Impact and Sensitivity Analysis of Soil Water and Heat Transfer Parameterizations in Community Land Surface Model on the Tibetan Plateau

Soil water and heat transfer is especially complicated during the freezing and thawing processes over the high‐altitude cold regions. In this study, four sensitivity tests of soil water and heat transfer parameterizations including replacing soil property data (SP1), soil resistance scheme modification (SP2), soil thermal conductivity scheme (SP3) and virtual temperature scheme (SP4), and four combination experiments (SP1+SP2+SP3/SP5, SP1+SP2+SP4/SP6, SP1+SP3+SP4/SP7, and SP1+SP2+SP3+SP4/SP8) were done using Community Land Model (CLM5.0) to examine its performances for soil water and heat transfer modeling on the Tibetan Plateau (TP) both in single‐point and regional simulations. The observed data from five eddy covariance sites, four soil moisture and temperature networks and 60 sites of soil temperature observations on the TP were used to evaluate the results. Single‐point simulations show that SP2 experiment reduced the wet biases of soil moisture in semiarid area, but enhanced the error of soil temperature. SP6 shows the best performances in simulating soil moisture, and SP3 in soil temperature. Regional simulations show that the SP7 experiment had the best performances for soil water and heat transfer simulation on the TP, and improved the simulation of soil freezing‐thawing processes. Compared to CLM5.0 default simulation, SP7 shows the best performances. For soil moisture, it reduced average Bias by 23%, Root Mean Square Error (RMSE) by 18%, and increased the Correlation Coefficient (Corr) by2%. For soil temperature, it reduced the Bias by 9%, 10%, 23%, and 13% at four soil depths on the TP, respectively.


Introduction
Interactions between the land surface and the atmosphere play a critical role in the weather and climate (Luo et al., 2017), especially it has been demonstrated that regional climate change is sensitive to land surface processes within a seasonal timescale (Dirmeyer et al., 2009;Zhang et al., 2011). Soil water and heat transfer is one of the important parts in water and energy partition between land surface and atmosphere and it is more complicated over the Tibetan Plateau (TP) due to the freezing-thawing processes. Soil moisture is a key variable in land surface processes due to its memory capacity, and it is important in numerous environmental applications, including hydrological modeling, weather and short-term climate forecasting (Meng et al., 2017). Soil temperature can reflect the thermal state of the soil, and directly affects the exchange of energy near the land surface as the upward longwave radiation, sensible and ground heat fluxes depend on it (Mahanama et al., 2008;Zheng et al., 2015). Heat flux exchanges at the land-atmosphere interface play an important role in controlling atmospheric heating and ground warming (Zheng et al., 2014), to influence regional and even global climate change.
The TP has an average elevation of more than 4,000 m, is the highest and the most extensive plateau in the world, and often referred to as the "Third Pole" (Deng et al., 2019;Wu et al., 2015). The TP was demonstrated to have profound influences on not only the local climate, but also the Asian climate, and even on the global climate by its dynamic and thermal forcing (Ma et al., 2014). Compared to other parts of China, the soil genesis layer is shallow and contains large soil particles and high gravel content over the TP. In recent years, longterm observation network has been gradually established over the TP, and provided a large number of reliable observation data for the verification of various soil moisture and temperature products. The soil freezing-thawing cycle in cold region is another feature on the TP. Soil moisture coupled with heat transport, and soil temperature changes interact with soil water during phase change (Niu & Yang, 2006;, which makes the mechanism of soil water and heat transport more complicated (Cuntz & Haverd, 2018). Yang et al. (2018) used virtual temperature instead of constant freezing point to determine the occurring of phase change. Results show that the modification can reproduce the features of daily and diurnal variations in soil moisture and showed closer simulation for freeze-thaw processes against to the observations. Soil thermal properties are highly important in land surface processes modeling, and also crucial for determining land surface energy partitioning and water budget. Generally, these parameters are empirically estimated from soil porosity, soil texture, and water content in current land surface models (LSMs; Farouki, 1981;Lu et al., 2007;Mccumber & Pielke, 1981). The soil thermal conductivity can quantify the rate of heat transfer between different soil depths, directly determines the vertical distribution of soil temperature and soil heat flux and further affects soil freeze-thaw cycles, soil water movement, and related processes such as root water uptake and transpiration (Cuntz & Haverd, 2018;Dai & Nan, 2019;. Dai and Nan (2019) incorporated seven typical soil thermal conductivity schemes in the literature into the Common Land Model to evaluate their application in land surface modeling. Among them, the Johansen scheme (1975) and its three derivatives, that is, Côté and Konrad (2005), Balland and Arp (2005), and Lu et al. (2007), are significantly superior to the other schemes. The Balland and Arp scheme ranks at the top among the selected schemes. Chen et al. (2012) investigated soil characteristics based on 77 soil profiles at 34 stations and found that the top layer soil of alpine grasslands included higher soil organic carbon content (Chen et al., 2012). The land surface model is the land component of earth system model, and its ability in simulating soil water and heat transfer determined the performance of climate model in regional and global climate simulations (Dickinson et al., 2010). The complexity of the land surface process in cold regions increased the uncertainties of the land surface model. The biases in the land surface model are mainly from three aspects, the first is imperfect of the parameterization scheme; the second is the uncertainty of initial and boundary conditions. The quality of the forcing data in land surface modeling is also an important factor (Deng et al., 2020;Yang et al., 2018).
Community Land Model (CLM5.0) is the latest version of CLM, which builds on progress made in CLM4 and CLM4.5 (Lawrence et al., 2019). Compared to CLM4.5, a dry surface layer-based (DSL) soil evaporation resistance parameterization was introduced in CLM5.0, which improves the soil evaporation and surface total water storage simulation in semiarid region (Swenson & Lawrence, 2014). In addition, new parameterization for fresh snow density (updated temperature effects and wind effects) was also added in CLM5.0 (Kampenhout et al., 2017). Deng et al. (2020) compared soil water and heat transfer modeling over the TP by using two CLM versions, and found that CLM5.0 overestimated soil moisture in arid and semiarid regions, but improved the simulation for soil temperature effectively. In addition, Deng et al. (2020) also found that soil property data affect model results, and the impact of soil property data is more important than the forcing data. Previous studies have improved soil water and heat transfer simulations on the TP.

of In Situ Observations and Land Surface Model Products Used in This Study
However, large biases in simulations still remain on the cold TP region, remaining the model to be developed using the observational based physical knowledge replenish. In this paper, we collected multiple observational data, including soil moisture and temperature networks data, multi-site soil temperature data, and the observed soil property data etc., to try to find out a combination of the parameterization schemes to better describe soil and water heat transfer processes on the TP.
Followed by the work of Deng et al. (2020) in which we assessed the performances of CLM 4.5 and CLM5.0 on the TP, in this study we sought to further test the most recent CLM version (i.e., CLM5.0) based on dry surface layer schemes (Swenson & Lawrence, 2014), Balland and Arp schemes (Balland & Arp, 2005) and virtual temperature schemes , to conclude a better combination of parameters and physical parameterization schemes in soil water and heat transfer modeling for the TP region. Following this introduction, the data and method are briefly described in Section 2. Section 3 introduced the description of the model and the model configuration. The assessments of soil water and heat transfer modeling are analyzed in Section 4. The conclusions and discussions are included in Section 5.

Study Area
The TP, which covers an area of ∼2.5 × 10 6 km 2 and with more than half of the area over 4,000 m above sea level. In this study, we mainly focused on the area of 25-40°N and 75-105°E. Figure 1 contains 60 conventional meteorological observation sites (the black points), four soil moisture and temperature monitoring network on the central TP, and five eddy covariance observational sites (Maqu, Maduo, Naqu, Amdo, and Dangx). The eddy covariance observations aim for long-term monitoring of atmospheric physics, energy and water partition, soil physics, hydrological and ecosystem carbon cycle. Detailed descriptions of the four eddy covariance sites (Maqu, Maduo, Naqu, and Amdo sites) are provided by Deng et al. (2020). The Dangx site is situated near Lasa, Xizang Province, China, with a cold and semiarid climate condition. The vegetation in Dangx is also a typical steppe. The mean annual temperature is 1.7°C, and the mean annual precipitation amount is 550.4 mm in 2004 (Yu et al., 2006). The Maqu soil temperature and moisture monitoring network is located at the north-eastern edge of the TP, and in a cold humid environment (Su et al., 2011). The Naqu network is located in a cold semiarid environment. The soil at Naqu network has a high saturated hydraulic conductivity positioned on top of an impermeable rock formation (Su et al., 2013). The Ali network is located in the western part of TP and in a cold arid environment (Su et al., 2011). The Pali soil temperature and moisture monitoring network in a cold and arid environment. These four soil temperature and moisture monitoring networks provide representative coverage of different climate and land surface conditions on the TP.

Data
In this study, the single-point atmospheric forcing data were from five field observational sites, and the regional atmospheric forcing data was the data developed by the Institute of Tibetan Plateau Research (hereafter ITP), Chinese Academy of Sciences (Yang et al., 2010). The ITP data starts from 1979, with 0.1° × 0.1° spatial resolution and 3 hourly temporal resolution. The forcing data used in this study include air temperature, wind, specific humidity, and pressure near surface, and downward shortwave radiation, downward longwave radiation. The wind, air temperature, pressure, and specific humidity data are obtained from Global Land Data Assimilation System (GLDAS1) and 740 China meteorological stations. Downward shortwave radiation data is from Global Energy and Water Exchanges (GEWEX), GLDAS1, and  observations. Precipitation data is merged by Tropical Rainfall Measuring Mission (TRMM) precipitation retrievals, GLDAS1, and observations (Yang et al., 2010).
The observed soil moisture and temperature data from the five stations were used to validate the single-point simulations of CLM. The observed soil moisture and temperature depths used in this paper include 5, 10, 20, and 40 cm. The observed soil moisture data from the Maqu, Naqu, Ali, and Pali soil moisture and temperature networks and soil temperature data from 60 stations over the TP ( Figure 1) were used to evaluate the performance of model (Table 1).
In the single-site offline simulations, the soil property data are observations from Maqu, Maduo, and Naqu sites. As for Amdo and Dangx sites and regional simulation, the soil property data is from Beijing Normal University (i.e., BNU). The original data of BNU soil property data are obtained from China 1:10 6 soil type  property data can be found in Shangguan et al. (2012). The use of observed and BNU soil property data can better represent the real condition of soil over the TP.

Analytical Method
In this paper, the simulations and the observations data were processed to daily averaged value. The neighboring grid matching method was used to analyze the simulated value in order to match with observations. The soil is separated into 25 layers in CLM5.0, the calculation of soil temperature and moisture is in the middle of each layer (Deng et al., 2020). Four-layer (5, 10, 20, 40 cm) soil temperatureand moisture from Maqu, Maduo, Naqu, Amdo, and Dangx sites are compared to CLM5.0. Observations at 5, 10, 20, and 40 cm were compared against CLM5.0 simulations at 4, 9, 16, and 40 cm respectively.
In order to quantify the differences between simulations and observations, four statistical features are calculated at each station: bias (Bias), root mean square error (RMSE), correlation coefficient (Corr), ratio of standard deviation (RSD).  To quantify the performances of simulations in terms of each statistical feature for all stations, a ranking scheme is utilized (Wang & Zeng, 2012). We extend Brunke ranking schemes to four statistical quantities calculated from soil moisture and temperature in this paper. At each station for each statistical of soil temperature and moisture and heat flux, different single-point simulations are ranked from 1 to 9, with 1 given to the simulation with the lowest value of Bias or RMSE in magnitude (or the highest Corr) and 9 given to the simulation with the largest value of Bias or RMSE in magnitude (or the lowest Corr). For RSD, 1 is given to the simulation with the ratio closest to 1, and 9 is given to the simulation with the ratio farthest from 1. All ranking scores are averaged at each station, to obtain the overall ranking score of different simulations at each station. The lowest values of the ranking score represent the closest relationships between simulations and observations. For different evaluations, some experiments perform better in some aspects and poor in others. Here, we extended a new comprehensive statistic metric, DISO (the distance between indices of simulation and observation) to evaluate. If the statistical metrics for n chosen are ( 1 E s, 2 E s ,… n E s ) and the corresponding metrics between the observed data and itself are ( 1 , then DISO is calculated as (Zhou et al., 2021): In this study, DISO, which is composed of four widely used statistical metrics: Bias, RMSE, Corr, and RSD. In order to make these four statistical metrics have the same weights, we normalized the observed data and different experiments data and then calculated Bias, RMSE, Corr, and RSD.
If the simulation perfectly performs the best, the best statistical metrics are: Corr = 1, Bias = 0, RMSE = 0, and RSD = 0, then DISO can be expressed: Now, the value of DISO expressed by the statistical metrics between the simulated and observed data are used to evaluate the performance of different experiments. A larger value of DISO indicates that the experiment has a poorer performance. In this paper, the most important and best advantages of DISO is that after normalizing the observed and simulated data, the value of DISO can express the performance of the same experiment at different sites. The value of DISO in this paper ranges from 0 to 2.

Description of CLM
CLM is the land component of the Community Earth System Model (CESM), which has been developed and expanded over the last decade (Lawrence et al., 2019), and the CLM5.0 is the latest version of the CLM. Compared to CLM4.0, the main parameterizations modified in CLM4.5 include hydraulic properties of frozen soil determined only by liquid water ; modified the snow cover setting to reflect hysteresis of the fractional snow cover at a given depth between accumulation and melt phases . In addition, most of the main components of the model have been updated, detailed descriptions of the updating are provided by Lawrence et al. (2019).

Dry Surface Layer Schemes
Previous studies show that the overestimation of soil moisture in CLM5.0 was mainly caused by the introduction of a dry surface layer based on soil evaporation resistance parameterization in the semiarid area over the TP (Deng et al., 2020). In this paper, we deprecate the dry surface layer and setup soil resistance to 0.
In CLM5.0, the soil evaporation is formulated as: Where atm E q and soil E q are the atmospheric specific humidity, soil specific humidity, respectively, atm E  is the atmospheric density, E raw and E r are the aerodynamic and soil resistance to water vapor transfer, respectively.

Balland and Arp Schemes
In CLM4.5 and CLM5.0, the soil thermal conductivity is mainly from Farouki (1981). Thermal conductivity E  is formulated as follows: Where sat E  and dry E  are the saturated thermal conductivity and the dry thermal conductivity, respectively. Ke E is the Kersten number, and S E r is the wetness of the soil with respect to saturation.
The Kersten number is a function of the degree of saturation Sr E and phase of water.
Farouki (1981) simplified the K e r E S  relationship by ignoring the effects of the soil particle size distribution. To solve this issue, Balland and Arp (2005) Where s E v, om E v , and g E v are the volumetric fractions of sand, soil organic matter, and gravel among all soil components, respectively. The data of sand, soil organic matter, and gravel content from the observed and BNU soil property data. This function takes soil property into consideration, which can lead to continuous variations in soil thermal conductivity predictions over the entire range of soil saturation levels and particle size distribution (Dai & Nan, 2019

Virtual Temperature Schemes
The snow and soil temperatures are evaluated to determine if phase change will take place as: w are the mass of ice and liquid water in each soil layer, respectively, and E Tf is the freezing temperature of water (K).
The phase change is determined by soil temperature, and the concept of supercooled soil water is estimated from Niu and Yang (2006 Where ,max, E wliq i is the maximum liquid water in layer i when the soil temperature is below the freezing temperature, E Lf is the latent heat of fusion, sa , E t i  is the soil texture-dependent saturated matric potential, sa , E t i  is the saturation water content, E Bi and E g are the Clapp and Hornberger exponent and the gravitational acceleration, respectively. Previous studies suggested that the occurring of phase change could affect the freezing temperature (Zheng et al., 2017), which can be calculated as follows Where sat E  is volumetric liquid water. We use of Tv instead of the constant freezing point temperature to determine the occurring of phase change.

Experiments Design
To investigate the performance of CLM in simulating soil moisture and temperature and surface heat flux over the TP, we conducted the following set of single point simulations at Maqu, Maduo, Naqu, Amdo, and Dangx sites using CLM5.0. (a) Sensitivity experiments of single-point simulations were conducted at each site by CLM5.0. The soil thermal conductivity parameterization is replaced with Balland and Arp schemes in these sensitivity experiments, and parameter of E  and E  is fitted to get the best results. (b) Replaced the soil property data using the observed soil property data and the BNU soil property data. (c) The dry surface layer scheme was deprecated, the selected parameter from Balland and Arp schemes was used, and the virtual temperature parameterization (Equation 11) was used to replace the original code in CLM5.0 (see Table 2 for details).  Figure 2 shows the time series of simulated and observed soil moisture in Amdo, Maduo, Maqu, and Naqu (Dangx omitted) at 5, 40 cm (10, 20 cm omitted). The correlations between the observed soil moisture and the simulations at these four sites all pass the correlation statistical significance at or above 99%. Soil property data, which represent the underlying heterogeneity and property, are crucial for land surface simulation. Compared to SP0, the soil property data was replaced by the observed data instead of CLM5.0 default in SP1 experiment. The default CLM5.0 simulation (SP0) overestimated soil moisture at these four sites for the entire year, especially during the thawing period, and the average Bias at these four sites in four layers is 0.104 m 3 ·m −3 . After replacing the soil property data, SP1 tended to decrease the wet biases at Amdo sites, the average Bias decreased from 0.095 to 0.078 m 3 ·m −3 , the Corr increased from 0.790 to 0.819, and the RMSE decreased from 0.111 to 0.094 m 3 ·m −3 at four soil depths (5, 10, 20, 40 cm). In terms of all five sites (Amdo, Maduo, Maqu, Naqu, and Dangx), the average Bias decreased by 2%, the Corr increased by 1%, and the RMSE decreased by 2% at four soil depths. In SP2 experiment, the increase of soil evapora- tion resulted in the decrease of surface water storage, and the decrease of soil moisture. When modifying soil resistance to 0, SP2 experiment significantly reduced the wet biases of the entire soil column during the entire year. At the semiarid site Amdo, soil moisture simulationsin SP2 experiment significantly decreased the wet biases. Compared to SP0 and SP1 experiments, the characteristics of soil moisture variation and soil freezing-thawing process in SP2 more coincided with that of observed. Compared to SP1, SP2 experiment decreased the wet biases of soil moisture during the freezing period, and increased the dry biases at 5, 10 cm depths during the thawing period. At Maduo, Dangx, and Naqu sites, SP2 experiment decreased the wet biases of soil moisture, and underestimated soil moisture during the thawing period at 5 cm depth. Overall, the average Bias of SP2 experiment at these five sites at four soil depths reduced by 91%, the average RMSE reduced by 45% compared to SP1 experiment. In SP3 experiment, we modified soil thermal conductivity scheme, and involved the gravel content into consideration. At these four sites (Amdo, Maduo, Naqu, and Maqu), the content of gravel at Maqu site in the shallow layer was less than Amdo, Maduo and Naqu sits, but more in the deep soil depths (20 and 40 cm). In SP4 experiment, we replaced virtual temperature instead of the freezing point  temperature. Compared to SP1, soil moisture simulations in SP4 showed no obvious differences at Maduo, Amdo and Naqu sites, with the time of soil freezing and thawing had been postponed.

Single-Point Simulation Results
In single-point simulations, SP3 and SP4 didn't show large improvements for soil moisture, but the description of physical processes in soil water and heat transfer is more complete. In this part, we combined SP2, SP3, and SP4 experiments to evaluate the performances of different scheme combinations (Table 2). Compared to SP1, the average Bias of SP5 for soil moisture reduced by 89%, the Corr increased by 1%, and the average RMSE reduced by 45%; the average Bias of SP6 reduced by 75%, the Corr increased by 2%, and the average RMSE reduced by 38%; the average Bias of SP7 increased by 16%, the Corr reduced by 8%, and the average RMSE increased by 16%; the average Bias of SP8 reduced by 71%, the Corr increased by 1%, and the average RMSE reduced by 40%. Figure 3 shows the statistical metrics DISO in these five sites at four soil depths. The value of DISO indicates the distances expressed by the statistical metrics between observations and simulations, with a larger value of DISO indicating a poorer comprehensive performance. When modifying soil resistance to 0, SP2, SP5, SP6, and SP8 experiments showed great improvements at Amdo, Dangx, Naqu sites at all four soil depths, especially for SP6 and SP8 experiments. At 40 cm soil depth, the value of DISO had a larger value in different experiments at Amdo, indicating the poorest performances in Amdo site at 40 cm soil depth. In Maduo site, there were no obvious differences among these nine experiments at 5 cm soil depth, but SP2 and SP5 experiments showed the best performance at 10, 20, and 40 cm soil depths. At Maqu site, SP5 experiment had the best performance at 5 cm soil depth, but SP0, SP1, and SP3 had better performances at 10, 20, and 40 cm soil depths. Overall, in single-point simulations, SP2, SP5, SP6, and SP8 showed better performance in semiarid region.

Regional Simulation Results
To further verify the universality of modified parameterizations for the whole TP regions, the comparison between simulations and observations in Ali, Maqu, Naqu, and Pali networks is performed. Naqu network locates in a cold semiarid climate, Maqu network locates in a cold humid climate and Ali and Pali networks located in a cold arid climate. These four soil moisture and temperature monitoring networks provide representative coverage of the different climates on the TP. We compared the results of adjacent grid points that were close to the observation sites, and calculated the average value in each network at different soil depths, respectively. Then we ranked soil moisture simulations of nine experiments based on Bias, RMSE, Corr, and RSD statistic metrics at each network. Figure 4 shows the time series of soil moisture at Ali and Maqu networks. Ali network is in a cold arid environment, and the default experiment SP0 in CLM5.0 overestimated soil moisture, while SP2 effectively reduced the wet biases of soil moisture in the whole year, especially in thawing period. Brunke ranking scores indicated that SP6 experiment had the best performances, followed by SP2 and SP7 experiments. Compared to SP0, SP6 experiment decreased the average Bias from 0.056 to 0.019 m 3 ·m −3 , increased Corr from 0.828 to 0.862, and decreased RMSE from 0.059 to 0.023 m 3 ·m −3 . Maqu network is in a cold and humid environment, with higher soil water content. SP2, SP5, SP6, and SP8 experiments underestimated soil moisture at Maqu network. Compared to SP0, soil freezing-thawing processes simulation in SP7 experiment shows better agreement with observations. Brunke ranking scores show that SP1 experiment had the best performances at Maqu network, followed by SP7. And SP7 experiment had the highest Corr compared to other experiments due to the better description for soil freezing-thawing processes. Compared to SP0, the average Bias of soil moisture in SP1 experiment was reduced by 91%, Corr was increased by 1%, and RMSE was reduced by 15%.
Naqu network is in a cold and semiarid environment, and the average soil water content in Naqu network is higher than the single-point Naqu site. Soil moisture simulated by SP0 experiment generally coincided with the observations (Figure 5). SP2, SP5, SP6, and SP8 experiments underestimated soil moisture during the thawing period. SP0, SP1, SP3, SP4, and SP7 experiments significantly overestimated soil moisture at 20 and 40 cm depths. Compared to SP0, SP4 experiment tended to increase the biases of soil freezing-thawing processes. Brunke ranking scores indicated that SP7 experiment showed the best performances at 5 and 10 cm depths, and SP2 experiment showed the best performances at 20 and 40 cm depths, followed SP7 experiment. Compared to SP0, the average Bias of soil moisture in SP7 experiment was reduced by 8%, the Corr was increased by 2%, and RMSE was reduced by 5%. Figure 6 shows the time series of soil moisture at Pali networks. SP2, SP5, SP6, and SP8 experiments reduced the dry biases of soil moisture at 5 cm depths, and underestimated soil moisture at 10, 20, and 40 cm. Compared to SP0, SP1, and SP3 experiments, SP4 and SP7 experiments reduced the wet biases of the soil moisture. According to Brunke ranking scores, SP2 experiment had the best performances at 5 cm depth, and SP7 had the best performances at 10 cm, 20, and 40 cm depths. Compared to SP0, the average Bias in SP7 for soil moisture was reduced by 30%, Corr was increased by 4%, and RMSE was reduced 29%. Figure 7 shows comparisons of soil temperature against observations between different experiments in Amdo, Maduo, Maqu, and Naqu sites (Dangx site omitted) at 5 and 40 cm depths (10 cm, 20 cm omitted). Soil temperature generally coincided with the observations, and time variations agreed with the observations. The average RMSE SP0 experiment at these four sites is within 3°C. SP2, SP5, SP6, and SP8 experiments underestimated soil temperature during the thawing period. At Amdo site, the average Bias of SP1 experiment for soil temperature decreased from 0.097°C to 0.015°C, Corr remains unchanged at 0.980, and the RMSE decreased from 2.270°C to 2.230°C. Brunke ranking scores indicated that SP3 experiment had the best performances, and SP6 and SP8 experiments had the worst performances. Compared to SP0 experiment, the average Bias of soil temperature in SP3 experiment increased from 0.097°C to 0.100°C, Corr increased from 0.980 to 0.982, and RMSE decreased from 2.270°C to 2.110°C. At Dangx site, SP7 experiment had the best performances, while SP0, SP6, and SP8 experiments had the worst performances. Compared to SP0, the average Bias of soil temperature in SP7 experiment reduced by 99%, Corr increased from 0.958 to 0.980, and the RMSE reduced by 58%. At Maduo site, Brunke ranking scores indicated the best performances of SP3, and the worst performances of SP0 and SP6. Compared to SP0, the average Bias of soil temperature in SP3 decreased from 0.559°C to 0.484°C, Corr increased from 0.984 to 0.985, and RMSE decreased from 1.707 to 1.585°C. At Maqu site, under a cold and subhumid environment, Brunke ranking scores indicated the best performances of SP0, followed by SP3, and the worst performances of SP6. At Naqu site, Brunke ranking scores indicated better performances of SP0, SP1, and SP3. Overall, SP0, SP1, SP3, and SP7 experiments showed better performances for soil temperature simulations. An updating of soil thermal conductivity scheme could improve soil temperature simulation. Figure 11. Same as Figure 6, but for soil temperature (Unit: °C). Figure 8 shows the statistical metrics DISO of soil temperature in these five sites at four soil depths. All experiments showed the worst simulations at Naqu site, but better performances at Dangx and Amdo sites. Overall, SP3 experiment had the best performances for soil temperature simulation in the single-point experiments, followed by SP7. Figure 9 shows soil temperature at Ali and Maqu networks from the regional simulations. At Ali network, soil temperature simulations in these nine experiments generally coincided with the observations, and the RMSE is within 3°C. SP4 overestimated soil temperature in winter, and underestimated soil temperature during May to July. Except for SP4, all experiments show an RMSE lower than 2°C, and Corr of 0.99. Brunke ranking scores indicated that SP1, SP2, and SP6 had better performances, followed by SP5 and SP7. Compared to SP0, the average Bias of soil temperature by SP1, SP2, SP5, SP6, and SP7 was decreased from 1.459°C to 1.312°C, 0.03°C, 1.377°C, 0.790°C, and 0.031°C, RMSE was decreased from 1.907°C to 1.797°C, 1.468°C, 1.871°C, 1.740°C, and 1.487°C, respectively. At Maqu network, in a cold humid environment, SP0, SP1, SP3, SP4, and SP7 experiments overestimated soil temperature. Brunke ranking scores indicated that SP5 experiments had the best performances, followed by SP8, while SP4 had the worst performances. Compared to SP0 experiments, the average Bias of soil temperature in SP5 experiment was decreased from 1.736°C to 0.341°C, Corr was increased from 0.967°C to 0.972°C, and RMSE was decreased from 2.955°C to 1.665°C.

Regional Simulation Results
At Naqu network, in a cold and semiarid environment, all experiments generally coincided with the observations with RMSE within 2°C excepted for SP4, which overestimated soil temperature ( Figure 10). According to Brunke ranking scores, SP7 experiment showed the best performances, followed by SP3. Compared to SP0, SP7 decreased the average Bias of soil temperature from 1.276°C to 0.876°C, and RMSE from 1.751°C to 1.410°C. Figure 11 shows the results at Pali network. SP2, SP5, SP6, and SP8 experiments underestimated soil temperature after modifying soil resistance to 0. Brunke ranking scores show that SP0, SP1, SP3, and SP7 experiments had better performances, with the average Bias at 5, 10, 20, and 40 cm are 0.158°C, 0.234°C, 0.09°C, and 0.087°C, RMSE were 0.967°C, 0.981°C, 0.996°C, and 1.002°C, Corr are 0.987, 0.987, 0.986, and 0.985, respectively.

Overall Evaluations
We calculated the statistical metrics of averaged soil moisture and temperature estimations at Maqu, Ali, Naqu, and Pali networks and then ranked the statistical results by using Brunke ranking scores ( Figure 12). Results show that SP7 experiment had the best performances at 5, 10 cm depths, and SP2 experiment had the best performances at 20 cm, 40 cm depths. Compared to SP0, the average Bias of SP7 was decreased from 0.066 to 0.051 m 3 ·m −3 , Corr was increased from 0.897 to 0.918, RMSE was decreased from 0.080 to 0.066 m 3 ·m −3 . For soil temperature, SP7 had the best performances, followed by SP1. Compared to SP0, the average Bias of soil temperature at these four soil depths was decreased from 0.668°C to 0.471°C, Corr was remained unchanged, while RMSE was decreased from 1.467°C to 1.284°C. Overall, SP7 experiments had the best performances in soil water and heat transfer simulation on the TP. In SP7 experiment, we replaced the default soil property data in CLM5.0 by BNU soil property data, and modifying soil thermal conductivity scheme, in which the influence of soil property and gravel content was involved, which belongs to part of the specific characters of TP soil property. In addition, we introduced virtual temperature instead of freezing point temperature to better represent the occurrence of phase change .
Regional soil temperature simulation results show that SP0, SP1, and SP7 had better performances on the TP. We selected 60 soil temperature observations sites (the locations of the soil temperature observations sites were shown in Figure 1) and compared with the closest adjacent grid points from 2009 to 2014 at 5, 10, 20, and 40 cm depths ( Figure 13). Soil temperature in SP0 experiment generally coincided with the 0 shallow layer (5 cm), there are 25 sites in SP7, 22 sites in SP1, and 13 sites in SP0 experiment showed closest to the observations. In 10 cm soil layer, there are 36, 9 and 15 sites in SP7, SP1 and SP0, respectively. There are 37, 8, and 15 sites in SP7, SP1, and SP0 closest to the observations in the 20 cm layer, respectively. In the 40 cm layer, it is 35, 5, and 20 sites in SP7, SP1, and SP0, respectively. Compared to SP0, the improvement of soil temperature simulation mainly locates in the northeast of the TP. Compared to SP0, the absolute value of biases of soil temperature in SP7 was reduced by 9%, 10%, 23%, 13%, respectively, at four soil depths on the TP. As a result, SP7 experiment shows an overall best simulation for soil temperature on the TP.

Conclusion
Previous studies focused on the evaluation of the performance in CLM5.0 the basic characteristic of soil moisture and soil temperature simulation on the TP (Deng et al., 2020). However, large biases of soil moisture still remain on the TP, and the influences and sensitivity of soil water and heat properties have not been fully investigated. In this study, offline simulations at single-point and regional considering the influences of soil texture, soil resistance, soil freezing, and soil thermal conductivity over the TP were conducted to evaluate the simulation for soil moisture and temperature. After setting soil resistance to 0, the increase of soil evaporation leads to the decrease of surface water storage, and the decrease of soil moisture, improving soil moisture simulation in semiarid area. The thermal processes depend on soil property data. After replacing the soil property data in CLM5.0 default by BNU product and the observations, soil moisture simulation was improved during the thawing period.
To further verify the universality of modified parameterizations for the whole TP, offline simulations at regional scales were conducted. The results show that SP7 can simulate soil moisture and temperature best comparing with other sensitivity experiments. In SP7, we replaced the default soil property data in CLM5.0 by BNU soil property data, modified soil thermal conductivity scheme considering the influence of soil property and gravel content, and introduced virtual temperature instead of freezing point temperature to represent the occurrence of phase change.
Compared to CLM5.0 default simulation, SP7 experiment gets better simulations, while large biases still remain in simulations for processes of freezing, thawing and the completely frozen time period. The biases of the model in soil water and heat transfer simulations for frozen soil are probably from the imperfection of parameterization scheme for the freezing-thawing processes, the deviation of the boundary condition (surface data), etc. Therefore, a detailed soil freezing-thawing processes observation is necessary on the complex land surface lying, including optimization of the parameterization describing these processes is still a main challenge in further work.