Changes in temperature extremes on the Tibetan Plateau and their attribution

The study focuses on the central and eastern TP because of the sparse station density on the western TP. We call the area the TP region hereafter. The observational data include daily maximum and minimum temperatures from 157 stations (figure 1) on the central and eastern TP from the China National Meteorological Information Center (NMIC) (available at http://data.cma.cn/). These data cover the period from 1951 to 2017, and have been controlled for quality and adjusted for homogeneity by Xu et al (2013). We use only the data after 1958 because of good quality and station density in this period.

Zoom In Zoom Out Reset image size

Based on these daily data, twelve extreme temperature indices (table 1) are calculated in each station following the ETCCDI definition (Zhang et al 2011) and are termed as ChinaDEX in this study. The anomalies of all the indices are calculated at each station relative to the 1961–1990 mean and then averaged onto a 2.5° × 3.75° latitude–longitude grid, which is consistent with the grid resolution for HadEX2 developed by Donat et al (2013) (https://climdex.org/gewocs.html). The extreme indices include 4 absolute intensity indices (TXx, TNx, TXn, and TNn), 4 percentile-based frequency indices (TX90p, TN90p, TX10p, and TN10p) and 3 fixed threshold exceedance indices (SU, FD, and ID) and the DTR. Among these indices, the absolute intensity indices represent the most extreme events and the percentile-based indices reflect the changes in the moderate extremes. The fixed threshold exceedances indices and DTR have substantial impacts on society, ecosystems and agriculture (Terando et al 2012, Cheng et al 2014). The 'number of tropical nights (TR)' is excluded because night-time temperatures hardly exceed 20 °C on the TP. The TR index includes too many small or zero values and is not reasonably covered for the TP region.

To test the robustness of different observational data for estimating climate change on the TP, the HadEX2 gridded extreme indices data during 1958–2010 are also used. There are 47 stations on the TP for HadEX2 (figure 1), which is much less than the ChinaDEX (157 stations). We present a comparison between ChinaDEX and HadEX2 during 1958–2010 in figure 4. In the other parts of our study, we will use ChinaDEX from 157 stations because the data have better spatial and temporal coverage than HadEX2. The trends of all the extreme indices on the TP during 1958–2017 are calculated using a nonparametric method (Sen 1968) and compared with those in the whole of China and East China.

The CMIP5 model simulations are used to estimate extreme temperature responses to external forcing and the role of natural internal variability (Taylor et al 2012). Sillmann et al (2013a, 2013b) calculated the ETCCDI indices based on daily data from the CMIP5 simulations. Here, we use the multi-model ensemble mean of 18 models from 82 historical runs driven by combined anthropogenic and natural forcing (ALL) and 6 models from 25 historical simulations driven by natural forcing only (NAT) (table 2). Because some CMIP5 historical simulations end in 2005, the RCP85 simulations for the years 2006–2017 are used to extend the data after 2005. The NAT simulations end in 2012 and there are no NAT simulations for subsequent years. To detect the influence of anthropogenic (ANT) forcing on recent climate change, we replicate the NAT simulation from 2008 to 2012 for the period of 2013–2017. This method should be reasonable because few volcanic eruptions in this period, and the error introduced by having the wrong solar forcing during 2013–2017 is expected to be small, too. Using 2005–2017 from RCP85 means that more than 20% of the time period is considered to have no NAT. We thus conduct a series of sensitivity tests to verify the robustness of detection results. We also use preindustrial control (CTL) simulations with no external forcing from 28 models to estimate the internal variability, in which only 7 models are available for DTR indices (table 2). All the models are interpolated onto the same 2.5° × 3.75° grid as the observations and then masked with the gridded observed temperature indices to mimic the availability of observational data in each grid. We compute the ensemble mean of all available models after calculating the ensemble means of all the runs from individual models. Then, regional averages are calculated based on the masked data.

We use an optimal fingerprinting method based on generalized linear regression (Allen and Stott 2003, Ribes and Terray 2013) to quantify the influence of external forcings on observed climate change. The method expresses the observed value change (Y) as a function of the climate responses to the external forcings (X), which is expressed as: Y = (X − v) β + u. Here, u accounts for the regression residual and reflects the effect of internal variability. The scaling factor β is estimated using the total squares method (Allen and Stott 2003, Ribes and Terray 2013). v represents the noise in the signal that depends on the model ensemble size because the multi-model average may not be able to remove all vestiges of internal variability. The residual consistency test (Ribes and Terray 2013) is used to evaluate whether the model simulated variability is inconsistent with the regression residuals. In total we extract the 285 chunks of 60 year long segments from half of the CTL simulations and within-ensemble differences (i.e. the residuals of the ensemble simulations after the ensemble mean is removed) to estimate the scaling factor β, and the 90% confidence interval. Scaling factors whose 5th percentile is greater than zero imply a detectable signal at the 5% significance level. Moreover, we consider scaling factors with 90% ranges including unity to indicate good agreement between observations and model simulations.

Our detection analyses include single-signal and two-signal analysis to detect the relative influences of the different forcings on the observed changes. The single-signal analysis yields a regression between observations and multi-model mean responses of a single factor (ALL, ANT, and NAT) to estimate whether the observed changes are agree with individual external forcing or combined forcings effect. We estimate anthropogenic forcing as the difference between ALL and NAT responses from all available simulations. In the two-signal analysis, we conduct a regression of the observations onto ALL and NAT responses simultaneously, and then perform a linear transformation to estimate the relative contributions of ANT and NAT signals to the observations and to determine whether these two signals can be detected separately.

Based on the ChinaDEX data, the observed trends of twelve temperature extreme indices during 1958–2017 are illustrated in figure 2. The increasing warm extremes and decreasing cold extremes are consistent with global warming. The intensity (TXx and TNx) and frequency (TX90p and TN90p) of the warm extremes have increased and the corresponding indices (TXn, TNn, TX10p and TN10p) for the cold extremes have decreased. The changes in the intensity of the coldest extremes (TXn and TNn) are greater than those in the intensity of the warmest extremes (TXx and TNx). For the fixed threshold exceedances indices, FD and ID have been decreasing, and SU increasing. The DTR changes on the TP have obviously decreased mainly due to the rapid increase in the nighttime extremes. If one compares changes in the daytime and nighttime extremes, changes in the nighttime frequencies (TN10p and TN90p) are larger than those in the daytime frequencies (TX10p and TX90p), indicating larger changes in the nighttime temperature. Previous studies have shown similar changes for these indices in China (e.g. Dong et al 2018).

Zoom In Zoom Out Reset image size

Figure 3 provides a comparison of trends in temperature extremes between the TP and China as a whole and East China (all of China that is east of 105°E). Except for TXn and SU, the magnitudes of the linear trends of all the other indices on the TP are larger than those in China and in East China. This result indicates that the warming of temperature extremes on the TP is more than in the other regions. More intense, more frequent warm extremes and less intense, less frequent cold extremes have occurred. The FD and ID have decreased more than in the other regions in China. The possible reasons are related to the changes in surface albedo feedback, surface radiation fluxes or amounts of clouds above the TP (Rangwala et al 2013, Pepin et al 2015), but the exact reason is still unclear. Additionally, we find that there are larger differences in SU, FD, and ID between the TP and other regions than in the other indices. The daily maximum and minimum temperatures more easily exceed 0 °C (fixed threshold for FD and ID), while for SU, the daily maximum temperature hardly exceeds 25 °C in the center of the TP. These variations may lead to large differences in the SU, FD, and ID indices.

Zoom In Zoom Out Reset image size

We also compare the changes in temperature extremes averaged on the TP from two observational datasets ChinaDEX and HadEX2 (figure 4). Note that HadEX2 covers only the period from 1958 to 2010. Generally, these two datasets correspond well for frequency indices, while the differences between the two datasets are obvious for other indices. We find that HadEX2 has some anomalous values in the southwestern part of the TP (black slash grid boxes in figure 1). When we include these grid boxes in regional averages, the HadEX2 time series show large changes in several indices, including intensity indices, SU and FD. If we omit these data from our regional average, the HadEX2 values show very good consistency with ChinaDEX (figure 4). The southwestern part of the TP has only 5 stations for HadEX2. We suspect that this large difference is due to very sparse station density and the related nonclimatic inhomogeneities in HadEX2 (Dunn et al 2014). For SU and FD indices there are some very high correlation among stations and very low variances. Correlations are not calculated if no data are present, so that the coverage can look different to other indices (Dunn et al 2014). These results indicate that the estimates of climate changes on the TP are affected by the data coverage in space and time and that different datasets may provide very different results. This possibility raises a caveat about the uncertainty of different observational datasets in analysing climate change in the data-sparse region.

Zoom In Zoom Out Reset image size

The model simulated linear trends during 1958–2017 for all the indices are shown in figures 5 and 6 based on multi-model ensemble mean from the ALL and NAT forcing experiments, respectively. For ALL forcing experiments (figure 5), the model-simulated trends reproduce the observed changes well. The absolute intensity and frequency of warm extremes have increased, while cold extremes have become less intense and less frequent. However, the models generally underestimate the observed amplitudes of trends (figure 2) for most indices, especially for cold extremes. The model-simulated spatial patterns are more uniform than the observations, indicating small variability and slight regional differences from the model-ensemble mean. On the other hand, the NAT patterns (figure 6) show slightly increasing trends for the warm extremes and decreasing trends for the cold extremes, which is consistent with the observations but with much smaller magnitude. It suggests a positive effect of NAT forcing on the changes in temperature extremes in the region.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The 5 year mean series of observational and model-simulated extreme temperature indices are shown in figure 7. The model-simulated long-term changes in these temperature extremes under ALL forcing are similar to the observations. For most of the indices, however, the simulated changes in absolute intensity, frequency and the fixed threshold exceedance indices are smaller than the observations, especially for the cold extremes, which are consistent with figures 2, 5, 6. This underestimation by model simulations may imply that the projected future changes in temperature extremes on the TP based on these multi-model ensembles are underestimated. The 90% ranges of ALL forcing runs (pink shading in figure 7) almost fully cover the observational changes, suggesting good performance of the models in simulating these indices. The observations are more similar to the simulated response to ANT forcing than that to NAT forcing alone, but the NAT simulations show small and positive contributions to the observed changes.

Zoom In Zoom Out Reset image size

We investigate the effects of external forcings on the observed spatiotemporal evolution of temperature extremes during 1958–2017 using an optimal fingerprinting method. Figure 8 shows scaling factors and their 90% confidence intervals for ALL, ANT and NAT forcings based on single-signal detection analyses. For all indices, both ALL and ANT signals are robustly detected on the TP. The confidence intervals of the scaling factors are generally smaller than those of the detection results in China (Lu et al 2016, Yin et al 2017, Yin and Sun 2018). The residual consistency tests are generally passed, suggesting a good consistency between observed and model-simulated variability, and the models do not miss any important forcings and/or responses to them. Without having analyses on the response to greenhouse gas and other anthropogenic forcing agents, in particular aerosols, in isolation, we speculate that aerosol forcing might be less important here than elsewhere so that the large uncertainties associated with it are less relevant here, a hypothesis that would need further research to scrutinize.

Zoom In Zoom Out Reset image size

For the intensity indices, the best estimates of scaling factors are close to unity for warm extremes (TXx and TNx), while they are greater than unity for cold extremes (TXn and TNn). These results indicate a good agreement between the observations and the forced responses of the models for the intensity of warm extremes and underestimation by model simulations for the intensity of cold extremes. These results are similar to the detection results of the intensity indices in other regions (Zwiers et al 2011, Kim et al 2016, Yin et al 2017, Dong et al 2018).

For the frequency and fixed threshold indices, both ALL and ANT signals can be clearly detected. The best estimates of the scaling factors are very close to or slightly larger than unity, indicating good agreement between observations and models or underestimation of the changes by the model simulations. The confidence intervals for the frequency indices are very small, indicating small uncertainties of the detection results. The residual consistency tests are passed except for TN10p, SU and DTR. For these three indices, the models underestimate the observed variability for TN10p and overestimate the observed variability for the SU and DTR, respectively. The smaller variability in the models implies that the models miss some physical process responsible for these kinds of extremes or they misinterpret some of the internal variability as a forced response. The detection results for ANT are similar to those for ALL but have larger confidence intervals. The NAT signal is detected for most extreme indices on the TP, although the 90% confidence intervals for the scaling factor have larger uncertainty ranges.

The scaling factors and their confidence intervals based on two-signal regression analyses are shown in figure 9. The ANT signal can be detected in all the indices while the NAT signal can be detected for most indices during 1958–2017. For the intensity indices, both the ANT and NAT signals are detected separately for warm extremes, but the NAT signal cannot be detected for cold extremes. This difference suggests a dominant influence of ANT forcing on the cold extremes. Similar to the single-signal results, the best estimates of scaling factors for ANT forcing are close to unity for the warm extremes while they are greater than unity for the cold extremes. For all the frequency and fixed threshold exceedances indices, the ANT and NAT signals are separately detected. The best estimates of the scaling factors are generally close to or greater than unity. The residual consistency tests are passed for most indices except for TN10p and DTR, in which model simulations have smaller variability for TN10p and larger variability for DTR than the observations, respectively. These results show good consistency with the single-signal results and provide clear evidence of human influence on the temperature extremes on the TP.

Zoom In Zoom Out Reset image size

The attributable contributions from different external forcings to the observed trends are shown in figure 10. The attributable trends are estimated as the linear trends in the signal from model simulations multiplied by the corresponding scaling factors. During 1958–2017, the observed changes in TXx, TNx, TXn, and TNn are 1.6 °C, 1.7 °C, 1.8 °C, and 3.5 °C, respectively. The median values for the contribution of ANT signal to these observed changes are 70%, 76%, 89%, and 95%, respectively. For the frequency indices TX90p, TN90p TX10p and TN10p, the median values from ANT contributions are 87%, 88%, 62%, and 85%, respectively. The observed changes in SU, FD, ID and DTR are 8.7 d, −26 d, −17 d and −1 °C, respectively and the median values of the ANT contribution explain 87%, 86%, 84% and 108% of these changes, respectively. NAT contributions are less than 15% for all the extreme indices except for TXx, TNx, and TX10p, in which NAT explains 30%, 24% and 38% of the changes, respectively. Note that these numbers are a lower limit but not necessarily good estimate of the contribution of natural forcing due to our unrealistic assumption of natural forcing during 2013–2017 to be same as in the period before. Compared with the results of previous studies across China (Lu et al 2016, Yin et al 2017, Yin and Sun 2018), the NAT influences on the TP are slightly larger than those in China.

Zoom In Zoom Out Reset image size

To evaluate the robustness of the detection results, we conduct a series of sensitivity tests (supplementary figure S1 and S2 is available online at stacks.iop.org/ERL/14/124015/mmedia). First, we test the sensitivity of detection and attribution results to the analysis period, since we use RCP85 data to extend the model simulations and replicate the 2013–2017 NAT results with the 2008–2012 model simulations. We redo all the analyses in different periods, including the periods from 1958 to 2005 and 1958–2012. We find that for both single- and two-signal analyses, the detection results for ANT are similar but that the detection of the NAT signal is affected by the studied time period for some indices, including the intensity and frequency of warm extremes, TN10p and fixed threshold exceedance indices. However, for the two-signal detection results during 1958–2012 from the same model subset under ALL and NAT forcing, only NAT is not detected anymore for the intensity of cold extremes, TX90p, TN10p, and ID. This justifies the robustness of our detection results. Second, we test the sensitivity of our results to different model ensembles, since we use all the available model runs for ALL and NAT experiments, and the number of runs and models for ALL and NAT do not exactly correspond. We redo all the analyses using the models that have the same ALL and NAT experiments. The detection results using signals from the same subset of models are quite similar to the results from all available model runs for most indices. However, we find that the detection for a few indices is affected by the choice of models and consequently the estimate of the naturally forced signal, especially when the studied time period is short. This outcome implies that our results are generally robust but affected by the caveats of using differing natural estimates for NAT than included in ALL, and by having wrong NAT forcing from 2005 onwards. In a longer period with more signals, the influences of ANT and NAT signals are more easily detected than those in a shorter period. With the improvement of data availability, including longer observation as well as model simulations, more studies are required to further scrutinize our results.