Asymmetric and heterogeneous frequency of high and low record-breaking temperatures in China as an indication of warming climate becoming more extreme

Authors

  • Zaitao Pan,

    Corresponding author
    1. Department of Earth and Atmospheric Sciences, Saint Louis University, St. Louis, Missouri, USA
    2. Jiangsu Key Laboratory of Agricultural Meteorology, School of Applied Meteorology, Nanjing University of Information Science and Technology, Nanjing, China
    • Corresponding author: Z. Pan, Saint Louis University, Earth and Atmospheric Sciences, Saint Louis, MO, USA. (panz@slu.edu)

    Search for more papers by this author
  • Bingcheng Wan,

    1. State Key Laboratory of Atmospheric Boundary Layer Physics and Atmospheric Chemistry, Institute of Atmospheric Physics, CAS, Beijing, China
    Search for more papers by this author
  • Zhiqiu Gao

    1. Jiangsu Key Laboratory of Agricultural Meteorology, School of Applied Meteorology, Nanjing University of Information Science and Technology, Nanjing, China
    Search for more papers by this author

Abstract

[1] The ratio of record highs to record lows and their rates of decay with time has emerged as a new metric for climatic warming and extremes complementing with the well-documented daily temperature range. This paper evaluates both the observed and model-projected high/low record behaviors in China, with a particular focus on an anomalous cooling region sometimes termed “warming hole.” We found that the observed frequency decay over time of high (low) record temperatures occurred much slower (faster) than the theoretically expected rate (1/n) of the independently and identically distributed (i.i.d.) time series. The high to low record (H/L) ratio is 3.0 in the recent years, about 50% larger than those in the U.S. and Europe. A drifting-mean i.i.d. model mimicking warming climate represents the observed record behavior much better than the i.i.d. sequence, but it still cannot explain the full record variability, implying that either the past mean warming rate has been nonlinear (accelerating) or temperature variance has become more extreme. The H/L record ratio by the mid-2040s with respect to 2006 could reach 6.8 under the high emission RCP8.5 scenario, whereas the increase of the record magnitude in the representative concentration path scenario remains largely similar to the current climate. Record frequency is highly correlated not only to Pacific Warm Pool temperature but also to Atlantic Multidecadal Oscillation index, with significant positive correlation in northern and southeastern China.

1 Introduction

[2] The global mean surface air temperature over land rose 0.76°C ± 0.19 °C during 1901–2005 with more warming in high latitudes during winter [Intergovernmental Panel on Climate Change (IPCC), 2007]. Over 70% of the global land area sampled showed a significant decrease (increase) in the annual cold (warm) nights, with some regions experiencing more than double of these occurrences [Alexander et al., 2006]. Climate change occurs not only in mean temperatures but also in its variance. There has been an increase in extreme events, most notably more frequent flooding and heat waves [Christensen and Christensen, 2003; Meehl and Tebaldi, 2004]. In July 2012 during the worst U.S. drought in the past five decades, 4420 stations broke or tied their daily high maximum temperature records/times, but only 325 stations broke or tied low minimum records—an over 10 to 1 ratio (http://www.ncdc.noaa.gov/oa/climate/research/records/). On the other hand, in a relatively cool July 2008, only 500 stations broke high maximum records whereas 667 stations broke low minimum records. Thus, the number of record-breaking temperatures can serve as a measure of climatic change.

[3] Ample studies have examined the spatial-temporal distribution of extreme temperatures in China [Xiong et al., 2009a, 2009b; Zhang et al., 2009; Sun et al., 2011]. Sun et al. [2011] showed two extreme heat centers. One is in southeastern China supported by the southerly low-level jet and the other is in Xinjiang Basin where lengthened daytime solar heating accumulates in the lower troposphere. The annual frequency of hot day (maximum temperature exceeds 35°C) has increased significantly over much of China, especially over the southeastern coast and northern China [Sun et al., 2011]. However, the frequency decreased significantly in portions of the lower reaches of the Yellow River Valley after 1960. Xiong et al. [2009a] examined the frequency and strength of record-breaking temperature during 1960–2005 and found that mean annual record-breaking temperature (RBT) frequency was 2.5 days higher than theoretical expectation (12.5 days) whereas low RBT was lower by 4.6 days during the 1976–2005 period.

[4] Comparatively, the studies of global warming effect on RBT frequency are relatively fewer. Redner and Petersen studied record frequency and magnitude of RBT at Philadelphia over 126 years and concluded that the current global warming rate is insufficient to measurably influence the frequency of RBT. On the other hand Zhang et al. [2009] found that a linear warming trend of 0.025°C per year can significantly increase RBT frequency in Nanjing. Krug [2007] and subsequent workers [Wergen and Krug, 2010] developed a mathematical expression to quantify the effect and demonstrated that the global warming rate can increase the RBT frequency by 40%.

[5] Numerous temperature metrics have been used to characterize and quantify climate change, and particularly global warming. They include percentile shift, threshold values, absolute change, and duration lengthening [Alexander et al., 2006]. The most commonly used direct metrics include mean, maximum/minimum, extreme temperatures, among others. The indirect measures of temperature change include the so-called bioindicators such as growing degrees, length of growing season, and time of the first frost day. Just recently, the frequency of record-breaking temperatures has emerged as a useful criterion of temperature change [Meehl et al., 2009; Wergen and Krug, 2010]. Meehl et al. [2009], among the firsts, compared the trends of record-breaking temperature frequencies with the expected 1/n curve (n is the length of record in years) for independently and identically distributed (i.i.d.) sequences. They found that the decays of both high and low record frequencies over time clearly deviate from the expected frequency, with record highs (lows) occurring more (less) frequently. The high-low temperature record ratio over the U.S. during the past decade or so (prior to 2009) was about 2. Wergen and Krug [2010] presented a linear drifting i.i.d. model to incorporate the global warming effects on the record-breaking frequencies and found that the mean warming trend in Europe can increase the high temperature record frequency by 40% over a 30 year period based on 1975–2005 observations. Even considering the mean trend effect, the high temperature record frequency observed was still higher than expected, suggesting that the mean warming rate was nonlinear. A similar result was found in Australia on a monthly scale [Trewin and Vermont, 2010].

[6] This paper examines the spatial-temporal variation of RBT frequency and amplitude in China, both in observations and future scenario climates and explores some causal processes contributing to the RBT behaviors. Because extreme temperatures are likely to respond substantially to both greenhouse gas forcing and local-regional climate systems [Diffenbaugh et al., 2005], we will also contrast the record-breaking temperatures in China with other parts of the world. Furthermore, over China the trends of extreme temperatures, durations of heat waves, and cold snaps have been extensively reported [e.g., Zhai, 1999; Zhai and Pan, 2003]. However, temperature record-breaking statistics has not been studied in the fashion above, nor has record characteristics in the future as revealed in the newly available fifth phase of the Coupled Model Intercomparison Project (CMIP5) models.

2 Data Sets

[7] The main data in this study are the daily observed and simulated maximum and minimum surface air temperatures. There are three global daily data sets freely available to the research community [Alexander et al., 2006]. In addition, some daily data sets are commercially available, such as the Earth Info Inc. (http:/www.earthinfo.com). However, the data coverage in these data sets is incomplete over China compared with the data set from China Meteorological Administration National Meteorological Information Centre (cdc.cma.gov.cn). Many studies have used the data set [e.g., Zhai, 1999; Qian and Lin, 2004; Zhang et al., 2011; Zhou and Ren, 2011]. One of the problems with this data set is the temporal discontinuity associated with station relocations. During the 1951–2004 period, more than 31% of the stations were relocated once and 41% moved twice [Li et al., 2004, 2009]. Thus, this study uses the China Homogenized Historical Temperature data set [Li et al., 2004, 2009]. Adapting the Easterling and Peterson (P-E) technique, Li et al. [2004] identified and removed the temporal discontinuity. This data set is relatively new and thus has not been used as widely as the nonhomogenized counterpart [e.g., Mao et al., 2010]. The number of stations started from about 180 at 1951 and kept a rough linear increase to above 650 around 1960 and roughly maintained at this level since then. Thus, our analysis starts from 1961 to 2004 when the homogenized data stops at. Those stations with 10% or more data missing were excluded. Another potential nonclimate effect on any trend analysis is the urban heat island effect on surface temperature that remains in the China Homogenized Historical Temperature data. Ren et al. [2008] estimated this effect to be relatively small compared to the climate change signals in most cases. In this study, we excluded the top 55 populous urban stations with population over 1 million, leaving 553 stations, out of raw 608. These 553 stations are more or less uniformly distributed across the country except for far western China, so the domain-average values in this study are simply the arithmetic mean of all stations without considering the density of these stations.

[8] The fourth version of the NCAR Community Climate System Model (CCSM4) [Bitz et al., 2012] simulated data are extracted from the CMIP5 data portal maintained by Program for Climate Model Diagnosis and Intercomparison [Taylor et al., 2012]. The reason for choosing CCSM4 is two-fold: first CCSM4 is one of most widely used atmosphere-ocean coupled general circulation models; second, it is the model used by Meehl et al. [2009] that studied the temporal variations of record high and low temperatures in the U.S., facilitating the contrast between the two regions of similar geographical locations and sizes. Only a small number of model runs saved daily resolution temperatures. A three-member ensemble of historical runs was integrated between 1850 and 2005, each member with a slightly different initial condition. Each initial condition was taken from a year near the end of the preindustrial control run and selected based on extremes in the Atlantic meridional overturning; details on this can be found in Gent et al. [2011]. The two representative concentration path (RCP) scenarios, RCP2.6 (lowest emission rate) and RCP8.5 (highest emission rate), are analyzed to embrace the breadth of the scenarios [Moss et al., 2010; Taylor et al., 2012]. Under the RCP 2.6 scenario the peak warming stabilizes at 2°C above 1901–1960 level around mid-21st century, whereas under RCP8.5 climate, global mean temperature can warm 6°C by the end of the 21 century. Each RCP scenario consists of two member runs. All members in historical and RCP runs are used to compute ensemble mean of each run and all model components have a nominal horizontal resolution of 1°.

3 Theory and Methodology

[9] The probability pn that the nth observation of a series xm= x1, x2, … xn has a higher or lower value than the previous observations, i.e., a record, (pn = Pr(xn > xi |i < n)) can be expressed as

display math(1)

provided that the values in the series are restricted to an i.i.d. sequence [Arnold et al., 1998; Benestad, 2003]. Here n is the length of the sequence in years and x represents daily Tmax (or Tmin) of a particular day in a calendar year. For example, Tmax of 1 January of each year constitutes a time series. The temperature of any 1 January will be a record if it is higher or lower than any prior 1 January temperatures.

[10] If x is not an i.i.d. variable, but it has a linear mean trend such as daily mean temperature increase under a global warming environment, it can be shown for a normally distributed x with n ≥ 7  [Krug, 2007; Wergen and Krug, 2010] that

display math(2)

[11] Here v is the mean temperature trend in C year−1 and σ is the standard deviation. For an i.i.d. sequence (without linear trend, v=0), the second term vanishes and equation ((2)) reduces to (1). If v is on the order of 0.01 C yr−1 and σ ≅ 3°C [Wergen and Krug, 2010], the trend term is about 40% of the 1/n term, meaning that the trend term would increase (decrease) the high (low) record rates by 40%.

[12] The magnitude or strength of extreme temperature is another measure of RBT. The strength of RBT of Tmax (Tmin) becomes greater (smaller) as the number of record increases. For a normal i.i.d., the strength of the kth record can be expressed as

display math(3)

where σ is standard deviation of daily temperature [Redner and Petersen, 2006; Zhang et al., 2009]. Equation ((3)) means that the magnitude of the kth record temperature is proportional to the square root of k.

[13] The main task of this study is to examine the frequency and magnitude of high and low record temperatures in the past and future decades. For comparison, however, we also compute the trends of the mean temperature itself by two methods. One is the traditional linear least-squares method. The other is the nonparametric Theil-Sen method [Sen, 1968]. In the latter method, the magnitude of a trend is the median of the trends among all possible data points

display math(4)

[14] Compared to linear least-squares, this independent Theil-Sen slope estimator is less sensitive to the outliers and data coverage duration [Kumar et al., 2009]. The statistical significance of a trend is determined using the Mann-Kendall trend test [Mann, 1945; Kendall, 1975].

4 Observed Record-Breaking Temperature in the Second Half of the 20th Century

4.1 Record-Breaking Temperature Frequency and High/Low Ratio

[15] Figure 1 plots the total number of days in the year when previous records, both high and low, were broken averaged over the 553 stations. The black solid curve represents the theoretical decay rate in record (1/n) of an i.i.d. sequence. The red and blue dashed lines incorporate the linear trend term in equation ((2)). In the first year (1961) every day is a record high and low temperature day, totaling 365 record highs and lows, because there are no prior records (data). Spatial averaging over all stations means we normalized the total number of record stations/days in a year by total number of stations. In the second year, the chance of record breaking is reduced by half (365/2), and so on. As the number of years increases, it becomes harder to break a record; the record frequency diminishes according the 1/n rate as n increases for an i.i.d. sequence [Arnold et al., 1998; Meehl et al., 2009]. By 2004, when n=44, pn=0.023 per day or 8.3 days record per year (black curve in Figure 1a). The red (blue) dots are high Tmax (low Tmin) record days. We also computed the low Tmax (high Tmin) records, but did not present them. Thus, the high (low) records hereafter refer to highest (lowest) daily maximum (minimum) temperatures, denoted as HiMax and LoMin, respectively. Figure 1 indicates that the expected 1/n decay is a good approximation to record frequency decay (both highs and lows) up to about 1978 roughly corresponding the Great Shift in 1976/1977 when the Pacific decadal oscillation shifted phase [Miller et al., 1994]. After the Shift when global warming accelerated [IPCC, 2007], the great majority of HiMax records (red dots) are above the 1/n curve, vice versa for LoMin. Before the Shift, during a globally slight cooling period, only one or two (out of 17) observed HiMax is located above the theoretical curve. It can also be seen that the measure (distance from the 1/n) is larger for the low record than high record, meaning that the daily Tmin has increased more than Tmax, in agreement with documented observations [Karl et al., 1993]. During the last 15 years or so, HiMax falls above the curve moderately, reaching around 15 per year compared with the expected 9–10, whereas LoMin falls quite below the curve down to 3–4 per year. This implies that the large HiMax/LoMin (H/L) record ratios were more due to the reduction in number of record lows, which agrees with a faster rising in Tmin than Tmax and Meehl et al. [2009] result.

Figure 1.

Number of record-breaking high and low temperature days aggregated over all stations in China. (left) The observed temporal decay with time of record days (dots) starting 1961, as compared with expected 1/n curve (solid black line). The dashed upper and lower lines are the same as the black line except that they incorporate the mean trend term in equation ((2)). (right) The ratio of high to low temperature records. The black curve is polynomial fitting of the ratios and the red line represents the ratio of 1. The dashed curves bracket the 5–95% CI. (a and b) For annual, (c and d) for summer, and (e and f) for winter.

[16] The theoretical curves with the mean trend added to pn in equation ((2)) fit observed records considerably better than the 1/n expectation, especially for the LoMin (dashed lines). The asymmetric deviation from the 1/n curve between HiMax and LoMin is because the trend v of Tmin is larger than that of Tmax. Even with the trend effects included, actual observed high record is still higher than the expectation, implying that the mean warming on Tmax is nonlinear (faster than linear) during the period, in agreement with Wergen and Krug [2010] in Europe.

[17] The H/L ratio increased from slightly less than 1 before 1973 to slightly above 1 by 1990, before increasing quickly to about 3 in 2000s (Figure 1b). A number of ratio dots fall below 5% confidence interval (CI) before the mid-1980s and they are above 95% CI after that. It should be noted that the heteroscedasticity (variance increase with time) of sample ratios toward the end is because the ratios are computed from increasingly smaller numbers. A same difference in ratios of large numerators/denominators results in less variability than in the ratios of small numbers [Meehl et al., 2009].

[18] To avoid potential effects of large records occurring in earlier time sequences on the subsequent record accounting [Wergen and Krug, 2010], we also computed backward record frequency starting from 2004 backward to 1961. Compared with forward record counting, the HiMax frequency is smaller than 1/n while LoMin frequency is higher than 1/n (not shown), just opposite of the forward counting as expected. This further shows that the high (low) record indeed had increased (decreased) during the past five decades. It also indicates that the number of records can be an effective indicator of climatic change.

[19] Summer (JJA) records in Figure 1c are similar to the annual one, but with larger spread during the early years partly due to the smaller sample size (3 vs. 12 months), but with a faster increase in H/L record ratio during the final year (Figure 1d) [Karl et al., 1993]. In winter (DJF), the LoMin falls further below the theoretical 1/n curve while in general the records are more scattered compared with summer (Figure 1e). The H/L ratio increased rapidly with time after around 1990, reaching 4 by the end of the time period (Figure 1f).

[20] The large H/L ratio can reach beyond 4 in later years. How anomalous are these “large” values? To assess the largeness of these values we carried out a statistical test using the moving block bootstrap (MBB) technique. The key difference between MBB and a conventional bootstrap resampling is that the former maintains the time sequence of the original time series while randomly generating the artificial time series [Vogel and Shallcross, 1996]. In an MBB, we choose a block length math formula , where n is sequence length, k is number of blocks to resample. Following Meehl et al. [2009] we simply choose λ = 3, partly to consider the serial correlation arising from El Niño-Southern Oscillation (ENSO) etc. signals, although some algorithms are available to optimize the λ. At each station we use MBB technique to resample the original sequence 1000 times. The test indicates that out of 1000 MBB resampled sequences only three curves have a larger end ratio than the actual observations (Figure 2), suggesting that the observed values of 3–4 are indeed large.

Figure 2.

Statistical confidence levels of annual observed high to low record temperature ratio. The blue dots are the ratio and the black curve is a nonlinear curve fit to the dots. The yellow lines are the curve fits of 1000 MBB bootstrap resample of the original time series (i.e., the blue dots). The 5–95% CI envelops (dotted) bracket the middle 90% of the yellow curves.

[21] To display the spatial distribution of temporally averaged record-breaking frequencies, we “recovered” temporal mean ratio at each station by reversing the 1/n factor, meaning that the weight of early years are smaller than later years during the temporal averaging. Figure 3 shows temporally averaged ratios across the country. The ratios are overwhelmingly larger than 1 across the country with larger ratios over the Tibetan Plateau. This result supports, in another perspective, Liu and Chen’s [2000] finding that climatic warming in the Tibetan region has been faster than surroundings. One area of interest is the blue region in the south-central China where the ratio is below unity. This region coincides with the so-called “warming hole (WH)” [Pan et al., 2009], a region of cooling during summer while global warming accelerated in the second half of the 20th century. Within the WH defined as H/L ratio <1 (delineated by the black lines) during summer in Figure 3a, low record-breaking days were more frequent than high records, opposite of the rest of the country. In winter, the ratio is overwhelmingly larger than 1 including the WH region, which implies the WH phenomenon is only present in summer, in agreement with Pan et al. [2004].

Figure 3.

(left) Temporally weighed average high to low record temperature ratios during 1961–2004. (right) Maximum temperature trend computed using the Theil-Sen method over 1961–2004. The irregular black lines in Figures 3a and 3b define the boundary of the warming hole (WH) where H/L ratio < 1. (a and b) For summer and (c and d) for winter.

[22] To assess the effectiveness of H/L record frequency in characterizing climatic change, we compared it with the typical linear trend based on least squares (not shown) and the Theil-Sen method (Figure 3b). The trends of ratio and of the mean base temperature itself are quite comparable, indicating the ratio is an effective measure of temperature changes (Figures 3c and 3d).

[23] The mechanisms responsible for abnormal cooling in the WH region are still under research [e.g., Qian et al., 2007; Meehl et al., 2012]; the behaviors of record temperatures may provide more insights into the WH mechanism. Separating the WH from the rest of the country showed that HiMax over the WH is below the 1/n curve, except for near the end, while LoMin scatters roughly around the 1/n curve (Figure 4a), suggesting that the WH feature is mainly associated with the reduction in Tmax. In particular in the WH summer, the negative trend term brings the mean-shifted expectation slightly below the 1/n curve. In the rest of the regions during summer, HiMax increase and LoMin decrease are more symmetric around the 1/n curve (Figure 4b). This symmetry seems to imply that the climatic warming in the rest of the country is likely caused by uniform shift of the probability function distribution of temperature. The wider spread of record values in the WH than the rest of the country could partly be related to the smaller number of stations participating in the average (120 in the WH vs. 433 outside). In winter, the difference between the WH and the rest is small except for larger scatters in the WH because of the smaller number of stations (Figures 4c and 4d).

Figure 4.

Same as Figure 1 left panels but for the warming hole region and the rest separately. The warming hole is defined in Figure 2. (a and b) For summer and (c and d) for winter.

[24] One might notice that the record decay over time tends to fit the theoretical 1/n curve initially and deviate from it gradually toward the end in all relevant plots. To see if this increasing departure from the 1/n expectation might be an intrinsic asymptotic feature, we plotted hypothetical time series consisting of normally distributed random numbers (Figure 5). The record frequency follows the 1/n curve exactly all the way to the end and did not show any increasing deviation. Thus, the deviations for record numbers seen in this study are real changes, not artifacts of the theoretical model.

Figure 5.

Validation of the expected 1/n decay of a hypothetical temperature time series consisting of normally distributed random numbers. The data points indeed follow the 1/n curve until record lengths increase to the end of time series, not diverging as seen in other plots.

4.2 Record-Breaking Temperature Magnitude

[25] Figure 6 shows the relation between the normalized RBT magnitude (Tk/σ) and the sequential order (k) of the RBT as expressed in equation ((3)) for the first 10 records. The increase (decrease), called slope (β), in Tmax (Tmin) amplitude tends to slow down as math formula increases. The RBT strength increase for Tmax is slower than its decrease for Tmin. For example, by the fifth record, the record high Tmax increases by 2.0 while record low Tmin decreases by 2.3. The relation Tk/σ vs. math formula is roughly linear as predicted by equation ((3)), but with k increase, β tends to decrease, deviating progressively from the linearity. Zhang et al. [2009] and Redner and Petersen, 2006 obtained also a similar concave (convex) relation for HiMax (LoMin) for the first 5–7 records at a single station. Our national average further confirms the approximate linear relation between Tk/σ and math formula for the first a few records.

Figure 6.

The change (slope) of RBT strength with the number of the records averaged over the whole country during the 1961–2004 period.

[26] During summer, the slopes show larger spatial heterogeneity than annual means (Figures 7a and 7c), likely related to the summer local convectiveness. Large slopes are over the lower reaches of the Yellow River and smallest slope over the WH region where the slope < 1.3 as compared with β > 1.7 in some surrounding areas. The slope magnitude (negative values) in winter is larger than summer, consistent with larger warming in winter than summer. The spatial distribution of LoMin RBT strength is overwhelmingly uniform (Figures 7b and 7d), suggesting that the spatially more coherent slopes are more related to large-scale weather/climate conditions. The mode of β for HiMax is 1.55 while that for LoMin is −1.80 (Figure 8). The β spectrum is wider for HiMax than for LoMin likely associated with more heterogeneous diabatic sources during daytime.

Figure 7.

Distribution of linear slope (β) of RBE strength averaged over 1961–2004 period. (a) Annual HiMax, (b) annual LoMin, (c) summer HiMax, and (d) summer LoMin.

Figure 8.

Annual histogram of the slope of RBT strength vs. the number of record k. Plotted are the proportionality coefficient β in equation ((3)) averaged over all stations. (a) HiMax and (b) LoMin.

4.3 Effects of Large-Scale Circulation on Record-Breaking Temperature

[27] To diagnose possible mechanisms responsible for the occurrence RBT, we examined the relationships of RBT with 44 major climate indices listed in NOAA/Earth System Research Lab. (http://www.esrl.noaa.gov/psd/data/correlation) and select three indices: Pacific Warming Pool (PWP), ENSO, and Atlantic Multidecadal Oscillation (AMO). PWP has strong positive correlation with RBT HiMax frequency over northern China with r>0.6 (Figure 9a), in agreement with Xiong et al. [2009a]. In the warm phase of PWP the regions of cyclogenesis shifts eastward along with the retreated subtropical high, resulting in a generally warmer (than normal) condition in southeastern China [Wu et al., 2004]. The strong positive correlation between RBT frequency is in agreement with the well-documented fact that northern China usually endures drought condition during El Niño years when the Asian moon weakens.

Figure 9.

Distribution of spatial correlations between HiMax RBT frequency and climate indices. (a) PWP, (b) ENSO, and (c) AMO.

[28] The correlation with ENSO index is largely negative and not significant (Figure 9b). The correlation of RBT with AMO is quite high with similar spatial distribution to that of PWP (Figure 9c). This suggests either both AMO and PWP have strong impacts on RBT, or AMO and WPW are highly correlated, which warrants further research.

5 The 21st Century CCMS4 Simulations

5.1 Record-Breaking Temperature Frequency

[29] Before presenting the future change of the record's spatial-temporal variations, we first examine CCSM4 skill in reproducing the observed behaviors. Overall, the general trend patterns simulated by CCMS4 are in good agreement with observations shown in Figure 1c. We will focus on only the summer temperature records simply because of its societal importance under warming climate. Unlike the observed decay of records where the record lows and highs scattered both below and above the 1/n curve, the number of record highs almost exclusively is above the 1/n curve while the record lows are below the curve (Figure 10a). Also, the model showed an earlier deviation from 1/n curve (stationarity). Compared to observed record decay over time, the model was unable to pick up the relative cooling before 1970s. Model failure to reproduce reasonable magnitude of the relative cooling is not uncommon among other CMIP5 models [Pan et al., 2013]. The domain average ratio over the last decade or so is about 5, compared with the observed of 3–4 (Figure 1b). In particular the model tends to capture the WH of low ratio in central China (smaller dots over the WH in Figure 10b).

Figure 10.

(left) Same as Figure 1, but from CCSM4 summer simulations. (a) 1961–2004 historical run, (c) 2006–2049 RCP2.6 Scenario, and (e) 2006–2049 RCP8.5 Scenario. (right) (b), (d), and (f) Corresponding HiMax/LoMin ratios over the whole period to the left panels (a), (c), and (e), respectively.

[30] For future scenario climates we chose the 2006–2049 period, close to the first half of the 21st century. Under the low emission RCP2.6 scenario, almost all the low record (blue) points are below the theoretical 1/n curve and vice versa for the high records (red dots). Both high and low records straddle the curve from 2015 to the end. The high records actually start to taper off after about 2040s, likely due to the leveling of the greenhouse gas forcing [Moss et al., 2010]. Under the high emission RCP8.5 scenario, the record trends started to depart from the curve slightly later than RCP2.6, but at a faster rate afterward (Figures 10c and 10e). At the end, the low records are far off the curve, while the high records are only moderately above it, accelerating the asymmetric feature. The accelerating departure from the i.i.d. sequence is consistent with lasting large anthropogenic forcing in the RCP8.5 scenario [Moss et al., 2010]. In both scenarios, the HiMax increase grows at a noticeably slower rate than the LoMin decrease, enlarging daily record temperature parity, consistent with well documented daily temperature range (DTR). It is worth mentioning that, typically, models have difficulty in reproducing the magnitude of DTR [Pan et al., 2013]. However, it seems CCSM4 is able to contrast the high and low records quite well. At the end of the simulation, the ratio can reach 20 (2) nationally in summer for the RCP8.5 (RCP2.6) scenarios. For comparison, during the hottest July 2012 in the U.S. in the past five decades, the H/L record ratio was 10 to 1, implying that the RCP8.5 climate of 2040s would be like 2012 if we assume the record references remain unchanged in the future.

[31] The temporal-averaged spatial distribution H/L ratio shows large values in far-western and northeastern China. Some large ratios are present in central northern China and along the south coast as observed. Ratios are greater than 1 everywhere, even for the low emission RCP2.6 scenario. The largest ratios of greater than 6 are over the southern China and parts of northwestern China (Figures 10d and 10f).

5.2 Record-Breaking Temperature Amplitude

[32] The CCMS4 simulated RBT strength slope in the historical run captured the observed structure fairly well, including larger values in northern-central China and smaller values in southern Xingjiang and along the southern boundary (Figure 11a). For LoMin, larger slope values mostly locate in the interior China (Figure 11b), resembling the observed (Figure 7). The slope for HiMax in the low emission RCP2.6 is larger (>1.6) over most interior regions. LoMin slope in southern China is mostly smaller than historical run (Figure 11d). Interestingly, the HiMax slope in RCP8.5 is generally smaller than RCP2.6. The frequency in RCP8.5 is almost double that in RCP2.6. This seems to suggest that under future warmer RCP8.5 climate the record temperature magnitude increment would be smaller than RCP2.6 climate, which seems counterintuitive. However, because under the much warmer RCP8.5 climate the RBT occurs more frequently, the temperature magnitude increment in successive record becomes smaller compared to the longer-gapped records in RCP2.6 climate. Slower decrease in RBT strength of LoMin (smaller -β) in southeastern China in the future scenario climates is indicative of the continuation of the more nocturnal warming currently observed in the region.

Figure 11.

CCSM4 simulated the RBT strength slopes β. (a, c, and e) For HiMax and (b, d, and f) for LoMin. Top row: historical; middle row: RCP2.6; and bottom row: RCP8.5.

[33] The histogram of slopes simulated by the model largely reproduced the observation with slightly underpredicted slope mode for LoMin (Figures 12a and 12b). The spectra of RCP scenario climates are narrower than past (simulated) climate with LoMin slopes apparently larger than HiMax (Figures 12c–12d).

Figure 12.

Annual histogram of the slope of RBT strength vs. the number of record k. Plotted are the proportionality coefficient β in equation ((3)) averaged over all stations. (a, c, and e) HiMax. (b, d, and f) LoMin. Top: historical; middle: RCP2.6; and bottom: RCP8.5.

6 Conclusions and Discussions

[34] In the past decades climatic warming has been evident, most likely attributable to anthropogenic sources. One of the climate change characteristics is more extreme temperatures. A wide variety of temperature metrics including percentile, threshold, absolute, and duration have been used [Alexander et al., 2006]. However, the record high maximum and low maximum temperatures, especially their ratio (H/L) have been seldom studied. This paper analyzed the temporal-spatial variation of high and low record temperatures in China during the past and future five decades. The observed H/L ratio increased from less than 1.0 before the 1970s to 3.0 for the decade (1995–2004) nationally overall, with 2.9 in summer and 3.2 in winter (Table 1). The H/L ratio in China during the most recent years is above 3, while it is about only 2 in both the U.S. [Meehl et al., 2009] and Europe [Wergen and Krug, 2010]. The relatively smaller ratio in the U.S. was likely attributed to the strong “warming hole” over a large area over the southeastern and central U.S. As for Europe, the lower ratio could be associated with less warming in general. The CCSM4 simulated an annual ratio of 2.8 (4.5 for summer and 2.2 for winter). The annual ratio under the low emission RCP2.6 scenario in 2040s would be 1.8, quite smaller than both the actually observed and model-simulated ratios in the 2000s. However, under the high emission RCP8.5 scenario, the annual ratio would be 6.8, much larger than current model simulated or observed.

Table 1. Ratio of High to Low Record-Breaking Temperature Ratios and Number of Record High Maximum Temperature Days per Year During the Last 10 Years of the Period (1995–2004 for Observation and Historical Run, and 2040–2049 for Future Scenarios)
 H/L Record RatioNumber of HiMax Day
 AnnualSummerWinterAnnualSummerWinter
  1. The expected values in parentheses are based equation ((2)).
Expected1.0(2.0)1.0(1.7)1.0(3.4)9.3(11.0)2.3(2.5)2.3(3.1)
Observed3.02.93.214.93.73.4
Modeled2.84.52.214.84.42.9
RCP2.61.82.11.412.33.52.7
RCP8.56.810.05.121.26.54.8

[35] While the observed record behaviors deviate from expected 1/n curve for an i.i.d. sequence, especially for the low minimum temperature, they fit the linear mean-shifting i.i.d. model much better. This means that the more frequent record-breaking temperature in the past decades can be explained by the warming trend of mean temperature. The remaining smaller deviation from the mean-drifting model implies that either the past warming trends were nonlinear or temperature became more extreme indeed.

[36] Although future Hi/Lo ratio will considerably increase in the RCP8.5 climate, the strength slope remain unchanged. This may be due, in part, to more frequent occurrence of record during shorter time intervals. The spatial distribution of RBT frequency resembles that of daily temperature trend well (Figures 3 and 7), suggesting that the RBT frequency is a reasonable measure of the climate warming trend in China. It seems counter-intuitive that it is the RBT frequency, not the magnitude, that matches the temperature trend. However, this may be partly reconciled by noting that frequency increase itself reflects the progressive increasing of temperature with each occurrence of the RBT.

[37] The observed RBT strength slope is smaller for HiMax than LoMin similarly to DTR. The more heterogeneity in summer is likely associated with small-scale convectiveness during daytime. The trend showed the larger warming in northern China and less warming, even scattered cooling, in southern China. The horizontal distribution of record frequency resembles that of temperature trend, including both large positive values in northern China and moderate negative values in the parts of southwest China. On the other hand, the distribution of record temperature strength increase with increasing record number differs notably from frequency and trend distributions. The large slopes mainly concentrated in the central portion of the country, suggesting that the warming in northern China is largely due to frequency increase. On the other hand, in south-central China, the strength of record temperature increased more. The RBT frequency is highly correlated to WPW in a similar fashion to the temperature trend itself with positive correction in the northern and southeastern China.

Acknowledgments

[38] We acknowledge the modeling groups, the Program for Climate Model Diagnosis and Intercomparison (PCMDI) and the WCRP's Working Group on Coupled Modeling (WGCM) for their roles in making available the WCRP CMIP5 multimodel data set. Support of this data set is provided by the Office of Science, U.S. Department of Energy. Z. Pan acknowledges the support of NOAA Climate Program Office “Modeling, Analysis, Predictions and Projections” (MAPP) Program under grant NA11OAR4310094 and WKC Foundation of Chinese Academy of Sciences. This study was also partly supported by the National Program on Key Basic Research Project of China (973) under grants 2010CB428502 and 2012CB417203, China Meteorological Administration under grant GYHY201006024, National Natural Science Foundation of China under grant 41275022, the CAS Strategic Priority Research Program grant XDA05110101, and the Priority Academic Development of Jiangsu Higher Education Institutions (PAPD). Authors are thankful to the three anonymous reviewers whose insightful and constructive comments strengthened this paper.