Advanced Search
Article Contents

Understanding the Surface Temperature Cold Bias in CMIP5 AGCMs over the Tibetan Plateau


doi: 10.1007s00376-017-6326-9

Get Citation+

Export:  

Share Article

Manuscript History

Manuscript received: 21 December 2016
Manuscript revised: 02 June 2017
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Understanding the Surface Temperature Cold Bias in CMIP5 AGCMs over the Tibetan Plateau

  • 1. State Key Laboratory of Numerical Modeling for Atmospheric Sciences and Geophysical Fluid Dynamics, Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing 100029, China
  • 2. University of Chinese Academy of Sciences, Beijing 100049, China

Abstract: The temperature biases of 28 CMIP5 AGCMs are evaluated over the Tibetan Plateau (TP) for the period 1979-2005. The results demonstrate that the majority of CMIP5 models underestimate annual and seasonal mean surface 2-m air temperatures (T as) over the TP. In addition, the ensemble of the 28 AGCMs and half of the individual models underestimate annual mean skin temperatures (T s) over the TP. The cold biases are larger in T as than in T s, and are larger over the western TP. By decomposing the T s bias using the surface energy budget equation, we investigate the contributions to the cold surface temperature bias on the TP from various factors, including the surface albedo-induced bias, surface cloud radiative forcing, clear-sky shortwave radiation, clear-sky downward longwave radiation, surface sensible heat flux, latent heat flux, and heat storage. The results show a suite of physically interlinked processes contributing to the cold surface temperature bias. Strong negative surface albedo-induced bias associated with excessive snow cover and the surface heat fluxes are highly anti-correlated, and the cancelling out of these two terms leads to a relatively weak contribution to the cold bias. Smaller surface turbulent fluxes lead to colder lower-tropospheric temperature and lower water vapor content, which in turn cause negative clear-sky downward longwave radiation and cold bias. The results suggest that improvements in the parameterization of the area of snow cover, as well as the boundary layer, and hence surface turbulent fluxes, may help to reduce the cold bias over the TP in the models.

摘要: 在28个CMIP5-AMIP试验的模式中, 大多数模式对高原气温的模拟存在着冷偏差, 多模式集合和半数以上的模式模拟的地表温度偏低, 主要特征为气温的冷偏差强于地表温度, 高原西部强于高原东部. 通过地表能量平衡分解法, 将地表温度的冷偏差定量分解为反照率反馈项、云辐射强迫项、晴空短波辐射项、晴空向下长波辐射项、感潜热项和地表热通量项. 结果表明, 这些分解项对冷偏差的贡献存在物理上相联系的过程, 积雪覆盖面积偏大引发的反照率反馈作用和晴空向下长波辐射强迫造成了地表温度模拟的冷偏差, 其物理过程是: 低温模式模拟的积雪覆盖面积偏大, 使得地表反照率增大, 地表吸收的短波辐射减少、地表感、潜热通量减少, 使得地表向大气输送的热量和水汽(尽管很小)偏少, 大气温度偏低、水汽含量减少, 导致晴空向下长波辐射减小, 地表温度偏冷. 鉴于积雪覆盖面积在模式模拟中的重要性, 有必要对积雪覆盖面积的参数化方案做出改进, 并提高地表湍流通量, 这可能会有助于减少模式模拟的地表温度冷偏差.

1. Introduction
  • The Tibetan Plateau (TP) is the highest and most extensive plateau in the world, with an area of approximately 2.5× 106 km2 and an average elevation of 4000 m. The TP affects not only the regional climate and environment in East Asia, but also global atmospheric circulation via thermal and mechanical forcing (Flohn, 1957; Yeh et al., 1957; Wu and Zhang, 1998; Duan and Wu, 2005; You et al., 2013). As a large mid-troposphere heat source during summer, the TP plays an important role in the development of weather systems over East China (Tao and Ding, 1981), as well as the onset and maintenance of the Asian summer monsoon (Wu and Zhang, 1998; Li et al., 2001; Liu et al., 2007; Wu et al., 2012). The net heating effect from the TP surface to the atmosphere is due to the net surface radiation, latent heat flux and sensible heat flux. Over the course of a year, the sensible heat transport is primary (Ye and Gao, 1979). The surface sensible heat flux is closely related to the difference between the surface temperature and air temperature, as well as the surface wind speed.

    Figure 1.  Scatter plots and correlation between T as and T s averaged over the TP in all four seasons [(a) March-May; (b) June-August; (c) September-November; (d) December-February] based on 28 CMIP5-AMIP models, their ensemble mean, and reanalysis (CFSR and ERA-Interim) data. The T as of the CRU dataset is presented by the bold purple lines.

    Studies on skin temperature (T s) and 2-m air temperature (T as) over the TP have long been largely dependent on satellite data, reanalysis data and model simulations, due to the unique geographical environment of the TP, limited observation data and unevenly distributed meteorological stations. (Frauenfeld et al., 2005) compared the T as dataset retrieved from ERA-40 with that obtained from 161 meteorological stations on the TP. They found that the reanalysis temperatures were lower than the temperatures observed at the TP stations. The temperature bias was almost exclusively due to differences between the grid cell elevations and the station elevations. (Wang and Zeng, 2012) indicated that the T as datasets of ERA-40, MERRA, NCEP-1, CFSR, ERA-Interim and GLDAS were all underestimated over the TP, and suggested that elevation differences largely accounted for the cold biases. (Duan et al., 2014) compared simulations from nine AGCMs using both observational data and five reanalysis datasets, and found that T as, T s and their difference (T s minus T as) were all underestimated. (Yang et al., 2007) attributed the surface cold bias to the absence of diurnal variations in the surface drag coefficient.

    Recently, (Chen and Frauenfeld, 2014) compared 20 GCMs from CMIP5 with observation data and found that the GCMs have substantially cold-biased over the TP, especially during the cold season. In addition, 24 GCMs from CMIP5 were assessed over the eastern TP by comparing the surface air temperature from the model outputs with observations from the China Meteorological Data Sharing Service System (http://cdc.cma.gov.cn/). The results showed a mean underestimation of 1.18°C-2.58°C for December-May, and an underestimate of less than 1°C for June-October (Su et al., 2013). Downscaling has also been used over the TP, indicating the temperature at 500 hPa has a sizable cold bias (Xu et al., 2017). Previous studies have shown that the cold bias in models is prevalent in high-elevation areas, especially the TP (Annan et al., 2005; Gao et al., 2011; Ji and Kang, 2013). (Chen and Frauenfeld, 2014) suggested that models may fail to properly simulate the snow-albedo feedback over the TP, causing significant cold bias during the cold season and a small bias during the warm season.

    Figure 2.  Annual mean T as (°C) differences between various models and CRU data averaged during 1979-2005. All air temperature values in the models have been corrected to real elevation at a resolution of 2.5°× 2.5°.

    Given that a cold bias exists over the TP in the majority of CMIP5 models, it is important to fully understand its causes and associated physical processes. The goal of this study, therefore, is to document the underlying processes of the T s cold bias towards model improvement. It has been shown that T s and T as are underestimated over the TP (Hua et al., 2014), and that the biases of T as are largely dependent on T s (Yuichiro et al., 2006; Duan et al., 2014). This is because the low-density air over the TP absorbs solar radiation very weakly and, climatologically, the surface air energy budget is mainly contributed by the underlying surface, i.e., the surface sensible heat flux, which is closely related to surface temperature. It can be seen from the scatter plots in Fig. 1 that T s and T as have a good linear relationship; indeed, the correlation coefficient is more than 0.8 in all four seasons. Therefore, to a certain extent, if the cold bias of T s can be improved, so too could that of T as. For this reason, we focus initially on the surface energy balance, rather than developing a detailed vertical profile of the surface and atmospheric temperatures. The surface energy budget method of (Lu and Cai, 2009a) is applied to evaluate the relative importance of various surface fluxes in the T s bias. Based on the surface energy balance, this method was successfully applied by (Lu and Cai, 2009a) to study the seasonality of a polar surface warming amplification via climate simulations. This study may help to explain the causes of the cold bias over the TP in CMIP5 AGCMs, and may also provide useful information for improving these models (Fig. 2).

    The remainder of the paper is organized as follows: Section 2 describes the models, data and analysis method used in this study. Section 3 uses the energy balance method to decompose the T s bias and determine the cause of the TP cold bias. A discussion is presented in section 4, followed by a summary in section 5.

2. Models, data and analysis method
  • The data used in this study are the outputs of the AMIP experiment from the CMIP5 dataset (https://pcmdi9.llnl.gov), in which the AGCMs are forced by historical SST and sea-ice data. The analysis uses the monthly mean outputs of the 27-yr (1979-2005) averaged climate data from 28 models, including sensible heat flux, latent heat flux, air temperature, surface temperature, downward shortwave (SW) radiation, upward SW radiation, clear-sky downward SW radiation, clear-sky upward SW radiation, clear-sky downward longwave (LW) radiation (DLR), and upward LW radiation data. Table 1 presents the institute, nation, modeling center (or group), model name, and atmospheric resolution associated with each model. Since the resolutions of the 28 models vary, a bilinear interpolation method is used to interpolate the model data onto a common horizontal grid of 2.5°× 2.5°.

    Because of the complex topography and severe weather conditions of the TP, most meteorological stations are located in the east-central part of the region. Plus, it is difficult to obtain long-term meteorological data series. We use the most recent dataset (CRU-TS3.10; New et al., 1999) from the Climatic Research Unit (CRU) as the "observation". The monthly dataset includes surface temperature and precipitation on a 0.5°× 0.5° latitude-longitude grid from 1901-2009 (New et al., 1999). Previous studies have found that the interannual temperature variations observed by the CRU are consistent with reconstructed proxy series from tree rings, ice cores and lake sediments in China, with the correlation between the data series reaching 0.84 (Wen et al., 2006).

    To evaluate the performance of CRU-TS3.10 data over the TP, (Wang et al., 2013) compared monthly mean CRU data for the region (25°-42°N, 70-105°E) with monthly mean data from 111 Chinese standard meteorological stations at elevations ≥ 2000 m from 1982-2001. A high interannual correlation (0.99) was found between the annual CRU and station temperature data, indicating the former can reflect the climate variability over the TP effectively.

    The reanalysis datasets used are ERA-Interim (Dee et al., 2011) and CFSR (Saha et al., 2010). ERA-Interim is a global atmospheric reanalysis dataset covering the period from 1979 to the present. It has a horizontal resolution of 1.5°× 1.5°, with 37 pressure levels ranging from 1000 hPa to 1 hPa. CFSR is a global high-resolution reanalysis product released in 2010 by the NCEP. The CFSR data used in this study include monthly mean data at a horizontal resolution of 0.5°× 0.5° from 1979-2005. (Wang and Zeng, 2012) compared the performances of six reanalysis products (MERRA, NCEP-1, CFSR, ERA-40, ERA-Interim and GLDAS) over the TP. The results showed that, compared with measurements from 63 China Meteorological Administration stations, ERA-Interim provides the best overall performance based on both daily and monthly air temperatures. Meanwhile, compared with measurements from nine Coordinated Enhanced Observing Period Asia-Australia Monsoon Project stations, CFSR provides the best overall performance.

  • Following (Lu and Cai, 2009a), the surface energy balance equation is as follows: \begin{eqnarray} Q&=&S^\downarrow-S^\uparrow+F^\downarrow-F^\uparrow-H-{\rm LE}\nonumber\\ &=&(1-\alpha)S^\downarrow+F^\downarrow-F^\uparrow-H-{\rm LE} , \ \ (1)\end{eqnarray} where Q represents the heat storage term for the land surface; S and S represent the surface downward and upward SW radiation, respectively; α is the surface albedo; F and F represent the surface downward and upward LW radiation, respectively, in which Fࣈ T s4; The term σ is the Stefan-Boltzmann constant, σ=5.67× 10-8 W m-2 K-4. T s is surface temperature, H and LE represent the sensible and latent heat flux, respectively. Here, we consider the difference () between the models with cold surface bias over the TP (cTP) and warm surface bias over the TP (wTP). By rearranging Eq. (1), we obtain the following: \begin{equation} \Delta F^\uparrow=\Delta[(1-\alpha)S^\downarrow]+\Delta F^\downarrow-\Delta Q-\Delta(H+{\rm LE}) . \ \ (2)\end{equation} We also introduce surface cloud radiative forcing (CRF), which is the difference between the net total-sky radiation and clear-sky radiation at the surface. This parameter has long been used to represent the effects of clouds on climate (Cess and Potter, 1988; Ramanathan et al., 1989). Because the difference of CRF includes the impacts of surface albedo-induced difference (SAF) and other feedback processes, (Lu and Cai, 2009a) excluded the SAF in the surface CRF difference \((\Delta\rm CRF_\rm s)\): \begin{eqnarray} \Delta{\rm CRF}_{\rm s}&=&\Delta[(1-\bar{\alpha})S_{\rm cld}^\downarrow+F_{\rm cld}^\downarrow-F_{\rm cld}^\uparrow]\nonumber\\ &=&(1-\bar{\alpha})\Delta S_{\rm cld}^\downarrow+\Delta F_{\rm cld}^\downarrow ,\ \ (3) \end{eqnarray} where ( A) cld=(A)-( A) clr, in which (A) represents radiation under total-sky conditions, ( A) clr represents radiation under clear-sky conditions, and \(\bar\alpha\) is the surface albedo of the unperturbed mean climate state. In Eq. (3), we utilize the fact that ∆ F cld↑=0 at the surface, because the mathematical expression for the radiation emitted from the surface is not directly related to clouds. According to Eq. (3), Eq. (2) can be rewritten as follows: \begin{eqnarray} 4\sigma\bar{T}_{\rm s}^3\Delta T\approx\Delta F_{\rm s}^\uparrow&=&-(\Delta\alpha)(\bar{S}^\downarrow+\Delta S^\downarrow) +\Delta{\rm CRF}_{\rm s}+(1-\bar{\alpha})\Delta S_{\rm cld}^\downarrow\nonumber\\ &&+\Delta F_{\rm cld}^\downarrow-\Delta Q-\Delta(H+{\rm LE}) . \ \ (4)\end{eqnarray} In Eq. (4), the overbar represents the reference mean climate state. The terms on the right-hand side of Eq. (4) represent the SAF, the differences in surface CRF, the non-SAF differences in clear-sky SW radiation (due mainly to atmospheric water vapor), the difference in DLR fluxes respectively, the difference in heat storage, and the difference in surface heat fluxes. Additional details are provided by (Lu and Cai, 2009a).

    Because of the resolution and elevation deviation between the models and the CRU data, utilizing the direct temperature difference between the models and CRU data will result in serious errors. Previous studies have also found that the elevation differences of different models can affect temperature simulation. We correct the simulated temperature to the actual elevation using the method of (Sheffield et al., 2006), reducing the error caused by the different model resolutions. First, the model temperatures are corrected to those at sea level based on a temperature lapse rate (6.5°C km-1). Then, all the model temperatures are interpolated to a resolution of 2.5°× 2.5°. The model temperatures are then corrected from sea level to the actual elevation.

    The inter-model empirical orthogonal function (EOF) used in this study is similar to the traditional EOF method (Jackson, 1991), but includes the continuous model-space matrix X(m,s) instead of the traditional time-space matrix X(t,s). Here, m, s and t denote different model, space and time components, respectively (Li and Xie, 2012, 2014; Wang et al., 2014).

3. Results
  • The region (25°-45°N, 70°-105°E), with an elevation >2000 m, is chosen as the study area. Since the model outputs are interpolated to the resolution of 2.5°× 2.5°, there are 60 grid boxes in the TP region (Fig. 3).

    Figure 3.  Spatial distribution of elevation over the TP. The solid lines indicate the TP region with an average elevation >2000 m. The locations of the 60 grid boxes are represented by squares.

  • The average of the simulated T s and T as in most models is lower than those based on the reanalysis data of CFSR and ERA-Interim in all four seasons (Fig. 4). The T as in all models and the two reanalysis datasets is below that of CRU, except in summer. The T as differences between each of the 28 CMIP5 AGCMs and their ensemble and the CRU data suggest that most models and the ensemble mean underestimate the summer (June-August) mean temperature over the TP (Fig. 4). The cold bias is more prevalent over the western TP than the eastern TP. Figure 4 shows that the simulated T as values are lower than the CRU values for the spring (March-May) mean and winter (December-February) mean (Figs. 4a and d). The differences between the simulated T as and that of the reanalysis datasets (ERA-Interim and CFSR) are also presented in Figs. 4a-d. Compared with the reanalysis data, the majority of the simulated T as values include a cold bias over the TP during all four seasons. The T as cold bias is larger in winter than in other seasons in most models, as well as the ensemble mean (Fig. 4d), which is consistent with the results of (Su et al., 2013).

    The simulated T s values also include a cold bias over the TP, in half of the models (Fig. 4e-h). Except in autumn, a cold bias dominates among the 28 models and the ensemble mean. The differences between the simulation and ERA-Interim or CFSR are larger for T as than they are for T s. This may be due to the larger variability in air temperature than ground temperature, associated with the greater heat content of soil compared with air. We also find the T s in CFSR and ERA-Interim differ from each other across the four seasons over the TP (Fig. 1 and Figs. 4e-h). In summary, by comparing simulations from 28 AGCMs with CRU data and two reanalysis datasets, we find that most of the models underestimate T as, while half of the models, as well as their annual ensemble mean, underestimate T s.

    Figure 4.  Seasonal mean temperature differences between CMIP5 models and observations [CRU and two reanalysis products (ERA-Interim and CFSR)] averaged over the 60 grid boxes from 1979-2005: (a-d) Tas differences between the models, CRU data and reanalysis products; (e-h) T s differences between the models and reanalysis products. All temperature data in the models and reanalysis products have been corrected based on actual elevation at a resolution of 2.5°× 2.5°. The numbers at the bottom of the plots indicate the models listed in Table 1.

  • We examine the inter-model variability of annual mean T s values over the TP via an inter-model EOF analysis of the 28 CMIP5 AGCMs. The first inter-model EOF (EOF1), which explains 63.2% of the total variance, suggests a general pattern of same-sign values over the entire TP, with a maximum in the southwest (Fig. 5a). EOF1 is consistent with the spatial distribution characteristics of the cold bias in most models (Fig. 2). The first principal component (PC1) is highly correlated with the mean T s of the TP, with a correlation coefficient of 0.98 (Fig. 5b). By comparing with reanalysis data, we check the necessary variables for Eq. (4) in the 28 models and identify individual models with large differences between the simulated results and observations.

    Figure 5.  (a) First inter-model EOF patterns of annual mean T s for the 60 gird boxes over the TP based on 28 CMIP5 AGCMs. (b) PC1 and surface temperature averaged from the 60 grid points over the TP based on 28 CMIP5 AGCMs. The numbers at the top of (a) and (b) indicate the explained variance and correlation coefficient, respectively.

    While a cold bias may be represented by the difference between a model simulation and observations, it may also be understood through the difference between the cTP and wTP models, i.e., through the model spread of the simulated temperature over the TP. This is particularly useful given that direct observations over the TP are still rare, and reanalysis datasets still contain large uncertainties. We choose three of the lowest PC1 values as the cTP models and three of the highest PC1 values as the wTP models. We choose the lowest values and highest values from the five PC1 sequences (i.e., for each of the four seasons and the whole year), and exclude two models with the coldest and warmest bias. As a result, the selected cTP models are INM-CM4.0, MRI-AGCM3.2H and MRI-AGCM3.2S; and the wTP models are CSIRO Mk3.6.0, MIROC5, and NorESM1-M (Table 1). The selected cTP (wTP) models also show a cold (warm) bias in their simulation of T s compared with the reanalysis data (ERA-Interim and CFSR).

  • We decompose the T s bias using Eq. (4). Figure 6a illustrates the T s differences between the cTP models and wTP models over the TP. Figures 6b-g correspond, respectively, to the terms on the right-hand side of Eq. (4), and Fig. 6h is the sum of these terms. The spatial and temporal distributions in Figs. 6a and h are relatively similar. The error between values is mainly due to the linearization of surface upward LW radiation adopted on the left-hand side of Eq. (4), as well as the interpolation calculation error. The differences between the cTP models and the wTP models suggest that the cold bias is widespread over the TP (Fig. 6a). The cold bias over the western TP is higher than that over the eastern TP (Fig. 6a). Additionally, the two largest terms contributing to the cold bias are SAF and the bias due to DLR (Figs. 6b and e).

    Figure 6.  Differences (K) in the various terms in Eq. (4) between the cTP models and the wTP models. The vertical coordinates represent months. The horizontal axis represents the 60 grid boxes arranged by increasing elevation: (a) T s difference, along with that due to (b) surface albedo feedback, (c) surface CRF, (d) non-SAF associated with clear-sky SW radiation, (e) non-SAF associated with net clear-sky LW radiation fluxes, (f) heat storage, (g) surface sensible and latent fluxes, and (h) the sum of (b) through (g).

    3.3.1. Albedo, surface heat flux and heat storage

    The SAF is negative for most of the grid boxes (Fig. 6b), which may be related to the albedo parameterization schemes of the models. (Qu and Hall, 2007) evaluated 14 models from the Fourth Assessment of the Intergovernmental Panel on Climate Change AR4 and found that models that did not explicitly consider the vegetation canopy in their surface-albedo calculations typically had high effective snow albedo values and strong SAF. When the surface albedo increases, less solar radiation is absorbed by the surface, resulting in net surface cooling. A sufficiently large albedo will inevitably lead to a diminished heat source over the TP. Therefore, the larger the albedo, the smaller the sensible heat flux. Figures 6b and g show that the sensible and latent heat flux patterns are similar to the SAF but with opposite signs, implying that the two terms offset each other. The cancelling out of these two terms leads to a relatively weak contribution to the cold bias (Fig. 7h). Comparison between Figs. 7f and g illustrates that the heat storage difference is a positive contribution during spring and a negative contribution during summer.

    The larger albedo differences mainly occur during spring and increase with increasing elevation (Fig. 6b). Figure 7b shows that the maximum albedo differences propagate from east to west over the TP, with seasonal evolution from winter to summer. These temporal differences may be related to the snow cover on the TP because snow cover significantly impacts the albedo of the area. Most models calculate the area of snow cover using an empirical formula based on snow depth. As (Yang et al., 2016) indicated, parameterized albedos have substantial discrepancies compared to observed albedo, particularly during the melt period. Spring is the snowmelt season over the TP; therefore, the largest albedo differences are observed during this season. Figure 8 illustrates the snow area fraction differences between the cTP models and wTP models over the TP. The consistent snow area fraction and SAF patterns indicate that the snow area fraction is larger in the cTP models and smaller in the wTP models. Consequently, the albedo is higher in the cTP models and lower in the wTP models. The higher cTP albedo leads to less solar radiation reaching the ground. The lower wTP albedo leads to more solar radiation reaching the ground and a warm bias. Given the importance of snow albedo in the majority of the CMIP5 models, (Qu and Hall, 2014) indicated that the snow area fraction parameterization scheme must be further improved.

    Figure 7.  As in Fig. 6, but the horizontal axis is the meridional mean averaged over the grid boxes with the same longitude. Panel (h) is the sum of (b) and (g), while (i) is the sum of (b) through (g).

    3.3.2. Surface CRF

    The surface CRF difference is positive for most of the TP grid boxes (Fig. 6c), and is negative over the central TP during winter (Fig. 7c). The surface CRF reaches a positive maximum during summer (Fig. 7c), which is mainly due to the increased solar radiation reflection by clouds (Yan et al., 2016). Although the surface CRF differences occur in-phase with the winter cold bias, the differences do not contribute much to the cold biases during the cold seasons compared to the total cold bias shown in Figs. 7a and i. Figure 7d shows the non-SAF net clear-sky downward SW radiation differences at the surface. This term is mainly impacted by the changed water vapor. Less/more water vapor reduces/increases the atmospheric absorption of solar radiation and increases/reduces solar radiation at the surface.

    3.3.3. Clear-sky DLR

    The annual DLR cycle differences (Fig. 6e) exhibit similar spatio-temporal distributions as those displayed in Fig. 6a. The increase/decrease in DLR at the surface is due to the net effect of the water vapor, atmospheric CO2 concentration, and atmospheric temperature differences. As we know, the AMIP5 models use the same CO2 concentration in the AMIP experiments. Therefore, the difference in DLR in Fig. 6e should be mainly attributable to the differences in atmospheric temperature and water vapor content. Atmospheric temperatures of the cTP models are lower than the temperatures of the wTP models below 300 hPa over the TP (Fig. 8b). The atmospheric temperature is partly affected by the surface turbulence (Fig. 7h) and partly by atmospheric processes. The latter is beyond the scope of the surface energy method in this paper. Recently, (Lu and Cai, 2009b) developed the Coupled Surface-Atmosphere Climate Feedback Response Analysis Method to quantify the contributions of individual model processes to the total temperature changes. This method will be used to investigate the effects of atmospheric processes on atmospheric temperature in a future study.

    The TP has experienced a winter warming trend of 0.4°C (10 yr)-1 during the last five decades, caused by increased near-surface water vapor, which has amplified the DLR (Rangwala, 2013). (Rangwala, 2013) found that, when an equal mass of water vapor is added into the atmospheric boundary layer during winter, a substantially greater increase (× 8) in DLR occurs at high-elevation sites relative to low-elevation sites. The water vapor contents of the cTP models are generally lower than the contents of the wTP models over the TP, which is associated with the lower temperature in the lower troposphere in the cTP models (Figs. 8b and c). Therefore, lower water vapor content (Fig. 7a) and negative DLR are generated (Fig. 7e), which result in a lower T s in the cTP model simulations.

    Figure 8.  Differences between the cTP and wTP models in (a) meridionally averaged snow area fraction, (b) annual mean atmospheric temperature in the vertical direction, and (c) column water vapor content (kg m-2). The meridional mean over the TP is used in (a) and (b).

4. Discussion
  • The method of subtracting wTP model results from cTP model results has been used previously to highlight the origins of SST biases in CMIP multi-model ensembles (Li and Xie, 2012, 2014). We use different models for our comparison in this paper, and different models have differences in their atmospheric state and/or model parameterizations. Therefore, is it reasonable that we use the results of low-temperature models minus those of high-temperature models?

    The use of the surface energy budget is justified by the fact that, in each model, the surface energy balance is maintained in the same way (i.e., use of the same equation), and the biases of the models with respect to the "unknown" real state (observations) may be considered as a perturbation of the "real state". Thus, the analysis presented in this paper is a kind of ensemble perturbation (bias) analysis, which is methodologically reasonable.

    Figure 9.  As in Fig. 7, but the differences are from the multi-model ensemble and CFSR.

    The results presented in section 3 refer mainly to the results of the cTP models minus those of the wTP models. Meanwhile, 15 other models (CanAM4, BCC_CSM1.1, CCSM4, CNRM-CM5, GFDL CM3, HiRAM-C180, HiRAM-C360, GISS-E2-R, INM-CM4.0, IPSL-CM5A-LR, IPSL-CM5A-MR, IPSL-CM5B-LR, MRI-AGCM3.2H, MRI-AGCM3.2S, MRI-CGCM3) that exhibit a cold bias over the TP are also used. Specifically, we analyze the bias of T s between the multi-model ensemble and CFSR (Fig. 9). The results show that the multi-model ensemble bias is contributed by the SAF (Fig. 9b), the LW radiation (Fig. 9e), and the non-SAF clear-sky SW radiation (mainly due to atmospheric water vapor) (Fig. 9d). We also compare the bias of each single model with CFSR. The results are basically the same as those for the ensemble mean (not shown).

    The use in this paper of the cTP model results minus those of the wTP models is done mainly because these models use the same forcing fields in their CMIP5 experiments. The performance of the multi-model ensemble is also tested. The two types of analysis show similar physically interlinked processes. The strong negative SAF is largely cancelled out by the surface heat fluxes. Therefore, these two terms contribute relatively weakly to the cold bias. Smaller surface turbulent fluxes lead to colder lower-tropospheric temperatures and lower water vapor content, which in turn cause negative DLR and a cold bias.

    The focus in this study is on the formation of a T s cold bias, which is important in causing a T as cold bias. However, the change in T as is not only due to change in T s; it is also associated with other physical processes, such as the absorption of radiation energy and latent heating/cooling related to condensation/evaporation. For several of the models that possess a warm T s bias but a cold T as bias, the negative T as bias should be related to either less water vapor content, which absorbs less radiation heating, or stronger sub-cloud/near-surface evaporation. Although such a case exists in only a few models, the mechanism requires further investigation.

5. Summary
  • In this work, we adopt a surface energy balance decomposition method to analyze the T s cold bias over the TP in CMIP5 AGCMs, and find that T as is underestimated over the TP by most of the models, while half of the models underestimate the annual T s. This cold bias is more pronounced over the western TP. The snow area fraction is larger in cTP models than wTP models, resulting in higher albedos. The analyses show a suite of physically interlinked processes contributing to the cold surface temperature bias. The strong negative SAF and the surface heat fluxes are highly anti-correlated, and the cancelling out of these two terms leads to a relatively weak contribution to the cold bias. Smaller surface turbulent fluxes lead to colder lower-tropospheric temperature and lower water vapor content, which in turn cause negative DLR and a cold bias.

    In most of the models, albedo is mainly associated with the parameterization of the area of snow cover, when there is snow, and this is calculated based on snow depth. (Swenson and Lawrence, 2012) indicated that when snowfall occurs, the area of snow cover and the snow depth generally exhibit a good relationship. However, during snowmelt periods, the empirical algorithm based on snow depth exhibits obvious discrepancies compared to the simulated snow area. Considering the importance of the area of snow cover in the albedo calculation, and the importance of the effects of albedo on the cold biases over the TP shown in this study, we suggest that improvements in the parameterization of the area of snow cover, as well as the boundary layer, and hence the surface turbulent fluxes, may help to reduce the cold bias over the TP in models.

    The resolutions of most existing CGCM models are continuously improving. However, corresponding physical process improvements will also be necessary. Resolution enhancements in steep mountain areas may increase the difficulty in simulating land-atmosphere interaction, and the consequent change in land processes will affect the simulation of circulation, water vapor transport, and water vapor distribution, which will ultimately affect the simulation of surface temperature via DLR variations. The results of the present study suggest that the T s cold bias over the TP is closely associated with snow cover and the SAF, and therefore provide a reference for reducing the cold bias and improving climate model performance.

Reference

Catalog

    /

    DownLoad:  Full-Size Img  PowerPoint
    Return
    Return