Evaluating and improving simulations of diurnal variation in land surface temperature with the Community Land Model for the Tibetan Plateau

Model description

In this study, we used CLM5.0 (Lawrence et al., 2018) as the land surface model, which is the latest version developed by the National Center for Atmospheric Research and serves as the land surface model in the Community Earth System Model version 2.0. CLM5.0 is a “big-leaf” model that conceptualizes the vegetation canopy as a single layer, the snowpack is simulated with up to five layers depending on snow depth, and soil can be divided into an arbitrary number of layers (ten layers in this study). Compared with earlier versions of the model, CLM5.0 includes new soil evaporation resistance parameterization by introducing the concept of the dry surface layer (Swenson & Lawrence, 2014). In CLM5.0, each grid cell is split into different land units, including vegetated surfaces, lake, urban, glacier, and cropland. Each land-unit can be further split into several different columns. Each column is divided into multiple plant functional types (PFTs) (Bonan et al., 2002) or crop functional types. The spatial distribution and seasonal climatology of those PFTs for CLM5.0 are derived from MODIS satellite land surface data products (Lawrence & Chase, 2007). This sub-grid structure can help more accurately resolve the surface heterogeneity in complex terrain regions (Ma et al., 2019b).

In this study, we mainly focused on three parameterizations in CLM5.0: ground sensible heat roughness length (z_0h,g), soil evaporation resistance (r_soil), and soil thermal conductivity (λ), because our results indicated that these parameterizations strongly impact LST (discussed in detail in ‘Results’). Previous studies also reported similar findings (Chen et al., 2018; Pablos et al., 2016; Trigo et al., 2015; Wang et al., 2014; Zeng, Wang & Wang, 2012). Ground sensible heat roughness length (z_0h,g) is of importance for the reliable calculation of the sensible heat flux from ground, which is a function of ground momentum roughness length (z_0m,g): (1)${\displaystyle z_{0h,g}=z_{0m,g}\ast e^{-a{\left(u_{\ast }{z}_{0m,g}∕\nu \right)}^b}}$ where z_0m,g = 0.01 for soil and glacier, z_0m,g = 0.0024 for snow-covered surfaces (m); u_∗ is the friction velocity, ν = 1.5∗10⁻⁵ m² s⁻¹ is the molecular viscosity, a = 0.13, and b = 0.45.

r_soil is used to represent the effect of soil resistance on soil evaporation, which is parameterized as: (2a)${\displaystyle r_{soil}=\frac{TDSL}{D_{\nu }\tau }}$ where D_ν is the molecular diffusivity of water vapor in air (m² s⁻²), and τdescribes the tortuosity of the vapor flow path through the soil matrix. These two parameters are related to soil type (Swenson & Lawrence, 2014). TDSL is the thickness of the dry surface layer (DSL, m) and given as: (2b)${\displaystyle TDSL=\left\{\begin{array}{c}{T}_{\max }\ast \frac{{\theta }_{init}-{\theta }_1}{{\theta }_{init}-{\theta }_{air}}\left({\theta }_1<{\theta }_{init}\right)\hfill \\ 0\left({\theta }_1\ge {\theta }_{init}\right)\hfill \end{array}\right.}$ where T_max= 15 is the maximal DSL thickness (mm); θ_init is the moisture value at which the DSL initiates and is equal to 0.8 times top model soil layer porosity, θ₁ is the moisture value of the top model soil layer, and θ_air is the “air dry” soil moisture value (mm³ mm⁻³) (Lawrence et al., 2018).

In CLM5.0, the soil thermal conductivity (λ) is assumed to be a weighted combination of the saturated (λ_sat) and dry (λ_dry) thermal conductivity (Lawrence et al., 2018): (3a)${\displaystyle \lambda =\left\{\begin{array}{c}{K}_e\ast {\lambda }_{sat}+\left(1-K_e\right)\ast {\lambda }_{dry}\left(S_r>10^{-7}\right)\hfill \\ {\lambda }_{dry}\left(S_r\le 10^{-7}\right)\hfill \end{array}\right.}$ where K_e is the Kersten number expressed as a function of water phase and saturation degree (S_r = θ∕ϕ, θ is the real soil moisture, ϕ is the soil porosity): (3b)${\displaystyle K_e=\left\{\begin{array}{c}l{\mathrm{og}}_{10}\left({\mathrm{S}}_{\mathrm{r}}\right)+1\left(T_{soil}\ge T_f\right)\hfill \\ {\mathrm{S}}_{\mathrm{r}}\left(T_{soil}<T_f\right)\hfill \end{array}\right.}$ Dry soil thermal conductivity (λ_dry) is estimated using the weighted mean of the thermal conductivities of dry mineral soil and dry soil organic matter (SOM), respectively: (3c)${\displaystyle {\lambda }_{dry}=\left(1-f_{om}\right)\ast {\lambda }_{dry,m}+f_{om}\ast {\lambda }_{dry,om}}$ where f_om = ρ_om∕ρ_om,max is the SOM fraction, ρ_om is the SOM density (kg m⁻³) acquired from input surface data, ρ_om,max= 130 kg m⁻³ is the bulk density of peat. λ_dry,om = 0.05 W m⁻¹ K⁻¹ is the dry SOM thermal conductivity, and λ_dry,m is the dry mineral soil thermal conductivity (W m⁻¹ K⁻¹), which depends on bulk density ρ_d= 2700 ∗(1 − ϕ)(kg m⁻³) as: (3d)${\displaystyle {\lambda }_{dry,m}=\frac{0.135\ast {\rho }_d+64.7}{2700-0.947\ast {\rho }_d}}$ Saturated thermal conductivity (λ_sat) depends on the thermal conductivities of the soil solid, liquid water, and ice constituents: (3e)${\displaystyle {\lambda }_{sat}={\lambda }_s^{1-\varphi }\ast {\lambda }_{liq}^{\frac{{\theta }_{liq}}{{\theta }_{liq}+{\theta }_{ice}}\ast \varphi }\ast {\lambda }_{ice}^{\left(1-\frac{{\theta }_{liq}}{{\theta }_{liq}+{\theta }_{ice}}\right)\ast \varphi }}$ where θ_liq and θ_ice are the soil liquid water and ice contents (mm³ mm⁻³), respectively; λ_liq = 0.57 W m⁻¹ K⁻¹ and λ_ice = 2.29 W m⁻¹ K⁻¹ are liquid water and ice thermal conductivities, respectively; and λ_s is the soil solid thermal conductivity, which is calculated using the weighted mean of thermal conductivities of mineral soil and SOM: (3f)${\displaystyle {\lambda }_s=\left(1-f_{om}\right)\ast {\lambda }_{s,m}+f_{om}\ast {\lambda }_{s,om}}$ where λ_s,om = 0.25 W m⁻¹ K⁻¹ is the SOM thermal conductivity and λ_s,m isthe mineral soil solid thermal conductivity: (3g)${\displaystyle {\lambda }_{s,m}=\frac{8.80\ast \left(\text{\%}sand\right)+2.92\ast \left(\text{\%}clay\right)}{\left(\text{\%}sand\right)+\left(\text{\%}clay\right)}}$ where %sand and %clay represent the gravimetric fractions of sand and clay in mineral soil, respectively, and are acquired from the input surface data.

LST calculation

LST is calculated with upward land surface longwave radiation (L↑), downward longwave radiation (L↓), and land surface emissivity (ε) as follows (Wang et al., 2014): (4)${\displaystyle LST=\sqrt[4]{\frac{L\uparrow -\left(1-\epsilon \right)\ast L\downarrow }{\epsilon \ast \sigma }}}$ where σ= 5.67*10⁻⁸ W m⁻² K⁻⁴ is the Stefan–Boltzmann constant. Atmospheric downward longwave radiation L↓ was obtained from the atmospheric forcing dataset (see ‘Climate forcing data’). Surface upward longwave radiation L↑ is the areally-weighted value with the fractions of all the PFTs in the grid cell of CLM5.0. Land surface emissivity (ε) was calculated with the ground emissivity (ε_g) and vegetation emissivity (ε_v) from CLM5.0: (5)${\displaystyle \epsilon ={\epsilon }_v+{\epsilon }_g\ast \left(1-{\epsilon }_v\right)+{\epsilon }_v\ast \left(1-{\epsilon }_g\right)\ast \left(1-{\epsilon }_v\right)}$ where ground emissivity ε_g is combined with the snow emissivity and the soil emissivity weighted by the snow cover fraction. Vegetation emissivity ε_v is parameterized with the exposed leaf and stem area indices (Lawrence et al., 2018). In this study, we used satellite data to evaluate the calculated LST with Eq. (4).

Modifications for three parameterizations

Data

Design of numerical experiments

A suite of numerical experiments was performed to assess the response of the simulated LST, soil thermal properties, and turbulent fluxes to the different representations of ground sensible heat roughness length, soil thermal conductivity, and soil evaporation resistance parameterizations with CLM5.0. Four offline CLM5.0 simulations for the TP for 1995–2018 were conducted with the prescribed satellite-derived phenology. Here, these simulations were conducted only for the TP in China due to the atmospheric forcing dataset CMFD only covering China. CLM5.0 was first run with the original parameterizations (‘Model Description’), defined as the control run (CTL). The second simulation, which employed revision z_0h,g (described in ‘Revision of ground sensible heat roughness length scheme’), was denoted as EXP1. The third experiment (EXP2) duplicated EXP1, but replaced the original soil thermal conductivity scheme with the modified scheme that was described in ‘Modification scheme for soil thermal conductivity’, On the basis of EXP2, soil evaporation resistance parameterization was improved as described in ‘Modification of soil evaporation resistance scheme’ (EXP3). The different experimental setups are summarized in Table 2. The spatial resolution of these four simulations was 0.1°, the same as that of the atmospheric forcing dataset, which reduced the error due to horizontal interpolation. Hourly model outputs for the period of December 2002 through November 2018 were used in the following analysis. The calculated hourly LST was interpolated to the two MODIS Aqua satellite overpass times (local solar time 1:30 a.m. and 1:30 p.m.) and then compared with MODIS Aqua LST. In this study, December, January, and February were defined as winter; and June, July, and August were denoted as summer. In addition, we employed statistical metrics such as the spatial pattern correlation coefficient (PCC), root-mean-square-error (RMSE), and average bias to evaluate the performance of CLM5.0.

Evaluation of the CTL simulation

We first present the performance of CLM5.0 with the original parameterizations (CTL) compared with MODIS LST products. Figure 1 shows the spatial distribution of the LST biases between the CTL and MODIS Aqua data averaged over 2003-2018 during daytime and nighttime in summer and winter, respectively. The LST biases displayed large spatial and diurnal variations. During daytime, the model produced cold biases for almost the entire TP region in winter (Fig. 1A), with a −8.81 K average bias and an RMSE of 9.77 K. In summer, cold biases appeared mainly on bare-ground regions (western TP and the Qaidam Basin), whereas warm biases mainly appeared on vegetated regions (southern TP; Fig. 1B). In addition, the average temperature bias was −1.61 K and RMSE was 5.03 K, averaged for the TP in summer. During nighttime, a warm bias covered the entire TP for both two seasons, whereas the cold bias was less pronounced over the northwestern TP in winter (Figs. 1C, 1D). The average nighttime LST for the TP varied from 2.27 K in winter to 6.43 K in summer. CLM5.0 reasonably captured the spatial patterns of LST for the TP, with PCCs greater than 0.70 for both seasons during day- and nighttime (Figs. S2 and S3), largely attributed to the high spatial quality of CMFD, which we used to drive the model. Interestingly, CLM5.0 accurately simulated land surface emissivity with a small RMSE (0.01) and average bias (<0.01) when compared to MODIS data (Fig. S4). Overall, CLM5.0 could reasonably reproduced TP LST in winter and summer, but with large biases in the magnitudes, indicating large rooms for improvement in TP LST simulation.

Impact of sensible heat roughness length on daytime LST

We applied the revision of z_0h,g (‘Revision of ground sensible heat roughness length scheme’) in CLM5.0, and the impact of this revision on LST for the TP is shown in Fig. 2. The EXP1 simulation increased the daytime LST compared with that of CTL simulations over some regions with bare-soil-underlying conditions, such as the Qaidam Basin and western TP (Figs. 2A, 2B), and the LST increases were more significant when the bare soil fraction increased, especially in summer (Fig. 2E). This mainly occurred due to the revision scheme (Eq. (6)), which reduced z_0h,gvalues over the same bare-ground surface, and then decreased the sensible heat flux; therefore, it increased daytime LST, which is consistent with the results of Wang et al. (2014). The EXP1 simulation reduced cold daytime biases in the CTL simulation over bare ground surface: the RMSE was reduced from 9.77 to 9.60 K in winter (Fig. 2C), and from 5.03 to 4.34 K in summer (Fig. 2D), demonstrating the important role of z_0h,g over bare-ground surface. Improvements in nighttime LST in this study with the revision of z_0h,gwere negligible (Fig. S5) due to the little impact of the z_0h,g revision on nighttime sensible heat flux (Zeng, Wang & Wang, 2012).

Impact of soil thermal conductivity on LST

Two soil thermal conductivity parameterizations (‘Model Description’ and ‘Modification scheme for soil thermal conductivity’) were applied to CLM5.0 and evaluated against the BJ station observations, which are referred to hereafter as CLM_ORI and CLM_NEW, respectively. These two offline simulations were driven by observational half-hourly in situ atmospheric forcing data during the summer in 2008. We compared the 10 cm soil thermal conductivity from in situ observations with CLM_ORI and CLM_NEW simulations, as shown in Fig. 3. CLM_ORI largely overestimated the soil thermal conductivity at the BJ site while CLM_NEW effectively reduced the average bias with CLM_ORI from 1.11 to 0.26 W m⁻¹ K⁻¹. The slight overestimations in CLM_NEW may be related to the inaccurate prediction of soil moisture (Pan et al., 2017).

Next, this modified soil thermal conductivity parameterization (‘Modification scheme for soil thermal conductivity’) was implemented in CLM5.0 to explore the impact of soil thermal conductivity on LST for the whole TP during day- and nighttime. Figure 4 shows the spatial distribution of day- and nighttime LST biases between EXP2 and MODIS Aqua data averaged from 2003 through 2018 in winter and summer. The biases presented large spatial and diurnal variations. In winter, cold biases (average bias: −5.50 K) were dominant over most of the TP during daytime, whereas warm biases appeared over the western TP, and cold biases were distributed along the TP edges during nighttime (Fig. 4C). In summer, warm biases were occurred for the whole TP during both daytime and nighttime, with the average bias values of 2.42 and 3.05 K, respectively (Figs. 4B, 4D). Compared to CTL, EXP2 considerably much improved the TP LST simulation during both the daytime and nighttime for the two seasons (LST RMSE: daytime: 9.77 compared to 7.18 K in winter, 5.03 compared to 4.76 K in summer; nighttime: 4.85 compared to 4.20 K in winter, 6.99 compared to 4.24 K in summer; Table 3). After modification, lower soil thermal conductivity permitted less energy transfer from the land surface to deep soil during daytime, resulting in more sustained energy on the surface, producing warming. During the night, less energy transferred from the deep soil to the land surface, leading to a reduction in LST. As a result, the introduction of a more realistic representation of soil thermal conductivity into CLM5.0 significantly improved LST simulations during day- and nighttime in the two seasons.

Impact of soil evaporation resistance on LST

Next, we further evaluated and improved the performance of CLM5.0 on simulated ET and surface SM during summer. Figure 5A plots the spatial distribution of seasonal ET differences between EXP2 and GLEAM_ET data averaged over the summers of 2003-2018. EXP2 underestimated summer ET for almost the entire TP, except for the Qaidam Basin and western TP, with an average bias of −46 mm/season (although simulated summer ET showed a similar spatial distribution patterns to those from GLEAM_ET data; PCC = 0.86). Such underestimation of ET leads to wetter soil in 87.6% of TP regions, with an average bias value of 0.08 mm³ mm⁻³ for SM when compared with GLEAM_SM data (Fig. 5B) because soil evaporation acted as the mainly component of ET (LAI values were generally well below 1.0 in the central and western parts of TP; Fig. S6). In addition, we found that summer ET was underestimated and SM was overestimated with EXP2 compared to other remote sensing observations (Figs. S7 and S8). The underestimation of ET (soil evaporation) and the prediction of wetter surface soil were improved by the modification of the soil evaporation resistance scheme as described in ‘Modification of soil evaporation resistance scheme’. We compared the ET simulated with EXP3 to that with EXP2 and to observed GLEAM_ET data. EXP3 generated more seasonal summer ET across the TP (red areas in Fig. S9B) than EXP2, especially in central TP. Moreover, seasonal summer ET estimated in EXP3 agreed better with GLEAM_ET data (Fig. 5C), with a smaller ET average bias (−18 mm/season) and larger ET PCC (0.88). The larger seasonal summer ET (soil evaporation) from EXP3 led to less soil water content when compared with the SM from EXP2 (blue parts in Fig. S9C), which was closer to the GLEAM_SM data (Fig. 5D), as EXP3 produced smaller SM average bias (0.02 mm³ mm⁻³) and larger SM PCC (0.29 larger than EXP2).

The larger seasonal summer ET from EXP3 implied that the lower daytime LST was due to the evaporation cooling effect when compared with daytime LST from EXP2 (Fig. S10B), being closer to MODIS daytime LST during summer, with an RMSE of 4.30 K (Fig. 6B). The lower SM from EXP3 corresponded to a smaller soil thermal inertia, implying less heat transfer from deep soil to the land surface during nighttime, producing lower nighttime LST when compared with that from EXP2 (Fig. S10D), being closer to MODIS nighttime LST during summer, with an RMSE of 3.40 K (Fig. 6D). In addition, the modification of the soil evaporation resistance scheme had a negligible impact on winter ET (Fig. S9A), but the lower soil liquid water content simulated by EXP3 in summer led to less soil ice content, and then to lower soil thermal conductivity in winter. Thus, EXP3 predicted higher daytime LST and lower nighttime LST during winter when compared to that of EXP2 (Figs. S10A, S10C). Furthermore, EXP3 produced better daytime and nighttime LST in winter when compared with MODIS observations (Figs. 6A, 6C; LST RMSE: 6.47 compared to 7.18 K in daytime, 3.80 compared to 4.20 K in nighttime).

Table 3 summarizes the LST average bias and RMSE between the four offline CLM5.0 simulations and MODIS/Aqua observations averaged over the TP during daytime and nighttime. Adjustments to these three schemes (ground sensible heat roughness length, soil thermal conductivity, and soil evaporation resistance) improved the simulations of the diurnal LST cycle for the TP during winter and summer.