Have human activities changed the frequencies of absolute extreme temperatures in eastern China?

We use a newly homogenized daily T_max and T_min dataset observed at 753 Chinese meteorological stations for 1960–2012 (figure S1, available at stacks.iop.org/ERL/13/014012/mmedia). The temperature observations we use have been quality-controlled and adjusted for most non-climatic biases due to the changes in the local observing system, such as station relocation (Li and Yan 2009, Li et al 2016).

Since the horizontal resolutions of climate models are in the range of 1°−3°, we divide the mainland of eastern China into 2° × 2° resolution grid boxes and construct a regional gridded temperature dataset using available observations within each grid box. Specifically, we first calculate the climatological mean annual cycle (base period: 1960–2012) and daily temperature anomalies at each station. Given that temperature is dependent on elevation, for the boxes where topography has a wide range and stations are unevenly distributed, there might be certain deviation in the extreme temperatures if the gridded temperature is developed by simple averaging of the individual station within each grid box. Hence, we need to correct the elevation-related bias in the temperature mean state within each grid box. Considering the lapse rate of near-surface air temperature is time-varying and region-dependent, use of a fixed temperature lapse rate could be problematic on the complex terrains in China. Following Li et al (2013), we divide the whole mainland China into 24 sub-regions (figure S1). We use the multiple linear regression method including the effects of latitude, longitude and elevation to estimate the lapse rates of T_max and T_min for each sub-region and each month (figure S2). The terrain-based global 0.25° × 0.25° land elevation and ocean depth dataset (TBASE) (http://research.jisao.washington.edu/data_sets/elevation/) is applied to estimate the averaged elevation within each grid box (figure S3). The local elevation bias in climatological mean annual cycle of the individual station is adjusted based on the spatiotemporal-varying temperature lapse rates. The final gridded dataset is obtained by adding the station average temperature anomalies to the station average elevation-bias-corrected climatological mean annual cycle for each grid box. Furthermore, to estimate the regional averages precisely, we establish a set of areal weights of land fraction by considering their latitude-dependent feature and the influence of coasts and islands (figure S4).

We use the CMIP5 simulations to estimate the responses of extreme temperatures to external forcings and the internal climate variability. Table 1 lists all the available CMIP5 models used in this study. All the experiments with specific forcings have three or more members and produce daily outputs. We first evaluate the skill of climate models with ALL forcing in simulating the climatological mean of T_max and T_min. As shown in figures S5 and S6, climate models tend to perform better over eastern China than western China. There are two explanations for this discrepancy: (1) the station density in western China is much lower than eastern China (figure S1); (2) the topography in western China is much more complex than eastern China, which is poorly captured in models with resolutions of around 1–3 degrees (figure S3). The gridded temperature values can be affected seriously by individual station with local effects. We focus our analysis on eastern China (east of 105 °E) also because the majority of China's people live in the eastern part of the country.

We calculate the time series of simulated regional mean frequency of extreme temperatures in eastern China, and compare them with the observed ones (figure S7). Results illustrate a good consistency between the observed and simulated frequencies of summer days and tropical nights in eastern China, though models tend to overestimate the frequencies of icy days and frosty nights in eastern China by on average 33.6% (14.1 days) and 13.8% (14.4 days), respectively. However, in western China, the spread of simulated frequencies of extreme temperatures is very large (figure S8). It implies that CMIP5 models can hardly capture the mean state and variability of surface air temperature in western China. Based on these evaluations, we focus on eastern China in this study.

We use 32 simulations from five models driven by combined anthropogenic and natural forcing (ALL); 23 simulations from five models driven by natural forcing only (NAT) and greenhouse gas forcing only (GHG) (table 1). All simulations end in 2012. More recent years are not included in this study, as most of the model simulations required for the detection analyses ended in 2012. It is assumed that the temperature extreme responses to historical anthropogenic (ANT) and NAT forcings are linearly additive and the difference between the ALL and NAT responses can be estimated as an ANT response. Annual anomalies, with respect to 1960–2012, are computed from the resulting regional average frequency of extreme temperatures from observations and individual model runs, using the same sets of space data masks and areal weights. We compute the ensemble means for individual models and then average the ensemble means to give the expected multi-model response to large-scale external forcings. Thus, the patterns we consider are the annual anomalies of the frequency of extreme temperatures.

Zoom In Zoom Out Reset image size

We use an optimal fingerprinting method in which observations (y) are expressed as a sum of scaled model-simulated fingerprint patterns (X) plus internal climate variability (ε) as y = Xβ + ε. The scaling factors β adjust the magnitude of the fingerprints to best match the observations. The multi-model ensemble averages of forced (ALL, GHG and NAT) simulations are used to estimate the fingerprints, and the pre-industrial control (CTL) simulations are used to estimate internal climate variability. The regression is fitted based on equation (4) in Allen and Tett (1999): ${\rm{\tilde \beta }} = {\left( {{{\rm{X}}^{\rm{T}}}{\rm{C}}_{\rm{N}}^{ -1}{\rm{X}}} \right)^{ -1}}{{\rm{X}}^{\rm{T}}}{\rm{C}}_{\rm{N}}^{ -1}{\rm{Y}}$. We compute non-overlapping 3-year-mean time series of the multi-model-simulated regional mean frequency of extreme temperatures as the forced response or signal for the specific forcing (X), which includes 18 data values for the period 1960–2012. Observations are processed in the same way as the simulations. Fitting and testing the regression models need two independent estimates of the inversed covariance structure of internal climate variability (${\rm{C}}_{\rm{N}}^{ -1}$). We use the CTL simulations and the inter-ensemble difference from forced simulations to estimate them. Time series from CTL simulations are divided into 60 non-overlapping 53 year chunks and similarly masked to be in accord with observations in space. Additional 79 non-overlapping 53 year chunks are constructed using inter-ensemble differences from forced simulations (ALL: 33; GHG: 23; NAT: 23). We separate each set of chunks from CTL simulations or forced simulations into two groups sequentially. The first group of chunks is used to pre-whiten the data and the second group is used for the uncertainty analysis on the estimation of scaling factors (${\rm{\tilde \beta }}$). Instead of decreasing the dimension via a projection on the first k leading empirical orthogonal functions, we use a regularized estimate of the covariance matrix of the internal climate variability (Ribes et al 2009). Regularized estimation of the covariance matrix can avoid the underestimation of the lowest eigenvalues that occurs in original covariance matrix and ensure the covariance matrix is full rank (Ribes et al 2013). We apply equation (19) provided by Allen and Tett (1999) to conduct residual consistency checks to detect model inadequacy. The results show that all the regression models can pass this test, which means that climate models are able to simulate the internal variability of the frequency of extreme temperatures in eastern China reasonably well.

Based on equations (6) and (7) in Allen and Tett (1999), we estimate the variance-covariance matrices of the internal variability noise by using the first set of non-overlapping 53 year chucks. We obtain the 5%–95% uncertainty range of scaling factors by assuming that the internal variability noise is normally distributed. To estimate the probability distribution functions of the contributions from different forcing agents, we generate random samples of 10 000 000 values from the normal distribution of estimated scaling factors and multiply the forced trends in different signals by these random numbers.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Figure 1 shows the spatial distributions of observed trends in the frequencies of the four extreme temperature indices. Summer days have increased significantly over the northeastern China (120–135°E, 40–55°N; +2.67 days decade⁻¹) and the middle and lower reaches of the Yangtze River (110–125°E, 28–32°N; +2.99 days decade⁻¹). The occurrences of tropical nights increased mainly over the Yangtze–Huaihe River basin (115–125°E, 28–34°N; +2.62 days decade⁻¹) and part of southern China (105–115°E, 18–24°N; +3.91 days decade⁻¹). Significant declining trends in icy days (−2.24 days decade⁻¹) and frosty nights (−3.35 days decade⁻¹) are found in the northwest of north China (105–115°E, 35–42°N). Frosty nights have also decreased significantly over northeastern China (−3.52 days decade⁻¹) and the Yangtze–Huaihe River basin (−4.22 days decade⁻¹). Figure 2 displays the time evolution of the observed and simulated frequency anomalies of the four indices in eastern China. The observed changes in extreme temperatures keep pace with the multi-model-simulated responses to ALL forcing, but not with the simulated responses to NAT forcing. We first apply the optimal fingerprinting method (Allen and Tett 1999) to scale the modeled time series of extreme temperatures in eastern China with ALL forcing to best fit the observations. As shown in figure 3, one-signal analysis suggests that climate models with ALL forcing can reproduce the observed frequency changes in summer days and icy days well, and have scaling factor estimates consistent with the value one though bear certain internal variability. However, climate models tend to overestimate (underestimate) the frequency change in tropical nights (frosty nights). This implies that models perform better in simulating the observed frequency change in daytime extreme temperatures than nighttime extremes. Though focusing on percentile-based extreme temperatures, Lu et al (2016) also found that climate models with ALL forcing do a better job in reproducing the frequency changes in daytime extreme indices than nighttime indices. It may be associated with the model's deficiency in reproducing the seasonality of warming trends in T_min in eastern China. Lewis and Karoly (2013) found that the T_min trends are noticeably subdued by the CMIP5 models, particularly in the boreal winter, when shallow boundary layers and soil freeze/thaw cycles are likely difficult to be simulated realistically. On the other hand, direct visual inspection of figure 3 illustrates that the uncertainty ranges in the scaling factors for cold extremes are larger than warm extremes, which implies smaller variability in the frequency of simulated cold extremes than that of observed ones. Other studies also found similar result existing in the changes in the annual maxima and minimum of daily temperatures (Morak et al 2013, Wen et al 2013, Yin et al 2017). A possible cause for this is that the strong internal variability of winter extreme temperatures in eastern China was underestimated by the CMIP5 climate models (figure 2 and figure S7). Increased GHGs enhance downward longwave radiation and hence increase the surface air temperature and change the frequency of temperature extremes. Meanwhile, the increased water vapor in warmer atmosphere can further increase downward longwave radiation. However, other anthropogenic forcing agents (e.g. aerosols) can decrease daytime temperature and change the frequency of daytime extremes directly by obstructing downward solar radiation and indirectly by changing the properties of clouds. Natural forcing agents, such as solar variability and volcanic eruptions, may also lead to the variations of surface air temperature and change the frequency of extreme temperatures by modulating solar radiation at the surface and the interaction between aerosols and clouds. The respective roles of anthropogenic and natural forcings in the change of extreme temperatures remain to be elucidated.

To detect the effects of ANT and NAT forcings in the same framework, we conduct two-signal detection analysis. As shown in figure 4, the 5%–95% uncertainty ranges of ANT scaling factors for the four indices do not include zero and the 90% confidence ellipse regions do not cover the origin of x-y coordinates. This indicates that the effect of ANT forcings can be clearly detected, and the climate responses of ANT and NAT forcings can be well separated from each other. In other words, the influence of human activities is detectable in the frequency changes of these four temperature extreme indices. Except for summer days and frosty nights, the 5%–95% uncertainty ranges of NAT scaling factors for other two indices include zero, suggesting that the effects of NAT forcings on their frequency changes are undetectable.

Zoom In Zoom Out Reset image size

To examine the influences of individual groups of anthropogenic forcing agents, we conduct three-signal analysis to scale the model responses of GHG, OANT (ALL minus the sum of GHG and NAT) and NAT for the optimal agreement with observed frequency changes in extreme temperatures. As shown in figure 5, the results reveal that the effects of anthropogenic increase in GHGs can be clearly detected in the frequency changes of these four indices. Models appear to underestimate the effects of GHGs on the changes in icy days and frosty nights by a factor close to two. It is inferred that the model's deficiency in the effects of GHGs on the changes in cold-season extreme temperatures is associated with an underestimation of GHG-forced temperature changes in the cold season in eastern China. Morak et al (2013) found that the HadGEM1 model significantly underestimates the changes in extreme temperatures in winter across large parts of Asia. Chen and Frauenfeld (2014) found that the winter warming in the CMIP5 models is only about half (one-fourth) of the observed warming in China for the period of 1901–1999 (1950–1999). The effects of OANT are also detectable, but with larger uncertainty. For all extremes indices, OANT effects are underestimated by the models. This may be due to the omission or simplification of the indirect effects of anthropogenic aerosols in some climate models, such as CanESM2 and IPSL-CM5A-LR (Hu et al 2014). Except for summer days, the influence of NAT forcings on other indices cannot be detected. These analyses demonstrate that the human-induced rise in greenhouse gas has imposed a detectable impact on the frequency change in extreme temperatures over eastern China.

Based on the estimate results of three-signal analyses, we quantify contributions to the frequency changes of extreme temperatures to individual factors through multiplying the simulated trends in GHG, OANT and NAT signals by the respective scaling factors. As shown in figure 6, we find that the observed frequency changes in extreme temperatures are the net result of the counteracting effects from GHG and OANT forcing agents, since NAT forcing imposes little influence on these changes. Among three individual components of ALL forcings, the effect of anthropogenic emission of GHGs is dominant and has changed the frequencies of summer days (tropical nights, icy days, frosty nights) by rates of +3.48 ± 1.45 (+2.99 ± 1.35, −2.52 ± 1.28, −4.11 ± 1.48) days decade⁻¹. Other anthropogenic forcing agents (dominated by anthropogenic aerosols) partly offset the effect from GHGs and changed the frequencies of these four indices by −1.53 ± 0.78, −1.49 ± 0.94, +1.84 ± 1.07, +1.45 ± 1.26 days decade⁻¹, respectively.

Zoom In Zoom Out Reset image size

To further evaluate the robustness of above results, we repeat these analyses based on the 5-year-mean series. As shown in figure S9, results from two-signal detection analyses are generally in line with those with 3-year-mean series. The influence of human activities can be clearly detected in the observed frequency changes of the four extreme indices. However, the effects of NAT forcing can no longer be detected in the change in summer days and frosty nights. Three-signal detection analyses based on 5-year-mean series also indicate that ANT influences (GHG and OANT) are detectable in the frequency changes of extreme temperatures (figure S10). And the influence of natural forcings cannot be robustly detected in any indices. All the detection analyses suggest that anthropogenic influences are responsible for the observed frequency changes of these four temperature extreme indices.