Interactive comment on “ Multi-proxy reconstructions of precipitation field in China over the past 500 years ” by Feng Shi et al

The PAGES Data Stewardship Integrative Activity seeks to advance best practices for sharing data generated and assembled as part of all PAGES-related activities. As part of this activity, a team of reviewers has been constituted for the “Climate of the Past 2000 years” Special Issue. The data team is reviewing the data handling within each of the CP-Discussion papers in relation to the CP data policy and current best practices. The team has identified essential and recommended additions for each paper, with the goal of achieving a high and consistent level of data stewardship across the 2k Special Issue. We recognize that an additional effort will likely be required to meet the high level of data stewardship envisaged, and we appreciate dedication and contribution of the authors. This includes the use of Data Citations (see example in supplement). We ask authors to respond to our comments as part of the regular open interactive discussion. C1

1 Introduction 30 High-resolution regional paleoclimate field reconstructions are able to accurately reproduce the fine spatiotemporal structure of regional climate change on multiple timescales for the period prior to instrumental records.Such reconstructions are an Clim. Past Discuss., doi:10.5194/cp-2017-2, 2017 Manuscript under review for journal Clim.Past Published: 16 January 2017 c Author(s) 2017.CC-BY 3.0 License.(Shi et al., 2015a) and the optimal information extraction (OIE) method (Yang et al., 2016).In theory, the PPR-based method maximizes the retention of spatial information, but this method requires a sufficient number of suitable proxy records near the objective grid points.
The precipitation (or the variable sensitive to precipitation) field reconstruction for a large-scale region using the PPR-based method is difficult when only one type of proxy records did not cover all reconstruction areas.For example, the tree-ring 5 based reconstruction of the MADA provides significant insights into past drought patterns in eastern Asia (Cook et al., 2010), but it performs poorly in reproducing dryness and wetness in eastern China because it only incorporates data from one short tree-ring width chronology from eastern China; consequently, it would be invalid to extrapolate objective gridded drought variability on the basis of remote tree-ring records in western China (Yang et al., 2013a).
Thus, one method to fully consider regional patterns of precipitation is to use the PPR-based method in conjunction with 10 multi-proxy records with good spatial coverage.Among those PPR-based methods, one based on the OIE method (version 1.2) has been successfully applied recently to reconstruct the precipitation field in western Qinling Mountains, China from a multi-proxy dataset (Yang et al., 2016).This OIE method included the ensemble LOC regression (Shi et al., 2012), which was inspired from the original LOC method (Christiansen, 2011).The idea behind the ensemble LOC regression method is that the regression coefficient and error terms are random variables with normal distributions within the ranges of the linear 15 regression and inverse regression.The LOC regression method has already been verified to efficiently retain low-frequency climate signals (Christiansen, 2011;Shi et al., 2012).
In eastern China, there are abundant historical records of precipitation variability (e.g., Hao et al., 2015).Recently, numerous tree-ring records in western China and surrounding regions have been archived in published databases (PAGES 2k Consortium, 2013;Yang et al., 2014).This presents an excellent and timely opportunity to integrate the data from tree-ring 20 records from western China and historical records in eastern China to reconstruct the precipitation field for the whole of China.Indeed, Feng et al. (2013) reconstructed the precipitation field in East Asia using composites of multi-proxy records, which are widely used to reflect the variability in precipitation (e.g., Liu et al., 2016).However, the distribution of proxy records in western China was not sufficiently dense in that reconstruction.In addition, spatial information may have been partially lost using the EOF-based method mentioned above.This is a limitation for developing a further understanding of 25 the dominant patterns of precipitation for the period before instrumental records and their possible driving mechanisms.In this paper, we incorporated additional tree-ring records from western China compared to previous studies and used the PPRbased framework with the OIE method to reconstruct the precipitation field for China.We present an empirical attempt to explore the dominant patterns of precipitation variability before the instrumental period and try to analyse their possible origin.30 Clim. Past Discuss., doi:10.5194/cp-2017-2, 2017 Manuscript under review for journal Clim.Past Published: 16 January 2017 c Author(s) 2017.CC-BY 3.0 License.

Data and methods
To reconstruct the precipitation field, we used a gridded instrumental precipitation dataset, sensitive precipitation proxy records, and the OIE precipitation field reconstruction method.Other datasets were used to validate the reconstruction and explore possible driving mechanisms.

Instrumental data 5
Owing to the highly heterogeneous and localized variability in precipitation, the accuracy of regional precipitation estimates depends mainly on the spatial density of the stations (Wan et al., 2013).Thus, a dense distribution of datasets is considered a priority.
We selected a monthly gridded precipitation dataset, China's Ground Precipitation 0.5° longitude by 0.5° latitude Grid Dataset V2.0 (Zhao and Zhu, 2015), covering the period AD 1961-2010, since this dataset is calculated based on nearly all 10 available national surface stations (n = 2474) in China.We targeted precipitation data from the warm season (May−September, MJJAS) because historical documents mainly record variations in MJJAS precipitation.The whole of China includes 4189 grid points.The calibration period was set to AD 1981-2000and the validation period was AD 1961-1980.Another instrumental precipitation dataset was used in this study to validate the reconstruction: The Homogenized Monthly 15 Precipitation Dataset in China for the interval 1900-2009 (Li et al., 2012), with a 5° longitude by 5° latitude grid resolution.This dataset includes data from all available national surface stations in China before AD 1960.The MJJAS mean precipitation anomaly for China for the interval AD 1920-1960 is selected here as an independent verification data.Some unusual values were evident for the period before AD 1920, as shown in original Figure 12 of Li et al. (2012), and these are likely a result of the sparse distribution of observation points.The two datasets are both developed by, and available from, 20 the National Meteorological Information Center, Chinese Meteorological Administration.

Tree-ring record
The use of tree-ring records has multiple advantages, including their annual resolution, easy replication, wide distribution, and significant corroboration from instrumental records.Consequently, such records are considered to be primary and practical archives for reconstructing precipitation fields in China over the past half millennium and are widely used to 25 reconstruct regional precipitation variability in China (e.g., Yang et al., 2014).
The candidate tree-ring records are required to satisfy two conditions: 1) each record needs to at least cover the period from AD 1876 to AD 1975; 2) if raw tree-ring width measurements are available, they must be based on at least five samples for each year to ensure good replication.
To maximize the overlap lengths of the instrument data and proxy records, all tree-ring records were extrapolated to AD 30 2000 using the RegEM algorithm (Schneider, 2001).Here, the truncation parameters for the RegEM algorithm were set to the same values as that used by Mann et al. (2008).A total of 242 of 491 tree-ring chronologies were extrapolated.The maximum and mean extrapolation lengths of the 242 chronologies were 24 years and 10.5 years, respectively.The extrapolation bias was ignored because of the short extrapolation length.
We synthesized 491 tree-ring records from China and surrounding area, as shown in Figure 1.The tree-ring records are located mainly in India, southeastern Asia, the Tibetan Plateau, Xinjiang, Mongolia, and Japan.There are only very few tree-5 ring records in northeastern China and eastern China.This indicates that the currently available tree-ring records are not sufficient for reconstruction of the precipitation field in eastern China.

Dryness/Wetness index (DWI)
In eastern China, abundant records of drought and flood disasters can be found in historical documents, which provide another opportunity to reconstruct past climate (Zhang, 1991;Ge et al., 2008).A valuable example is the Yearly Charts of 10 Dryness/Wetness in China for the Last 500-Year Period dataset (Chinese Academy of Meteorological Science, 1981).The DWI dataset, also known as the drought/flood indices (DFI), has been widely used to assess precipitation variability in eastern China (e.g., Qian et al., 2003a;Hao et al., 2015).This DWI dataset includes data from 120 locations that are distributed mainly in eastern China and northeastern China, with a few in western China.Herein, 12 DWI are excluded because their covering too short time periods.The DWI for each year 15 has five grade values: very wet (grade 1), wet (grade 2), normal (grade 3), dry (grade 4), and very dry (grade 5).This dataset has been extended using the annual average and standard deviation of observed rainy season precipitation (May-September, MJJAS) after AD 1980 (Zhang and Liu, 1993;Zhang et al., 2003).Thus, it does not need to be extrapolated to AD 2000.
However, 51 DWI needs to be interpolated because of missing values.The RegEM method was also used to interpolate the DWI data.The maximum (mean) length for interpolation was 400 (79) years.Any interpolation bias was ignored, as the 20 historical documentary data record a regional drought or flood event, rather than a local phenomenon, and they show good regional spatial consistency.
In total, we assembled 600 proxy records, which included 489 tree-ring width chronologies, 2 tree-ring oxygen isotope chronologies, 108 DWI, and an instrumental precipitation record from South Korea covering AD 1771-2000.Each record is required to be significantly correlated with one or more instrumental precipitation record at the 90% (p < 0.1) confidence 25 level during the overlap period, based on both raw data and linearly detrended data.Moreover, previous studies have repeatedly verified a significant relationship between tree-ring precipitation reconstructions and the DWI on a regional scale (e.g., Zhang, 2010).The common period to all of the proxy records is AD 1875-1978.The number of proxy records has a visible changing point from AD 1470 to AD 1469 after extrapolation/interpolation, with a 9.50% decrease from 155 to 98.
The details for each record are provided in Table s1 and Figure  timescales based on instrumental analysis (Huang and Wu, 1989;Ma, 2007;Qian and Zhou, 2014).Multiple ENSO and PDO reconstructions over the past millennium have been assessed in our previous work (Shi et al., 2016a).Without loss of generality, we selected three reconstructed ENSO indices (Cook et al., 2008;McGregor et al., 2010;Li et al., 2013b) and three PDO indices (D'Arrigo et al., 2001;D'Arrigo and Wilson, 2006;Shen et al., 2006) that have good performances in relationship with instrumental data.Note that three ENSO indices have a strong significant relationship during the common 5 period AD 1650-1977 with the range of correlation coefficients [0.58, 0.66, 0.84], but three PDO indices are only very weakly related during their common period AD 1700-1979 with the range of correlation coefficients [0.03, 0.09, 0.13].As argued recently (e.g.Newman et al., 2016), the PDO cannot be considered as a single dynamical process but results of the combined influence of remote tropical forcing and local North Pacific atmosphere-ocean interactions.This makes it a particularly challenging target for proxy-based reconstructions, explaining the poor agreement between the available series. 10

Climate model simulation
Six coupled climate models were used to assess whether their past1000 modeling experiments for the interval AD 850-1849 are consistent with our reconstruction.These are BCC-CSM1.1 (Wu et al., 2010), CCSM4 (Landrum et al., 2012), FGOALS-s2 (Man et al., 2014), GISS-E2-R (Schmidt et al., 2014), IPSL-CM5A-LR (Dufresne et al., 2013), and MPI-ESM-P (Jungclaus et al., 2010).The description of the six models, sponsoring institutions and main references was shown in Table 15 s3 of Shi et al. (2015a).For details and data of the past1000 experiments, see the websites of the Paleo Modelling Intercomparison Project Phase 3 (PMIP3) and the fifth phase of the Coupled Model Intercomparison Project (CMIP5).All simulated results were interpolated to the same temporal and spatial resolution as the reconstruction in this study.

Reconstruction method
The OIE method has been successfully used to reconstruct the South Asian summer monsoon index over the past millennium 20 (Shi et al., 2014), the Northern Hemispheric temperature over the past two millennia (Shi et al., 2015b), and the precipitation field over the past 500 years in western Qinling Mountains, China (Yang et al., 2016).The drawback of this method is an overfitting tendency.An independent test data is needed for cross validation to avoid it.This method (version 1.2) within the PPR-based framework comprises three steps.The first step is to search for the candidate proxy records.We used the search radius method to select those following standard applications of the PPR 25 method (Cook et al., 1999).The search radius is initially set at 450 km according to the precipitation decay correlation distance (New et al., 2000).Five is the minimum number of candidate predictors to ensure good replication of the reconstruction (Cook et al., 2013).If the number of candidate predictors is less than 5, the search radius is extended by 50 km at a time until 5 candidate predictors is found.The maximum search radius is fixed at 3500 km.
The second step is to determine the weighting for the proxy record using the correlation coefficient between the candidate 30 proxy record and the reconstructed target.The third step is regression of the proxy record using the ensemble LOC regression method (Shi et al., 2012) Clim.Past Discuss., doi:10.5194/cp-2017-2,2017 Manuscript under review for journal Clim.Past Published: 16 January 2017 c Author(s) 2017.CC-BY 3.0 License.
Traditional accuracy and skill parameters, including the explained variance (r 2 ), reduction of error (RE), and coefficient of efficiency (CE) (Cook et al., 2010), were used to evaluate the reliability of the reconstructions, and the uncertainty was characterized using the standard deviation of the instrumental precipitation anomaly and the correlation coefficient between the reconstructed and instrumental precipitation anomaly (Mann et al., 2008).The significance of the correlation for the filtered time series was accessed using the effective number of degrees of freedom following Zhao et al. (2016).5 The ensemble empirical mode decomposition (EEMD) method (Wu and Huang, 2009) was used to analyse the reconstructed mean MJJAS precipitation time series for eastern China, western China, and the whole of China.The eastern and western China is simply divided along -the longitude 105°.Following Mann et al. (1995), the interannual timescale was set to < 8 years.The interdecadal timescale was defined as ≥8 years and <35 years.The multidecadal timescale was defined as ≥35 years and <100 years, and the centennial scale was >100 years.10

Results
The number of grid points as a function of the number of predictors is shown in Figure 2a.As mentioned above, the minimum number of predictors is five.The 1737 grids with five candidate proxy records account for 41.5% of the grids.The 3368 grids with ≤10 candidate proxy records account for 80.4% of the grids.The maximum number of predictors is 33.This indicates that all grid points have enough predictors.The number of grids for different search radius is shown in Figure 2b.15 The 1962 grids with a 450 km search radius account for 46.8% of the grids.The 3417 grids with search radii of ≤1000 km account for 81.6% of the grids.The maximum search radius is 3450 km.This implies that precipitation in most of the grid points can be reconstructed using nearby proxy records.
Figure 3 presents a summary of the reconstruction skills.Figure 3a, 3b and 3c show that the similarity in the patterns among the r 2 , the RE and the CE maps, characterized by a better quality of the reconstruction in eastern China (with the exception of 20 some regions in northeastern China) than in western China.The maximum explained variance is 0.91.The number of grids for which both the RE and CE values are greater than zero is 3971, accounting for 94.8% of the grids.This indicates that most of the grid points pass the cross-validation process.The uncertainties associated with the grids in southeastern China are greater than those for the grids in northwestern China in Fig. 3d because of large precipitation anomalies in southeastern China.25 Figure 4a compares the reconstructed MJJAS mean precipitation anomalies with the instrumental MJJAS mean precipitation anomalies (Zhao and Zhu, 2015) for China for the interval AD 1961-2000.The reconstructed MJJAS mean precipitation anomalies mostly agree with the instrumental data.The correlation coefficient is 0.91 (n = 40), which is significant at the 99% confidence level.Figure 4b compares the reconstructed MJJAS mean precipitation anomalies with the instrumental MJJAS mean precipitation anomalies (Li et al., 2012)  The different components of the MJJAS precipitation anomalies for eastern China (east of 105°E), western China (west of 105°E), and whole China over the past 521 years (AD 1475-1995) are obtained using the EEMD method (Fig. 5 and Fig. s1).
The amplitudes of interannual and interdecadal components in eastern China are much larger than in western China (Fig. s1), 5 but the differences of the amplitudes of other components between eastern and western China are less clear in Fig. 5.The drought/flood changes in eastern and western China are generally consistent, but the long-term trends have opposite signs.
The long-term trend in eastern China can be broadly divided into three periods: the first phase before 17th century is a partial drought condition, the second stage is wet condition from the early 17th century to the early 20th century, and then, the third stage gradually becomes a significant drought condition until now, which is consistent in previous studies (e.g., Zheng et al., 10 2006;Pei et al., 2015).The long-term trend in western China can be divided two stages, the first stage from the late 15th century to the early 19th century is a drought condition, and then, then second stage gradually become a significant wet condition until now.The long-term trend for whole China has similarities with the one in eastern China, but with a much weaker amplitude.
The centennial components in eastern and western China describe both a relative wet climate during the 16th century and a 15 drought during the 17th century.The 17th century drought is also reported in previous studies (e.g., Wang et al., 2002).
Moreover, the correlation coefficient of the centennial component in eastern and western China during the interval AD 1475-1849 is 0.67, but the correlation coefficient during the interval AD 1849-1995 is -0.39.This may suggest that the driver of the centennial component has changed after AD 1849.
Figure 6 shows the reconstructed 9-year running mean MJJAS precipitation anomalies for China for the interval AD 1470-20 2000 compared with six climate model simulations.We focus on the period during AD 1470-1849, which is a compromise between the common period of the past1000 climate model experiments (AD 850-1849) and the plausible reconstruction (AD 1470(AD -2000)).We only use the reconstruction after AD 1470 since there are no historical documents before that date.
The reference period is AD 1961-1990.Only the FGOALS-s2 results are significantly correlated with the reconstructed results during AD 1470-1849 at the 95% 25 confidence level, but the correlation is low (r = 0.23).Furthermore, there is distinct shift between the mean values of the FGOALS-s2 results and the reconstruction over that period.This is related to the choice of the reference period and the increasing trend in FGOALS-s2 model since the mid-19th century in the reconstruction.All climate model simulations show low correlations with the reconstructed result.The correlation of the simulated times series is also weak between the different models.This suggests that MJJAS mean precipitation anomalies over the past 380 years in China may be largely 30 controlled by the internal variability rather than by external forcing during the interval (AD 1470-1849).Similar conclusions were derived from the comparison of reconstructed and simulated hydroclimatic variables over the past millennium in North America (Coats et al., 2015) and in East Africa (Klein et al., 2016).The second leading mode of the MJJAS precipitation field (Fig. 7b) demonstrates a south-north anomalous rainfall dipole pattern, with drying in the middle and lower reaches of the Yellow River, and increasing rainfall across and to the south of the Yangtze River.This mode accounts for 12.6% of the total variance and also appears in the EOF2 of the reconstructed 20 MJJAS precipitation anomalies during the interval (AD 1850(AD -2000)).The South-Flood North-Drought pattern is commonly referred in previous studies from an analysis of DWI (e.g., Wang and Zhao, 1979;Qian et al., 2003b) and instrumental data (e.g., Huang et al., 1999;Yu and Zhou, 2007;Ding et al., 2008;Zhou et al., 2009).Moreover, the variations have the same sign in most of western China and southeastern China.The EOF1 from three climate models (CCSM4, FGOALS-s2, GISS-E2-R), the EOF2 from BCC-CSM1.1 model, and the EOF3 from MPI-ESM-P model reproduce a similar south-north dipole 25 pattern in eastern China to the reconstructed results, but the specific range for each model is different.Moreover, their corresponding time coefficients show that no climate model simulation demonstrates a significant relationship with the reconstructed result (figure not shown).As mentioned for EOF1, this may be perfectly well justified if the variability of the south-north dipole pattern is dominated by internal variability.
The third leading mode illustrates a "sandwich" triple precipitation pattern with increasing rainfall in the area covering the 30 middle and lower reaches of the Yangtze River valley, drying over southern and northern China, and variations of the same sign in most of western China and central China (Fig. 7c).This mode accounts for 8.1% of the total variance.The "sandwich" triple mode in eastern China has been reported based on the analysis of DWI (e.g., Wang and Zhao, 1979;Qian et al., 2003b) and instrumental data (e.g., Ding et al., 2008).The EOF1 from two climate models (BCC-CSM1.1 and MPI-ESM-P), the The superposed epoch analysis (SEA) between the precipitation, its PC1, and 35 large eruption events during AD 1470-1849 shows that volcanic activity as one important external forcing may affect the MJJAS precipitation anomalies variability for China (Fig. 8).Nevertheless, the signals are barely significant and there are similar averaged scores before and after the 10 volcanic eruption year, suggesting a weak influence of the eruption.The solar activity, as another potentially important external forcing, may also be part of the driving mechanism.This view is supported by the fact that the PC1 shows a weak significant relationship with solar activity index (Wang et al., 2005) (r = 0.17, n = 239) at the 95% confidence level for the interval AD 1611-1849.The correlation coefficient reaches 0.33 after the 11-year running mean filter.In summary, the influences of volcanic eruption and solar activity on PC1 are not very strong in our results.A pattern showing some 15 similarities to the PC1 of the reconstructions appears in three climate models (Fig. 7), but the differences, in particular in western China, are too large to ensure that it has the same dynamical origin and to use model results to determine the origin of the reconstructed pattern The second mode of annual precipitation field is the north-south dipole mode in eastern China, with variations of the same sign in most of western China and southeastern China.In fact, the north-south dipole pattern of the precipitation in eastern 20 China was found over centennial timescale during the Medieval Warm Period and the Little Ice Age from the historical documents and speleothem records (e.g., Wang et al., 2001) and was one of dominant modes over interdecadal timescale (e.g., Ding et al., 2008).
In order to explore its possible origin, we calculated the correlation of the precipitation field with the annual mean (over the months July-June) ENSO index of McGregor et al. (2010), as shown in Figure 9.The results show a similar pattern with a 25 north-south dipole mode in eastern China, and the precipitation anomalies in most of western China have a positive correlation with ENSO at the 90% confidence level.In addition, PC2 is significantly correlated with the ENSO index reconstruction (McGregor et al., 2010) at the 99% level (r = 0.22, n =200) during the interval AD 1650-1849.Moreover, the other two ENSO indices (Cook et al., 2008;Li et al., 2013b) give similar correlation maps with the precipitation field (Fig. s2), but a lower correlation coefficients with PC2.This indicates that the north-south dipole in eastern China and variations 30 of the same sign in most of western China and southeastern China are likely influenced by ENSO variability before the Industrial Revolution in our reconstruction.Those results are consistent with previous studies based on PIMP3 model simulations suggested that La Niña (El Niño)-like conditions may explain the north-south dipole in eastern China on centennial timescale (e.g., Shi et al., 2016b) and with instrumental observations indicating that ENSO was associated with summer rainfall in eastern China (e.g., Huang and Wu, 1989;Guo et al., 2012;Schubert et al., 2016).20 Based on instrumental data analysis, three general views, are used to explain the precipitation variability in eastern Asia linked to ENSO, and all of them emphasize the important bridge role of the anomalous western North Pacific anticyclone.
The first one is the equatorial Rossby wave response to ENSO via the Pacific-East Asia teleconnection (Wang et al., 2000;Zhang et al., 2011;Karori et al., 2013;Feng et al., 2016).The second one is equatorial Kelvin wave response to Indian Ocean warming during El Nin õ decaying summer which is named "Indian Ocean capacitor effect" (Xie et al., 2009).The 25 more recent third one is the nonlinear atmospheric interactions between ENSO and the annual cycle (Stuecker et al., 2013;Zhang et al., 2016).
We calculated the correlation map of the precipitation field with the PDO index (D'Arrigo et al., 2001) applying a 9-year running mean filter (Fig. s2).The results at the 90% confidence level show a pattern similar to EOF2 with a north-south dipole mode in eastern China.Moreover, the relationship between PDO index (D'Arrigo et al., 2001) and PC2 is strongly 30 significant (r = 0.41, n = 150) during AD 1700-1849 at 95% confidence level after a 9-year running mean filter.However, the other two PDO indices (D'Arrigo and Wilson, 2006;Shen et al., 2006) give different correlation maps with the precipitation field (Fig. s2), and lower correlation coefficients (-0.36 and 0.07) with PC2 after a 9-year running mean filter.This indicates that the EOF2 mode is also possibly related to variations in the PDO, but the result is sensitive to the choice of the reconstructed PDO index.
Based on the instrumental data analysis, the "sandwich" triple mode in eastern China is likely associated with a meridional tripolar teleconnection in eastern Asia: the Pacific-Japan (PJ; Nitta, 1987), Pacific-East Asia (EAP, Huang and Li, 1988), or Indo-Asia-Pacific (IAP; Li et al., 2013a).The PJ/EAP/IAP teleconnection pattern can be considered as an internal mode 5 mainly controlled by atmospheric processes (Hirota and Takahashi, 2012;Zhang and Zhou, 2015).It also can be forced by the external heating such as the anomalous convective activity in the western Pacific and tropical Indian Ocean during the El Niño decaying year (Huang and Li, 1988;Xie et al., 2009;Wu et al., 2010;Li et al., 2013a).However, there is no district evidence from the correlation maps of the reconstructed precipitation field with the ENSO and PDO indices to support this kind of mechanism for the "sandwich" pattern in this study.Alternatively, a new hypothesis was proposed recently to 10 explain the "sandwich" triple mode through the interannual change in the strength of moisture transport from the Bay of Bengal to the Yangtze corridor across the northern Yunnan Plateau (Day et al., 2015).The increased latent heating associated with an increase in water vapor along the Yangtze corridor may generate the triple mode in eastern China and variations of the same sign in most of western China and central China.
Our results indicate thus that the south-north mode variability of precipitation anomalies in China carries very likely the 15 fingerprint of ENSO evolution over the past 500 years, but the origin of the EOF1 and EOF3 patterns are not clearly established yet.This implies that the other factors such as North Atlantic Oscillation (NAO) (Wu et al., 2009;Zheng et al., 2016), interdecadal Pacific oscillation (IPO) (Song and Zhou, 2015), North Atlantic triple SST pattern (Ruan and Li, 2016) through the North Atlantic-Eurasia Teleconnection (AEAT) (Li et al., 2013a), the snow cover change of the Tibetan Plateau (Ding et al., 2009;Wu et al., 2012), and changes aerosol concentration (Li et al., 2016) may contribute to the reconstructed 20 precipitation field modes during the pre-industrial period.Some climate models (e.g., CCSM4, MPI-ESM-P) can broadly reproduce some of the dominant spatial patterns of variability of the reconstructed precipitation field for period studied.Nevertheless, the corresponding time coefficient does not match with the reconstructed series.A possible reason for this is that the precipitation changes are controlled by internal variability (e.g.related to ENSO).By constraining model result to follow the observed time series, data assimilation may then provide 25 an interesting opportunity to analyse in more detail the mechanisms at the origin of the reconstructed changes (e.g., Widmann et al., 2010;Hakim et al., 2016).

Conclusions
The precipitation field for all of China was reconstructed for the past half millennium using the OIE method and additional proxy records compared to previous studies.The reconstruction shows good performance through the cross-validation 30 process and comparison with "out-of-sample" instrumental data.
1. 30 The two dominant modes of natural climate variability, the El Niño-Southern Oscillation (ENSO) and Pacific Decadal Oscillation (PDO), are used to explore the possible connection between our reconstruction and large-scale variability, since the two indices have already been shown to affect precipitation in China on interannual, interdecadal, and multidecadal Clim.Past Discuss., doi:10.5194/cp-2017-2,2017 Manuscript under review for journal Clim.Past Published: 16 January 2017 c Author(s) 2017.CC-BY 3.0 License.
in China during AD 1900-2000.The reconstructed MJJAS mean precipitation 30 anomaly is significantly correlated to the instrumental independent data during the interval AD 1920-1960, with a correlation coefficient of 0.57 (n = 41), also significant at the 99% confidence level.This indicates that the reconstruction passes the Clim.Past Discuss., doi:10.5194/cp-2017-2,2017 Manuscript under review for journal Clim.Past Published: 16 January 2017 c Author(s) 2017.CC-BY 3.0 License.out-of-sample validation on the mean MJJAS precipitation anomalies for China.A part of the disagreements before AD 1919 can come actually from the uncertainties of Li et al. (2012), as specially explained in the instrumental data section.
Clim.Past Discuss., doi:10.5194/cp-2017-2,2017 Manuscript under review for journal Clim.Past Published: 16 January 2017 c Author(s) 2017.CC-BY 3.0 License. the corresponding time coefficients of model EOFs shows no obvious significant relationship with the reconstructed data (figure not shown), which was expected if natural variability is the main driver of the changes.
Clim.Past Discuss., doi:10.5194/cp-2017-2,2017 Manuscript under review for journal Clim.Past Published: 16 January 2017 c Author(s) 2017.CC-BY 3.0 License.largest loading in eastern China indicating a co-variability of precipitation between monsoon and non-monsoon regions.We firstly consider the influence of the external forcing on the MJJAS precipitation anomalies variability during AD 1470-1849.
Clim.Past Discuss., doi:10.5194/cp-2017-2,2017 Manuscript under review for journal Clim.Past Published: 16 January 2017 c Author(s) 2017.CC-BY 3.0 License.We calculated simulated Niño 3.4 in different seasons including the annual mean (over the months July-June), previous July to current June, previous December-January-February (DJF), current March-April-May (MAM), current June-July-August (JJA) and current MJJAS seasons.The correlation maps of five simulated MJJAS mean precipitation anomalies for China with the five-corresponding simulated annual mean Niño 3.4 indices are shown in Fig. 9.They display similar south-north dipole correlation patterns in eastern China, similar to the one from the reconstruction, for three climate models (BCC-5 CSM1-1, CCSM4, and MPI-ESM-P).The Niño3.4 indices in previous July to current June, previous December-January-February (DJF) and current March-April-May (MAM) seasons during AD 1470-1849 in CCSM4 model are significantly related to its PC1 at the 99% confidence level, and the correlation coefficients are 0.30, 0.30 and 0.27, respectively.The Niño3.4 indices in current June-July-August (JJA) and MJJAS seasons in FGOALS-s2 model during AD 1470-1849 are significantly related to its PC1 at the 99% confidence level, and the correlation coefficients both are 0.16.The Niño3.4 10 indices in previous July to current June, previous DJF, current MAM, JJA and MJJAS seasons during AD 1470-1849 in MPI-ESM-P model are significantly related to its PC3 at the 99% confidence level, the correlation coefficients are 0.23, 0.22, 0.20, 0.18, and 0.19, respectively.The EOF1 of CCSM4 model, the EOF1 of FGOALS-s2 model and the EOF3 of MPI-ESM-P model all show a similar south-north dipole mode, even the specific ranges of their spatial patterns are different.This demonstrates that ENSO has likely an imprint on the south-north dipole mode of the precipitation pattern in eastern China 15 during AD 1470-1849 in those simulations.
Clim.Past Discuss., doi:10.5194/cp-2017-2,2017 Manuscript under review for journal Clim.Past Published: 16 January 2017 c Author(s) 2017.CC-BY 3.0 License.The precipitation field reconstruction reveals three leading modes for the period AD 1470-1849 before the Industrial Revolution.The first dominant mode shows consistent variation across most of China, with the exception of the northeastern and western margins of the Tibetan Plateau.This mode does not appear to be associated to the response to volcanic eruption or the solar activity.A hypothesis is that such homogenous precipitation variations in various climate regions in China have their origin in the internal variability of the system but it was not possible to determine in the present framework through 5 which mechanism.The second mode, comprising a north-south dipole in eastern China and variations of the same sign in most of western China and southeastern China.The correlation with different reconstructions of ENSO index indicates that this dipole is likely related to variations in ENSO.The third mode is a "sandwich" triple mode in eastern China and variations of the same sign in most of western China and central China.Moreover, the precipitation field reconstruction was used to assess the skill of PMIP3 coupled climate models.For most 10 models, the dominant mode of variability is not characterized by relatively homogenous changes over all China, in contrast to the reconstructed fields.The correlation map between the five simulated MJJAS mean precipitation anomalies for China with the five-corresponding simulated annual mean Niño 3.4 indices shows that the ENSO has likely an imprint on the south-north dipole mode of precipitation anomaly in eastern China over the half past millennium in the simulations too.However, there is a clear model-reconstruction mismatch in reproducing the corresponding time development as they are not 15 able to reproduce the timing of events associated to internal variability.Acknowledgements.This work was jointly funded by the Ministry of Science and Technology of the People's Republic of China (Grant No. 2016YFA0600504), and the National Natural Science Foundation of China (Grants No. 41505081, No. 41690114, and No. 41430531).Feng Shi was supported by the "MOVE-IN Louvain" Incoming Post-doctoral Fellowship, 20 co-funded by the Marie Curie actions of the European Commission.Hugues Goosse is Research Director with the FNRS-Fonds de la Recherche Scientifique, Belgium.We thank François Klein for his help in the processing of the data.Thanks are extended to Kevin Anchukaitis, H. P. Borgaonkar, Achim Bräuning, Brendan Buckley, Edward R. Cook, Zexin Fan, Keyan Fang, Narayan P. Gaire, Xiaohua Gou, Minhui He, Katsuhiko Kimura, Paul J. Krusic, Jiangfeng Li, Jinbao Li, Eryuan Liang, Hongbing Liu, Yu Liu, Chun Qin, Jonathan Palmer, Tatyana Papina, Jianfeng Peng, Somaru Ram, Masaki Sano, Margit 25 Schwikowski, Xuemei Shao, Paul Sheppard, Jiangfeng Shi, Shri A.B. Sikder, Olga Solomina, Jinglin Wang, Chenxi Xu, Bao Yang, Fengmei Yang, Koh Yasue, Yujiang Yuan, Muhammad Usama Zafar, Masumi Zaiki, Qibing Zhang, Haifeng Zhu, and other paleoclimatic scientists who published the various tree-ring chronologies used in this study.Clim.Past Discuss., doi:10.5194/cp-2017-2,2017 Manuscript under review for journal Clim.Past Published: 16 January 2017 c Author(s) 2017.CC-BY 3.0 License.

Figure 1 :
Figure 1: Map showing the locations of proxy records (a) and plot of the first year covered by each proxy record (b).

Figure 2 :
Figure 2: The number of grid points as a function of a given number of proxy predictors (a) or a given search radius (b).

Figure 3 :
Figure 3: Skills of the reconstructed MJJAS mean precipitation anomalies in China for the 1961-1980 verification period and the 1981-2000 calibration period.

Figure 5 :
Figure 5: The intrinsic mode functions (IMFs) of the mean MJJAS precipitation anomalies for eastern China (blue line), western China (red line), and whole China (yellow line), over the past 521 years (AD 1475-1995) using the Ensemble Empirical mode decomposition (EEMD) method, including the multidecadal (a), centennial (b), and long-term trend (c) components.

Figure 6 :
Figure 6: Comparison of the 9-year running mean reconstructed and six simulated MJJAS mean precipitation anomalies for China during the interval AD 1470-2000.

Figure 8 :
Figure 8: Superposed Epoch Analysis results applied to the precipitation (a) and its PC1 (b) response to 35 large volcanic events (Sigl et al., 2015) with 90%, 95% and 99% confidence limits of the mean given as dashed, dotted, and dashed-dotted lines, respectively.5

Figure 9 :
Figure 9: Correlation maps of five simulated and a reconstructed MJJAS mean precipitation anomalies for China with the fivecorresponding simulated annual mean Niño 3.4 indices and a reconstructed annual mean ENSO index (McGregor et al., 2010).