly precipitation data from 26 stations (Table 2), derived from a nationwide 160-station dataset beginning in the 1950s [courtesy of the National Climate Center (NCC) in Beijing; http://ncc.cma.gov.cn/cn/]. Since the data for the period 1991–2005 are missing for the Shanba station, we only include data from 1954 to 1990 for that station (Table 2). The regression models based on the data since the 1950s are verified with independent data of monthly precipitation prior to the 1950s from the Global Historical Climatology Network (GHCN) (www1.ncdc.noaa.gov/pub/data/ghcn/v2). Our selection of these degraded pre-1950s data (Li et al. 2009) for verification was made to take advantage of the minimal post-1950s data for calibration. Missing values were replaced by mean values of neighboring years. For those stations without historical data prior to the 1950s, we use the interpolated data of the nearest precipitation grid from the Climatic Research Unit (CRU) TS3 dataset at 0.5° × 0.5° resolution (New et al. 2000) for verification purposes. Because the pre-1950s data could be interpolated from distant stations if there were no concurrent data nearby, we only selected the period when station historical records could be found within a reasonable distance (~500 km) or used an arbitrary 15-yr (the average length of historical data before 1951 is 14.3 yr) verification period (Table 2). For example, because data collection at the Lanzhou station began in 1933, the gridded data for verification used near the Lanzhou station were truncated at 1933.
d. Regression methods
We use the PPR method (Cook et al. 1999) to reconstruct the precipitation data of each station from the surrounding tree-ring chronologies. PPR is essentially a variant of the principal component regression (PCR), which is used to reconstruct multiple predictands from multiple predictors (Briffa et al. 1986), restricted to the one-predictand case (Cook et al. 1999). Previous PPR-based reconstructions chose a “search radius” method to locate candidate tree-ring chronologies (e.g., Cook et al. 1999, 2004), whereby candidate tree-ring predictors are those that fall within the “circle” of a given search radius centered at the tree-ring site. If insufficient tree-ring chronologies are found within the spatial area of a search radius, then the search radius is extended to a specific distance to gather more tree-ring data for regression.
One innovation of this study is the use of search spatial correlation contours to locate candidate tree-ring chronologies for the calibration of a given climate station, instead of the search radius. Spatial correlation contours are calculated between precipitation data of the target climate station for reconstruction and the surrounding gridded precipitation records from the CRU TS3 dataset. This spatial correlation map indicates the similarity of precipitation variability between the location of the target climate station and surrounding areas. Tree rings located in areas with comparable precipitation to the target climate station tend to show higher correlations with the precipitation records of that station. Tree-ring chronologies falling within the spatial ranges of the initial spatial correlation coefficient contour (0.4) are selected as candidate predictors. Similar to the search radius method (Cook et al. 1999), we expand the areas surrounding the climate station if insufficient candidate tree-ring chronologies are found. In this study, the search spatial correlation contours expand in 0.05 increments until at least 20 candidate tree-ring chronologies are available (Fig. 1). The search radius is based on the hypothesis that the climate is evenly distributed in space (Cook et al. 1999, 2004). However, the climate of a given region could be highly variable because of factors such as complex topography. We consider this method more appropriate in NW China, where variable climate conditions and complex topographic features are common. In this study, only tree-ring sites proximal to a given station were involved in regression (Fig. 1) because the spatial correlations with precipitation of distant areas are less meaningful (Cook et al. 1999).
Four steps were used in regression after the candidate chronologies were located (Cook et al. 1999). First, the candidate tree-ring chronologies were screened using correlations between climate and tree rings with a cutoff two-tailed probability, to exclude chronologies insensitive to the precipitation at a given station. Because tree rings in our study region may show a 1-yr lag climate response, tree rings of both the current year (t) and the year before (t − 1) were taken into account to generate doubled sets of predictors (Cook et al. 1999, 2010). The target precipitation data for reconstruction are the averaged monthly data from January to October. We chose not to use the data in November and December because temperatures in these months may fall below freezing, and tree rings are already formed before the cold seasons. Second, we tested the reconstructions with and without the prewhitening of the autocorrelations of both predictors and predictands. The pooled autocorrelations of various chronologies modeled through this prewhitening process need to be added back into the reconstruction, to preserve the low-frequency variability of climate change. Third, the retained tree-ring predictors were processed by PCA to generate orthogonal eigenvectors. To reduce the dimension of the predictors and to highlight the common climate signal, we excluded the higher-order tree-ring eigenvalues (less than 1) that accounted for little variance. Last, the retained orthogonal tree-ring eigenvectors were entered into the regression model based on a minimum Akaike information criterion (AIC; Akaike 1974).
Statistics employed to evaluate the model calibration and verification include correlation coefficients of the calibration period (calibration r) and verification period (verification r), as well as the reduction of error (RE). Different from the statistics of explained variance, the calculation of RE is mainly based on the actual and estimated data in the verification period. This statistic is thus sensitive to the quality of the climate records during the verification period. The theoretical range of RE is from −∞ to 1, and a RE statistic with a value greater than zero is an indicator of acceptable model accuracy (Cook et al. 1999). To test the homogeneity and quality of the verification data, we simply examine the abrupt shifts by calculating the mean differences (differences divided by the mean) between the calibration and verification datasets. Visual comparisons are made between the averaged actual and reconstructed data during the warm season (from January to October) from all stations. Following the methodology of Cook et al. (1999), we additionally calculate the Pearson correlations between the instrumental warm-season precipitations at all stations in a given year and the reconstructed data at various stations in the same year. That is, the observations for this correlation calculation are the precipitation values of various stations in a given year, instead of a time series of precipitation records between stations. The above-mentioned procedures check for model fidelity at each station over time. However, there is no guarantee that the reconstruction recovers large-scale spatial features because the only spatial component incorporated into the PPR method is the search radius (Cook et al. 1999) or, at the local scale, the searching spatial coefficients. Verifications of the reconstruction in a spatial sense are made using the varimax- (Kaiser 1960) and promax-based (Hendrickson and White 1964) rotated principal component analysis (RPCA; Richman 1986) to consider the variable (meteorological station) and the observations (precipitation records of each station).
3. Results and discussion
a. Precipitation reconstruction
A number of experiments were conducted for each target station to generate the best possible regression model with different settings (Table 2). The search spatial correlation coefficients normally did not drop below 0.4 to locate sufficient numbers of candidate tree-ring chronologies, except for the stations of Dunhuang, Qiemo, Ruoqiang, and Yulin (with the lowest search correlation coefficients ranging from approximately 0.2 to 0.3). The screening probability varies from 0.05 to 0.3 to include a sufficient number of candidate moisture sensitive chronologies to produce a reliable reconstruction. The number of chronologies retained for the reconstruction of precipitation at a given station ranges from 3 to 23 (Table 2), which is related to the screening probability and the distribution of chronologies. For example, more chronologies are involved in the regression models for Aletai and Tacheng stations, where a denser tree-ring network is found nearby (Table 2).
Autocorrelations of the retained tree-ring chronologies for regression vary from site to site. This prewhitening procedure is designed to correct the difference in autocorrelations of the candidate tree-ring chronologies within a region probably cause by an issue such as growth disturbances, which vary between sites (Cook et al. 1999). However, there is no guarantee that the reconstruction with prewhitening overwhelms the model without prewhitening (Cook et al. 1999). Most of the reconstructions (16 stations) are better calibrated after prewhitening each chronology before the modeled pooled “red noise” signal was added back in (Table 2), suggesting that corrected autocorrelation modeling is more suitable for reconstruction in this area with sparsely distributed chronologies. The reconstructed data were then scaled to the instrumental data to equalize the mean and variance during the period of overlap (Esper et al. 2005). The most recent instrumental data, that were not included in the regression process, were added to the scaled reconstructed data to generate the longest possible precipitation time series. In so doing, the common period for all the reconstructions is from 1802 to 1990, and the longest single reconstruction is from 1711 to 2008 at the Hetian station (Fig. 3).
b. Calibration and verification
The explained variance varies from 28.3% at Wuwei station to 55.7% at Lanzhou (Table 1). The relatively poor verification results found in southern Xinjiang, as indicated by low verification r and RE at Hetian, Qiemo, and Ruoqiang, may be related to the poor quality of verification precipitation data because the differences between the verification and calibration periods are particularly high for southern Xinjiang (Fig. 4). For example, precipitation for Ruoqiang station in the verification period (1941–52) is 10 times higher than during the calibration period (1953–92). Actually, there is a significant (95%) abrupt shift around 1950 for most stations (21 stations). as indicated by Mann–Kendall tests (Gerstengarbe and Werner 1999), suggesting a degraded quality of the instrumental records before 1950 for most stations. The big difference in mean values between the estimated and actual data leads to abnormally low RE values according to its definition (e.g., Cook et al. 1999). Although RE is negative for the Qiemo and Ruoqiang stations, the two stations are well calibrated with reliable precipitation data since the 1950s. RE statistics for the Kuche and Yining stations in central Xinjiang are negative, but the values of the verification r for these two stations are relatively high (0.44 and 0.52, respectively). We consider these four reconstructions reasonably accurate in modeling past precipitation, and the poor verification results may be an artifact because of the degraded quality of interpolated precipitation before the 1950s. However, there is limited success in either calibration or verification for the Hetian reconstruction, suggesting special caution should be taken in the case of this reconstruction (Fig. 4).
There are reconstructions (stations Hami, Jiuquan, Lanzhou, Wuwei, and Xining) in the central and eastern NW areas with relatively poor verification results. The reconstructions at the Lanzhou and Xining stations still have some potential for recovering past climate because the calibration r for Lanzhou is extremely high and the verification r for Xining is not very low. However, the reconstruction models for stations Hami, Jiuquan, and Wuwei show rather poor model fitness for both calibration and verification, suggesting that caution should be used when considering these three stations. Poor calibration and verification results for the stations over central NW China may be related to the sparseness of tree-ring chronologies over that region. Future spatial reconstruction models could be improved when more tree-ring chronologies are available, particularly for areas with sparse coverage, such as central NW China. In summary, we advise caution when using the reconstructions at four stations (Hetian, Hami, Jiuquan, and Wuwei).
For a given year, correlations between the instrumental and reconstructed data of all stations are highly significant (0.000 01 level), with some relatively low values in 1941, 1942, 1947, 1955, and 1972 (Fig. 5). The map correlations indicate that the actual and estimated data agree with each other better after the 1950s, as indicated by a pattern of generally increasing correlation values (Fig. 5). This is probably an indicator that the instrumental data after the 1950s are more reliable. Relatively low correlations in 1941, 1942, 1947, and 1955 are indicated by clear mismatches of the averaged actual and estimated values in these years (Fig. 5). However, the relatively low correlation in 1972 does not correspond to an obvious mismatch of the same year (Fig. 5). This is probably because the minor disagreements (e.g., abnormally high and low values) between the instrumental and reconstructed data during that year were averaged out, leading to close mean values. It should be noted that the reconstructed averaged monthly precipitation at some very dry stations in Xinjiang fall below zero for a few years (Fig. 3). This largely resulted from the scaling process, a linear transformation that is designed to equalize the mean and variance of reconstruction with instrumental data (e.g., Esper et al. 2005).
c. Spatial patterns of the actual and estimated data
The first four principal components account for 61.5% and 49.5% of the total variance of the instrumental and estimated data, respectively, which were retained for rotation using varimax and promax methods. Since the varimax and promax rotations showed similar results, we only show results for the varimax rotations for instrumental and reconstructed data (Figs. 6 and 7). For the varimax rotation of the instrumental data over the common period 1941–90, factor 1 yields the highest loading on eastern NW China, which is thus referred to as the “eastern NW factor” (Fig. 6). Similarly, factor 2 is named the “southern Xinjiang factor,” factor 3 is the “northern Xinjiang factor,” and factor 4 is the “central NW factor” (Fig. 6). Similar spatial patterns occurring in the instrumental and reconstructed data suggest that the reconstructed model does well at recovering these spatial features (Figs. 6 and 7). Varimax factor (VF) 2 for the reconstructions corresponds to VF1 of the instrumental data (eastern NW factor), though the reconstruction VF1 is more concentrated on the southeastern region. VF2 of the reconstructed data (northern Xinjiang) and reconstruction VF4 (southern Xinjiang) correspond to instrumentation VF2 and instrumentation VF3 (Fig. 6), respectively. However, instrumentation VF4 (Fig. 6), the central NW factor centering at Jiuquan, has moved eastward for the reconstruction VF4 with the highest loadings over Wuwei (Fig. 7). This might be another indicator of the relatively poor reconstructions associated with some central NW China stations (Jiuquan and Wuwei), which modify the actual spatial precipitation patterns over these areas. Meanwhile, we need to consider the possibility that the spatial ranges of different climate divisions might shift over time.
Both the instrumentation VF1 and reconstruction VF1 (Figs. 6 and 7) are similar to the previous drought classifications in eastern NW China (Li et al. 2009; Shi et al. 2007; Zou et al. 2005). The previously classified western “Xinjiang” factor in NW China is separated into southern (instrumentation VF2 and reconstruction VF4) and northern (instrumentation VF3 and reconstruction VF2) Xinjiang factors. This is confirmed by previous climate classifications at a more local scale (Qian and Qin 2008). The transitional area between western and eastern NW China, as identified in previous studies (Qian and Qin 2008; Shi et al. 2007), is comparable to the central NW China factor in this study. It should be noted that the southern Xinjiang factor shows a lower explained variance (8.9%) for the reconstruction data (reconstruction VF4) than the explained variance (16.8%) for the instrumental data (instrumentation VF2), which may be due to the limited number of tree-ring chronologies in southern Xinjiang. This indicates that future improvements are needed to sample more evenly distributed chronologies.
d. RPCA factor scores
Moisture variables in NW China are highly diverse—for example, between its western and eastern parts (Chen et al. 2008; Shi et al. 2007); therefore, we did not extract the first principal component to represent the leading mode of the entire NW China, instead we discuss the varimax rotated factors. The four varimax factor scores of the reconstructed data and the corresponding varimax factors of the instrumental data are shown in Fig. 8. The scores of the first three varimax factors of the reconstructed data match the corresponding varimax factor scores of the instrumental data well. However, there are discrepancies between the scores of the southern Xinjiang factor, that is, reconstruction VF4 and instrumentation VF2. This may be related to the degraded quality of instrumental data before the 1950s in the southern Xinjiang area, as indicated by the large differences in mean values before and after 1950, which results in high normalized varimax scores before approximately 1950 and relatively suppressed low normalized varimax scores after 1950 (Fig. 8). Meanwhile, the instrumentation VF2 with higher explained variance shows high loadings over the larger area, whereas the reconstruction VF4 has smaller explained variance, leading to changed factor scores.
The most severe dry years for VF1 are in 1862 and 1932. The dry conditions in 1932 fall within a well-known drought event from the 1920s through the 1930s (Fang et al. 2009; Liang et al. 2006). The widespread drought from the 1920s through the 1930s was most severe for the east-central factor (factor 3 in Fig. 8), an event that is particularly significant for stations Baotou, Shanba, Yinchuan, Yulin, Zhangye, and Zhongning (Fig. 3). The extreme drought in 1829 appears to cover the entire Xinjiang area, which is significant for both the northern (instrumentation VF4) and southern Xinjiang (reconstruction VF4) factors. A remarkable drought in 1944–45, evident in previous studies (e.g., Li et al. 2006), is most extreme in northern Xinjiang (VF2 in Fig. 8) but not significant in southern Xinjiang. The extreme drought in 1945 is particularly apparent in Aletai, Hami, Wulumuqi, Wusu, and Yining (Fig. 3). The extremely wet conditions in 1958, seen in reconstruction VF2 and reconstruction VF3, correspond to the wet period in the 1950s, as revealed in previous studies for this area (e.g., Fang et al. 2010a). For reconstruction VF4 in southern Xinjiang, the driest year is 1956, followed by 1960 and 1917. According to the instrumental precipitation records, it was particularly dry in 1956 in Hetian, Kashi, and Ruoqiang.
4. Concluding remarks
We have developed a tree-ring network of 132 tree-ring chronologies from ring-width data (limited density data) in NW China and the vicinity. This tree-ring network was utilized to reconstruct a 26-station precipitation dataset in NW China, derived from a nationwide 160-station dataset, using a PPR spatial reconstruction method. We introduced a search spatial correlation contour method to locate the candidate tree-ring chronologies, instead of the previously used “search radius” method. The calibration and verification procedures generally follow the standard methodologies used in previous studies. Independent data for verification were extracted from historical or gridded datasets. Based on the calibration and verification results, we advise caution when using the reconstructions at Hami, Hetian, Jiuquan, and Wuwei. Additional comparisons of spatial patterns of the actual and estimated data suggested that the reconstructions generally match the spatial patterns of the instrumental data, although the spatial boundaries of some factors vary between the instrumental and reconstructed datasets. For example, the central NW factor is shifted a long way eastward in the reconstructed data. This may be related to the poor reconstructions at a few stations (Wuwei and Jiuquan). Some severe drought events are significant in the factor scores of the reconstructed data. A spatial reconstruction for a climate dataset with much denser spatial coverage could be beneficial to the study of climate variations over space and time, but it will require the generation of additional long tree-ring chronologies.
The authors acknowledge Jianfeng Peng, Yongxiang Zhang, Yong Zhang, Qinghua Tian, Tao Yang, and other persons for their kind help in the field and in the lab. We are grateful to Fritz Schweingruber, Gordon Jacoby, Paul Sheppard, Qibing Zhang, Xuemei Shao, Neil Pederson, and other scientists who have contributed their tree-ring data in the study region to the publicly available database. We appreciate the helpful comments from Peregrine Gerard-Little and two anonymous reviewers. This research was supported by the National Basic Research Program of China (973 Program) (2009CB421306), the National Science Foundation of China (40971119 and 41001115), the NSFC Innovation Team Project (41021091), and the Chinese 111 Project (B06026).