Signals In Temperature Extremes Emerge in China During The Last Millennium Based On CMIP5 and CMIP6 Simulations

Though the magnitude of any change is important, regions which have a larger signal of change relative to the background variations will potentially face greater risks than other regions, as they will see unusual or novel climate conditions more quickly (Frame et al. 2017). Providing more information about signal and noise on regional scales, and the associated attribution to particular causes, is therefore important for adaptation planning (Chen et al. 2021). However, whether a detectable signal in temperature extremes emerges in China at the local or regional level during 850−2005 has not been discussed. Based on six selected and bias-corrected global models under the Coupled Model Intercomparison Project phase 5, relative to pre-industrial levels (ca 1850), we show that the temporal information of signal-to-noise ratio (S/N) in annual temperature extremes are consistent with annual mean temperature variations in China during 850−2005. Before 1850, absolute values of regional mean S/N in temperature extremes under cold climatic conditions are generally larger than that under warm climatic conditions. At the level of S/N > 1, local increasing signals of cold extremes emerge in the second half of 13th century and in the early 19th century after intensively volcanic eruptions in 1257 and 1815 in most part of China, especially in southern China and Tibet Plateau. Over the past 150 years under global warming, absolute values of regional mean S/N in temperature extremes have increasing trends. The regional mean increasing signals of warm extremes over China begin to exceed natural variability in 1969 at the level of S/N > 1, and local warm signals rst occur in 1929 in Tibet Plateau. These warming signals are related to greenhouse gas forcing.


Introduction
Knowledge of the temperature variability during the last one to two millennia provides a context for future climate change and is important for determining climate sensitivity and the processes that control warming. Since systematic instrumental temperature records only extend back to the 19th century, such knowledge mainly relies on proxy data and model simulations.
Over the past two decades, many proxy-based temperature reconstructions in East Asia or China covering the past one to two millennia have been published (Yang et Nguyen et al. 2018). However, to the best of our knowledge, whether a detectable signal in temperature extremes emerges at the local or regional level during the last millennium has not been discussed.
One approach to doing this utilizes the signal-to-noise ratio (S/N) which is the ratio of signal in temperature extremes to its natural variability (Hawkins and Sutton 2012). The S/N describes the magnitude of the climate change signal relative to its natural variability and may be useful for climate impact assessments. Though the magnitude of any change is important, regions which have a larger signal of change relative to the background variations will potentially face greater risks than other regions, as they will see unusual or novel climate conditions more quickly (Frame et al. 2017). The S/N is also important for many hazards (Loarie et al. 2009). Providing more information about signal and noise on regional scales, and the associated attribution to particular causes, is therefore important for adaptation planning (Chen et al. 2021).
Therefore, the purpose of this study is to address whether signals in temperature extremes over China exceed its natural variability during the last millennium and if yes where and when the signals will occur rst.
Data And Methods

Data
Daily minimum and maximum near-surface air temperatures of seven global climate models from the CMIP5 of the World Climate Research Programme and two CMIP6 models are available at this stage. All seven models from CMIP5 have performed the pre-industrial control run, the 20th century experiment with all forcing (or historical run) and the last millennium experiment (or past1000 run) described by Taylor et al. (2012), and similar information of two CMIP6 models is described by Eyring et al. (2016). These climate model data are freely accessible at the website https://esgf-node.llnl.gov/search/. We use only the rst run for each model to treat all of the models equally. Basic information on the nine models and their experiments is presented in Table 1.
To bias correct the performance of models in simulating climate extremes over China, we used observation data for daily minimum and maximum temperatures at the 0.25° × 0. Given the range of horizontal resolutions of the nine models and observation data, all of simulations and observation data are regridded to a relatively mid-range horizontal grid resolution of 2° × 2°. Seasonal analyses were performed according to standard procedures for winter (December-January-February, DJF) and summer (June-July-August, JJA).

Climate extreme indices
We considered 12 indices of temperature extremes in this work ( Table 2). They are the coldest night (TNn), the warmest night (TNx), tropical nights (TR), frost days (FD), warm nights (TN90p), and cold nights (TN10p), which are based on daily minimum temperature, and the coldest day (TXn), the hottest day (TXx), summer days (SU), icing days (ID), warm days (TX90p), and cold days (TX10p), which are based on daily maximum temperature. The indices are generally divided into absolute indices (TNn, TNx, TXn, and TXx), absolute threshold indices (TR, FD, SU, and ID), and percentile-based threshold indices

Signal-to-noise ratio
The S/N is an important metric for comparing local signals with natural variability (Hawkins and Sutton 2012). First, the modeling periods of pre-industrial control run shown in Table 1 Then, the signal and the noise of temperature extremes are calculated separately for each model at each grid cell. Thus, the S/N of each temperature extreme is obtained from the ratio of signal to noise in each temperature extreme for each model or CMIP5 multi-model median in each grid cell. In our work, we consider that signals emerge from natural variability or it is outside natural variability, when its absolute value of S/N is larger than 1.
Lacking of pre-industrial control run for LMR, the pre-industrial period 1651-1850 is considered as the reference period.  the multi-model median S/N in annual mean temperature are smaller than -1 to the whole country are from 0.4% to 25.1%, from 0.4% to 11.0% and from 0.5% to 4.9%, respectively. These large S/N may be driven by volcanic events (Fig. 1d), particularly after the large eruptions in 1257, 1452, 1600, and 1815 (Hartl-Meier et al. 2017). The results from six individual models are similar to the multi-model median except CSIRO-Mk3L-1-2, which has some warm signals in annual mean temperature during 850−1850. In addition, CCSM4, IPSL-CM5A-LR, and MPI-ESM-P have stronger signals than multi-model median and the other three models with respect to volcanic forcing during 850−1850 ( Fig. 1a and Fig. 2a).
During the period 1851−2005, regional mean signal in annual minimum, mean, and maximum temperature begins to be larger than natural variability in 1971, 1972, and 1980 at the level of S/N > 1, respectively ( Fig. 1a and Fig. S4a). At local scale, the warm signals in annual mean, minimum and maximum temperatures begin to emerge in 1929, 1933, and 1944 at the level of S/N > 1, respectively, and occur in most of China until 2005 (Fig. 2a and Fig. S4a). These warm signals are mainly caused by greenhouse gas forcing (Fig. 1a and Fig. 1d, Schmidt et al. 2011). All six individual models have warm signals in the 20th century, and the time of emergence in signals in annual temperature for IPSL-CM5A-LR is earlier than that for other ve models at the level of S/N > 1 (Fig. 1a).
Next, S/N in annual mean temperature from CMIP5 models are compared with that from LMR. On one hand, median signals in annual mean temperature of LMR over China do not emerge from natural variability at the level of S/N > 1, either (Fig. 1a). Speci cally, median regional mean S/N in annual mean temperature from LMR over China during 850−1850 are in the range of -0.7 to 0.6. Moreover, the correlation coe cient between median S/N in annual mean temperature from CMIP5 models and that from LMR during 1400−1850 (0.25) is higher than that during 850−1399 (-0.30). In addition, both of them show that the early 19th century is a cold period. However, local signals from the median runs of LMR do not emerge from natural variability during 850−1850 after the volcanic forcing in the CMIP5. On the other hand, during period 1851−2005, both CMIP5 simulations and LMR have the increasing trend for S/N in annual mean temperature, and their correlation coe cient is 0.87 (Fig. 1a). Nevertheless, the time of emergence in median regional mean signals of annual mean temperature over China for LMR is 10 years later than that for CMIP5 models at the level of S/N > 1.

Cold extremes
The temporal information of annual cold nights (TN10p) is consistent with annual mean temperature variations in China, with larger (smaller) S/N in cold (warm) climatic conditions before 1850 ( Fig. 1a and  Fig. 1b). The regional mean S/N in annual cold nights range from 0.1 to 1.5 before 1850. Specially, during both 1247−1276 and 1810−1828, regional mean S/N in annual cold nights in China are larger than 1. At local scale, the increasing signals in annual cold nights emerge from natural variability in the 13th century and from the end of 16th century to the rst half of 19th century at the level of S/N > 1. It is worth mentioning that during 1261−1276 and 1817−1827 the increasing signals in annual cold nights emerge from natural variability in more than half of China at the level of S/N > 1. The correlation coe cient between S/N of cold nights and radiative forcing from volcanic aerosols is -0.50, -0.95, and -0.54 during the period 900−1200, 1201−1449, and 1450−1850, respectively, and the correlation coe cients between multi-model median S/N of cold nights and other radiative forcings are low ( Table 3). The results from six individual models are also similar to the multi-model median during 850−1850 except that CSIRO-Mk3L-1-2 has some warm signals in annual cold nights, and that CCSM4, IPSL-CM5A-LR, and MPI-ESM-P have stronger signals than multi-model median and the other three models with respect to volcanic forcing at the level of S/N > 1 (Fig. 1b and Fig. 2b).
During 1851−2005, regional mean median S/N in annual cold nights in China are from -1.2 to 0.4, with decreasing signals exceeding natural variability from 1984 (Fig. 1b). All six individual models also have decreasing signals in annual cold nights, and the time of emergence in signals of annual cold nights for IPSL-CM5A-LR is earlier than that for other ve models (Fig. 1b). Local decreasing signals in annual cold nights emerge from noise in 1970s for multi-model median (Fig. 2b).
The other cold extreme indices, such as cold days (TX10p), frost days (FD), and icing days (ID) have similar features to cold nights (TN10p), but their absolute values of S/N are generally smaller than that of cold nights (Fig. S4c−S4d). To be speci c, regional mean increasing signals in annual cold nights and cold days over China emerge from natural variability during 1247−1276 and 1810−1828, and there are no regional mean signals in other temperature extremes for the whole country before 1850. At local scale, the increasing signals in annual cold nights, cold days, frost days, and icing days exceed natural variability in the second half of the 13th century and in the early 19th century. The local increasing signals in annual cold nights and cold days are outside natural variability from the end of the 16th century to the end of the 18th century.  (Fig. S5a). Both of them show that high frequency cold nights in winter in southern China occurs in 1800−1850 (Fig. S5a), but simulation presents that the rst half of this period has more cold nights than the second part has, and that is opposite to Zheng et al. (2012). In addition, they nd the intensities of some cold events in southern China are strong, such as those during 1653-1654, 1670, 1690, 1861, 1892 and 1929. The simulation can reproduce that both in 1690 and in 1892, with a weak magnitude (Fig. S5b-S5c).

Warm extremes
The warm extreme indices, such as warm days (TX90p), summer days (SU), tropical nights (TR), hottest day (TXx), and warmest night (TNx) have similar features to warm nights (TN90p), but their absolute values of multi-model median S/N are generally smaller than that of warm nights (Fig. S4b−S4d). Consistent with annual mean temperature changes, larger (smaller) absolute values of multi-model median S/N in warm nights occur in cold (warm) climatic conditions during 850−1850 ( Fig. 1a and Fig.  1c). However, there are no regional mean multi-model median signals in these warm extremes for the whole country before 1850, with regional mean multi-model median S/N in annual warm extremes being from -0.6 to 0.03. At local scale, the decreasing signals in annual TNx and TR exceed natural variability in the second half of the 13th century and in the early 19th century. The local decreasing signals in annual TNx are also larger than natural variability in the middle of the 15th century. For individual models, CCSM4, IPSL-CM5A-LR, and MPI-ESM-P have decreasing signals in warm nights with respect to volcanic forcing during 850−1850 (Fig. 1c and Fig. 2c).
As for multi-model median results, regional mean increasing signals in annual warm nights, warm days, and tropical nights for the whole country emerge from natural variability during 1851−2005, with the regional mean signals rst occurring in 1969 for annual warm nights (Fig. 1c, and Fig. S4c−S4d). There are some local signals in almost all annual warm extremes during 1851−2005, with local signals rst occurring in 1929 for annual warm nights (Fig. 2c, and Fig. S4b−S4d). The local increasing signals in annual warm nights emerge from natural variability until 1989 for all grid points of the whole country (Fig. 2c). All six individual models also have increasing signals in annual warm nights, and the time of emergence in signals of annual warm nights for IPSL-CM5A-LR is earlier than that for other ve models (Fig. 1c, and Fig. 2c).
According to the research of Zhang and Gaston (2004), the northern China heat wave in summer 1743 is severe, and another strong case is in 1215. From the view of S/N, the signals of summer maximum temperatures in 1215 and 1743 do not exceed natural variability over China based on simulation (Fig.  S6). Moreover, models cannot capture the features of hot events in 1215, but represent a weakly warmer summer in northern China in 1743 (Fig. S6b−S6c). The other warm extreme indices have similar results with summer maximum temperature (Fig. S6d−S6e).

The S/N in annual temperature extremes in subregion of China relative to the pre-industrial level
We examine the spatial patterns of S/N in temperature extremes by dividing China into four roughly equal area land regions, including northwestern China (NWC), northeastern China (NEC), Tibet Plateau (TP), and southern China (SC) (Fig. 3). On one hand, the noise variance decreases with averaging (Hawkins and Sutton 2012). On another hand, these regions were determined by annual mean temperature (Fig. 3a), noise of temperature extremes (Fig. 3b−3c), administrative boundaries and societal and geographical conditions.
Time series of area-weighted regional mean S/N in annual mean temperature, cold nights, and warm nights during 850−2005 over the land grid point of four regions in China for multi-model median relative to the pre-industrial level are shown in Fig. 4. The multi-model median S/N in annual mean and extreme temperatures in subregions of China during 850−1850 have similar features to that in China (Fig. 4−5).
The annual mean and extreme temperatures have some differences among four regions with respect to volcanic forcing during 850−1850 (Fig. 4−5). For example, northeastern China has local signals in annual cold nights with respect to strong volcanic forcing, such as after the large eruptions in 1257 and 1815; northwestern China, Tibet Plateau, and southern China have local signals in annual cold nights with respect to both weak and strong volcanic forcing (Fig. 1d and Fig. 5b). The other cold extremes, such as cold days, frost days, and icing days, have similar regional features during 850−1850, but their areas where signals emerge from natural variability are smaller than that for cold nights (Fig. S7). Compared our results with the reconstruction by Li et al. (2021), both of them show that the early 19th century is a cold period in Tibet.
As for warm signals during the 20th century, the time of emergence in signals of annual warm nights in Tibet Plateau, southern China, and northwestern China is earlier than northeastern China (Fig. 4−5). All the time of emergence in signals of the other temperature indices is latter than that of warm nights in each region of China (Fig. S7−S8).

Concluding Remarks
Based on six models chosen from CMIP5, we examine bias-corrected multi-model median S/N in temperature extremes over China during 850−2005. Our major conclusions are as follows.
(1) The temporal information of annual temperature extremes in China are consistent with annual mean temperature variations, with larger (smaller) absolute values of regional mean S/N in cold (warm) climatic conditions during 850−1850. The regional mean increasing signals in annual cold nights and cold days in China emerge from natural variability during 1247−1276 and 1810−1828 at the level of S/N > 1.
(2) At local scale, some increasing signals of cold extremes and decreasing signals of warm extremes in most of China especially in southern China and Tibet Plateau exceed natural variability in the second half of the 13th century and in the early 19th century at the level of S/N > 1. Some local increasing signals of cold extremes are also larger than natural variability in the 17th century and in the 18th century. These emerging signals are qualitatively consistent with extremely cold conditions and may be partially due to intensively explosive volcanism during the last millennium.
(3) Over the past 150 years, regional mean decreasing signals of cold extremes and regional mean increasing signals of warm extremes over China emerge from natural variability after 1969 at the level of S/N > 1, and local signals rst occur in 1929 in Tibet Plateau. These warming signals may be related to greenhouse gases forcing.
Furthermore, the agreement between the behaviors shown by the reconstructions of LMR and models is the highest during period 1851−2000, and it is higher during 1400−1850 than during 850−1399. First, the minor contribution of volcanic forcing found in LMR is in disagreement with ndings for CMIP5 simulations for the period 850−1399. This might partly be due to different responses of volcanic effect in reconstructions and model simulations (Hartl-Meier et al. 2017). Another may be partly due to less proxy records except for the uncertainty of models. Before 1400, the number of global proxy sites is less than 150, and it is between 250 to 450 during the period 1500−1850 (Tardif et al. 2019). Moreover, the proxy records over China are mainly in Tibet Plateau. Second, both reconstructions by LMR and Li et al. (2021) and simulations present that the early 19th century is a cold period, but it is related to the AMO for reconstructions (Ratna et al. 2019, Li et al. 2021) and to volcanic forcing for simulations. This might because that analyzed models lack some critical forcing or have missing or too-weak feedback mechanisms (Ljungqvist et al. 2019 10; here, TX in 10 is the calendar day 10th percentile centered on a 5-day window for the base period Table 3 The correlation coe cients between bias-correctted multi-model median S/N of cold nights and the radiative forcings from individual components: well-mixed greenhouse gases (CO 2 , CH 4