Monopole Mode of Precipitation in East Asia Modulated by the South China Sea Over the Last Four Centuries

Precipitation in East Asia affects one quarter of the global population. However, the mechanisms governing precipitation changes at the century scale remain unclear. Reconstructions of warm season precipitation over the last 531 years show that the dominant mode of variability is a monopole covering most of China. However, this mode is mostly absent from Coupled Model Intercomparison Project Phase 5 results. In contrast, experiments using data assimilation reproduce this monopole mode well. Results show that sea surface temperature in the South China Sea is a major driver of the monopole mode of precipitation via a Gill‐type response. Warm sea surface temperatures induce a distinct baroclinic structure over the central part of eastern China comprising a low‐pressure cyclone in the lower troposphere and a high‐pressure anticyclone in the upper troposphere with rising airflow, resulting in water vapor convergence and increased precipitation in East Asia.


Introduction
Investigations of the characteristics and mechanisms governing precipitation changes are important for predicting large-scale trends in droughts and floods in East Asia (Ding et al., 2009;B. Wang et al., 2001). Precipitation variability in this region has been explored over both the recent past, using abundant instrumental meteorological data sets (e.g., Webster & Yang, 1992), and the last several millennia (e.g., Cheng et al., 2016;Guo et al., 2002). However, relatively little is known about precipitation variability over the last few centuries (F. Shi et al., 2017).
Two hypotheses have been proposed in previous studies to describe monsoon precipitation variability in East Asia over the past 500 years. The first suggests that precipitation variability over East China is dominated by multipole meridional patterns that depend on the latitudinal band (e.g., a dipolar or tripolar mode; Feng et al., 2013;Ge et al., 2013;Shen et al., 2009;Wang & Zhao, 1979), whereas the second suggests a homogeneous, monopole mode (F. Shi et al., 2017;H. Shi et al., 2018;Wang & Zhao, 1979).
The first hypothesis was mainly derived from reconstructions based on historical documents. The "north drought and south flood" pattern in eastern China was first observed in the Yearly Charts of  , 1981;Shen et al., 2009;Wang & Zhao, 1979). Subsequent analyses of 63 long drought/flood proxybased indices show drought (flood) in the northern (southern) part of eastern China during four warm periods (650-750, 1000(650-750, -1100(650-750, , 1190(650-750, -1290(650-750, , and 1900(650-750, -2000Hao et al., 2016;Zheng et al., 2014). This pattern was redefined in Ge et al. (2016) as a "drought-flood-drought" tripolar mode for a small drought region in southwestern part of eastern China. This inverse phase of the "northern wet and southern dry" pattern during the Medieval Warm Period (1000-1300 CE) was also found by J. Chen et al. (2015) using a multiproxy record comparison. This meridional precipitation pattern over eastern China persisted over the last 530 years according to a multiproxy gridded warm season precipitation reconstruction for Asia (Feng et al., 2013).
Several studies have provided evidence for the existence of the monopole mode (F. Shi et al., 2017;H. Shi et al., 2018). The monopole was identified as the leading mode using an empirical orthogonal function (EOF) analysis applied to a May-September (MJJAS) precipitation variability reconstruction covering China over the past 500 years, based on tree ring data and historical documents (F. Shi et al., 2017). The EOF main loading is located in the middle and lower reaches of the Yangtze and Yellow Rivers, as shown in Figure 1a1. This was subsequently confirmed and extended to East Asia by H. Shi et al. (2018), based on EOF analyses applied to a summer precipitation reconstruction using mostly the same tree ring data and historical documents (H. Shi et al., 2018).
In contrast, models participating in the fifth phase of the Coupled Model Intercomparison Project Phase 5 do not simulate a monopole mode as the dominant mode of precipitation variability in China over the last five centuries (F. Shi et al., 2017). Here, we use a data assimilation method based on a particle filter (Dubinkina et al., 2011) to drive model output toward signals reconstructed by proxy records. This allows an investigation of the physical mechanisms behind the reconstructed precipitation monopole pattern that are consistent with climate model simulations.

Data and Methods
Experiments applying data assimilation based on a particle filter method (Dubinkina et al., 2011)

Instrumental and Reconstructed Precipitation Data
The high-resolution (0.5°latitude × 0.5°longitude) MJJAS precipitation reconstruction over China used here is based on 479 proxy records, which include 371 tree ring total ring width chronologies, one tree ring δ 18 O chronology, and 107 drought/flood indices from historical documents using the optimal information extraction method (F. Shi et al., 2017). This data set can successfully reproduce the leading mode of precipitation variability during the instrumental period 1961-2000. This data set is similar to the summer precipitation reconstruction developed by H. Shi et al. (2018), because both data sets are based on many common proxy records (e.g., Chinese Academy of Meteorological Science, 1981; Zhang et al., 2003).
The other seven instrumental precipitation data sets are used to assess the presence of the monopole mode during various periods. The three sea surface temperature (SST) data sets are employed to evaluate the relationship between precipitation and SST (see supporting information Text S1).

Simulation Results
Data assimilation experiments use model ensembles based on the results of three climate models (CESM, CCSM4, and MPI-ESM-P). The forcings used to drive the simulations throughout the last millennium are those recommended by the third phase of the Palaeoclimate Modelling Intercomparison Project Phase 3 (Schmidt et al., 2012). A brief description of the three models are given in Text S2. Model results are only used for the period 850-1849 CE.

Data Assimilation Method
The data assimilation method is based on a particle filter and has been applied to several previous paleoclimate assimilation studies (e.g., Two approaches have been applied in data assimilation in paleoclimatology: offline and online methods. The difference between these methods lies in how the model ensemble is generated. In the offline approach, the ensemble is built once for the entire simulation period, and may come from a preexisting ensemble of simulations, as is the case here. In the online approach, the ensemble is sequentially generated, based on the simulation analysis of previous subperiods, (e.g., Matsikaris et al., 2015). The performance of the online method is expected to be better than the offline method in which the assimilation leads to changes in the states of slowly varying components of the climate system (e.g., assimilating a slow variation of ocean temperature) because the climate model would be able to propagate this information forward in time (Matsikaris et al., 2015).
In this study, an offline approach is used to assimilate the precipitation data in China with 480 members, using the methodology of Klein and Goosse (2018). This approach is used because precipitation variations have no long-term memory. Moreover, a large ensemble of simulations covering the past millennium using a general circulation model would be prohibitively expensive. As only 12 simulations are available from CESM-LME, climate states from different time steps of the simulations are used to produce more members. This approach is only valid if the forcings play a negligible role compared with natural variability (Klein & Goosse, 2018). Based on the frequency over which the various simulations are sampled to produce more members (in addition to the actual year of the data assimilated), experiments using 120, 240, 480, 960, and 2,400 particles are performed.

Sensitivity of Data Assimilation-Based Reconstructions to Input Parameters
The impact of data error, the number of data points assimilated, and the number of particles contained in the model ensembles on data assimilation-based reconstructions is assessed (Text S3 and Figure S1). Results are based on data errors calculated as 2 × RMSError, where RMSError is the root-mean-square error of the proxy-based reconstructions. This is a good compromise between deviations from the instrumental data and adequate constraints to drive the model ensemble. Based on several tests, 480 particles and 43 grid points are used in the data assimilation experiments analyzed here. The choice of the model used to produce the ensemble is also assessed, with ensembles built from simulations of the past 1,000 years performed with MPI-ESM-P and CCSM4, instead of CESM. The three reconstructions with data assimilation using these three model ensembles show similar monopole modes of precipitation variation in China ( Figure S2). In addition, their three time coefficients (principal components [PCs]) are significantly correlated at the 99% confidence level based on a Student's t test, but the correlation coefficient between the data assimilationbased reconstructed PC and the proxy-based reconstructed PC for CESM (r = 0.7) is larger than those of CCSM4 (r = 0.5) and MPI-ESM-P (r = 0.4). This indicates that the monopole mode can be roughly reproduced by the three climate models after data assimilation, and the monopole reproduced by the data assimilation reconstruction using multimember simulations of CESM is more accurate than those from reconstructions based on single simulation ensembles of CCSM4 and MPI-ESM-P. This may be due to specific model characteristics or to the number of simulations (see the discussion of model biases in Text S4 and Figures S3-S6).

EOF Analysis of the Reconstruction With Data Assimilation
To further assess the validity of the data assimilation, a comparison of EOF analyses of the reconstruction based on proxies and the reconstruction with data assimilation is shown in Figure 2. The leading modes are ordered differently in the proxy-based and data assimilation-based reconstructions. The monopole pattern is the first leading mode in the proxy-based reconstruction, whereas it is the second leading mode in the data assimilation-based reconstruction. In the reconstruction with data assimilation, the first EOF has a dipole pattern, as it does in the free model run without data assimilation. After adjusting the order, the spatial pattern of the three leading modes of MJJAS precipitation are similar in the data assimilation-based and proxy-based reconstructions. The correlation coefficients between the PCs of the reconstruction based on proxies and the reconstruction with data assimilation are also significant: 0.70, 0.73, and 0.45, respectively. The monopole pattern is also reproduced in the reconstruction with data assimilation (Figure 2), and this can be used to determine its origin.

Atmospheric Circulation Corresponding to the Monopole Mode
To understand the physical mechanisms driving the monopole mode of precipitation variability over East Asia in the data assimilation-based reconstruction, Figure 3 shows the circulation (Figures 3a-3c), cloud (Figure 3d), and moisture transport (Figure 3e) anomalies associated with the second PC of precipitation variability over East Asia. The vertical velocity at 700 or 500 hPa is typically used to describe vertical motion in the midlayer atmosphere. Here, results for 700 hPa are similar to those for 500 hPa ( Figure S7). Thus, the 700-hPa vertical velocity was chosen for Figure 3c.
Correlation patterns between circulation and the PC of the monopole mode of precipitation variability over East Asia reveal a distinct baroclinic structure over central eastern China, with a cyclonic low in the lower troposphere (Figure 3a), an anticyclonic high in the upper troposphere (Figure 3b), and strong ascending anomalies in the middle troposphere (Figure 3c). In addition to the upward-motion anomalies, significant convergence anomalies of vertically integrated moisture flux accompanied by cyclonic circulating moisture flux anomalies cover most of the middle and lower reaches of the Yangtze and Yellow Rivers (Figure 3e), which enhance the climatological northeastward moisture transport from the Bay of Bengal and the South China Sea (SCS). Accordingly, significant increases in total cloud fraction and precipitable water are seen in the central part of eastern China (Figure 3d and 3e), indicating that anomalous circulation dynamically enhances convection and contributes to increased rainfall over the central part of eastern China.

Origin of the Monopole Mode
What causes the change in circulation responsible for the monopole mode? There is a significant relationship between warm SST anomalies over the SCS and the monopole mode (PC2; Figure 4a) in the reconstruction with data assimilation. The PC2 exhibits strong interannual and decadal variations in phase with an index computed from SST anomalies averaged over the SCS (100.5-123.5°E, 0.5-23.5°N), with correlation coefficients of 0.63 ( Figure 4b1) and 0.74 for a decadally smoothed version (Figure 4b2). This suggests that SSTs in the SCS may affect PC2 variability. The correlations between the vertical velocity (meridional velocity) along 70°-140°E and the PC of the monopole mode are shown in Figure 4c1 (Figure 4c2), and the correlation of the regional mean SST anomalies in the SCS with vertical velocity (meridional velocity) along 70-140°E is shown in Figure 4d1  From the above results, we hypothesize that warm SST anomalies in the SCS act as a heat source and excite anomalous cyclonic circulation to their northwest (East Asia). The circulation anomalies shown in Figures 3, 4c, and 4d resemble a Gill solution for the atmospheric response to tropical heating with a center away from the equator (Gill, 1980;Xing et al., 2014). When the heating is symmetric about the equator, anomalous westerlies (easterlies) are induced to the east (west) of the heat source via a Rossby (Kelvin) wave response, with a low-level cyclonic circulation pair to the west and poleward sides. When the heating is located north of the equator, the anomalous cyclonic circulation to the northwest (southwest) of the heat source is strengthened (weakened). Importantly, the pressure field is in approximate geostrophic balance with the winds, suggesting that the low-pressure center over the central part of eastern China (Figure 3a) is dynamically consistent with warm SST anomalies in the SCS (Figure 4a).
A significant relationship between SSTs in the SCS and the PC of the monopole mode of proxy-based reconstructed MJJAS precipitation in China can also be found in the three observational data sets for the twentieth century ( Figure S8). Furthermore, the SCS is the dominant moisture source for precipitation in the Yangtze River Basin region in boreal summer during the instrumental period 2004-2009CE (Chen et al., 2013. The direct source of subtropical water vapor converges with midlatitude water vapor above the main rainbelt between the lower reaches of the Yangtze and Yellow River valleys during the instrumental period 1951-1999 CE (Zhou & Yu, 2005). The Indian Ocean is also an important moisture source for the other leading mode of East Asian summer monsoon rainfall (e.g., Baker et al., 2015;Zhou & Yu, 2005). However, the monopole mode is not dominant during the twentieth century. Several hypotheses may explain this behavior. The first is that this monopole is only evident at long time scales. The evidence for this is that the monopole is not only dominant over 380 years in the recent past (1470-1849 CE) but is dominant for each century over the last half millennium, although the primary loading zone varies slightly from century to century ( Figure S9; e.g., the main loading zone moves to southeastern China during the twentieth century). Further evidence for this hypothesis is that the monopole mode becomes more and more visible in instrumental data sets of MJJAS precipitation as the period covered by these data sets increases (Figures S10 and S11). Moreover, the moisture transport pattern of the monopole mode of MJJAS precipitation during 1470-1849 CE is different than the first leading mode of precipitation during 1961-2000 CE (Text S5 and Figure S12). Alternatively, additional processes may be specific to the late twentieth century (e.g., those related to aerosol forcing) that may dampen the variability associated with the monopole pattern or enhance a dipole pattern (Menon et al., 2002;T. Wang et al., 2013). However, this is difficult to verify given that the effect of aerosols included in these simulations has large uncertainties Wu et al., 2016).
An additional question is the origin of warm SST anomalies in the SCS. The El Niño-Southern Oscillation (ENSO) is a potential source as it is the largest mode of interannual climate variability in the region (Timmermann et al., 2018). Our results show that the ENSO index calculated from the data assimilationbased reconstructed SSTs is significantly and negatively correlated to the data assimilation-based reconstructed precipitation PC2 (r = -0.57) and the proxy-based reconstructed precipitation PC1 (r = -0.50) shown in Figure S13. However, the hypothesis that the monopole mode may be affected directly by ENSO through modulating SSTs in the SCS is not valid, as discussed below.
First, the link between the monopole mode and ENSO is less robust between models than the link between the monopole mode and SSTs in the SCS. The data assimilation-based reconstruction using the CCSM4 model ensemble shows a K-shape SST pattern in the Pacific ( Figure S14a), which may be related to mega-ENSO variability (B. . The reconstruction using the MPI-ESM-P model shows that the monopole mode is related to a large-scale positive SST anomaly over the whole tropical Pacific Ocean ( Figure S14b). In addition, the data assimilation-based reconstruction using the CESM model ensemble displays a "La Niña-like" SST pattern ( Figure S14c). The reason for this may be related to model bias (see the discussion of model biases in Text S4 and Figures S3-S6). Second, these patterns relating the monopole mode and SSTs shown in Figure S14 are distinctly different than the Niño 3.4 index and the SSTs shown in Figure  S15. Third, there is no significant relationship between the ENSO index reconstructed from data assimilation and the proxy-based reconstructed ENSO index. There is also no significant relationship between the ENSO index reconstructed from data assimilation and the instrumental index shown in Figure S16. This suggests that the precipitation data reconstructed from data assimilation for China is not strongly linked to ENSO and thus the data assimilation does not improve ENSO evolution compared with the instrumental ENSO index. Alternatively, this may indicate a problem with the assimilation or in the teleconnection patterns in CESM (see Text S4 and Figure S6). Finally, the correlation between the proxy-based reconstructed ENSO index and the proxy-based reconstructed precipitation in Figure S17a and in Figure 9 of F. Shi et al. (2017) follows a north-south dipole pattern rather than a monopole pattern in East China. The ENSO index is defined and calibrated by the annual mean (July-June) Niño 3.4 region SST anomalies (McGregor et al., 2010). The reconstructed precipitation based on data assimilation also shows this north-south dipole of correlation with ENSO, except in northeast China ( Figure S17b). In addition, the correlation pattern between The proxy-based precipitation reconstruction (IGGPRE.1.0.anom.nc) is archived at the NOAA website (https://www.ncdc.noaa.gov/paleo/ study/23056). This work was jointly funded by the NSFC (41888101, 41690114, 41877440, and 41430531)