Is land surface processes representation a possible weak link in current Regional Climate Models?

We evaluate reanalysis-driven RCM simulations performed as part of the EURO-CORDEX project and downloaded from the Earth System Grid Federation (ESGF) archive. The nine RCMs considered (table 1) provided simulations at 50 km (0.44 degree on a rotated grid) spatial resolution on a common analysis domain encompassing Europe in its entirety. The 6-hourly ERA-Interim reanalysis (Dee et al 2011) is used in all models to prescribe lateral boundary conditions and sea surface temperatures. The longest period common to all models (1990-2008) is used for the analyses.

In addition to the aforementioned official EURO-CORDEX simulations we also analyse a simulation performed with COSMO-CLM². COSMO-CLM² is an alternative configuration of COSMO-CLM featuring a different LSM.

COSMO-CLM is a non-hydrostatic regional atmospheric model jointly developed by the COnsortium for Small-scale MOdelling (COSMO) and the Climate Limited-area Modelling Community (CLM-Community) and is one of the participating EURO-CORDEX RCMs (table 1). In its standard configuration, COSMO-CLM includes TERRA_ML as its LSM. In COSMO-CLM², however, TERRA_ML is replaced by the more complex Community Land Model (CLM). Earlier versions of COSMO-CLM² where based on CLM3.5 coupled as a sub-routine to COSMO-CLM (Davin et al 2011, Davin and Seneviratne 2012). Here we use the more recent version CLM4.0 (Oleson et al 2010, Lawrence et al 2011) coupled to COSMO-CLM via the OASIS3-MCT coupler (Valcke et al 2013).

The main conceptual differences between TERRA_ML and CLM4.0 concern both biogeophysical and hydrological processes. Unlike in TERRA_ML, an explicit canopy layer is considered in CLM4.0, resulting in specific vegetation temperature and fluxes. The linkage between transpiration and photosynthesis is considered in CLM4.0 while an empirical formulation of stomatal conductance is used in TERRA_ML. Sub-grid scale surface heterogeneity is ignored in TERRA_ML and is represented using a tile approach in CLM4.0. CLM4.0 additionally considers groundwater and calculates runoff taking into account sub-grid scale topographic heterogeneity using a TOPMODEL-based approach. A more complete description of CLM4.0 and its input datasets is provided in Oleson et al (2010). In the present study, CLM4.0 is used in its biogeophysics-only configuration without carbon and nitrogen dynamics. The only modification we included to CLM4.0 concerns two hydrological parameters influencing surface and subsurface runoff (i.e., exponential decay factor influencing the saturation excess component of surface runoff and maximum subsurface drainage). Namely, we reverted to the values used in CLM3.5 for these parameters (Lawrence et al 2011) since preliminary tests indicated slightly more realistic evapotranspiration rates over Europe for this parameter choice.

The simulation performed with COSMO-CLM² follows the EURO-CORDEX protocol and is a sister simulation of the one performed with COSMO-CLM (table 1). That is, the same model version and model parameter set are used for the atmospheric component, the only difference being the LSM used. In doing so, comparing the COSMO-CLM and COSMO-CLM² simulations strictly isolates the effect of land processes representation, all else being identical.

The various reference products used for evaluation are described in table 2. The selected products cover temperature, precipitation and surface heat and radiation fluxes. When possible, different products are considered for a given variable in order to account for uncertainties in observation-based datasets. All the products are used at a monthly resolution and were regridded, using bilinear interpolation, to a common half-degree regular grid for comparison with the EURO-CORDEX models. For products not covering the full 1990-2008 EURO-CORDEX common analysis period a shorter time period is used instead.

In this section, we evaluate overall RCM performance using synthetic scores applied over the entire European continent. Model performance with respect to temperature, precipitation and surface fluxes is assessed for each of the EURO-CORDEX RCMs (figure 1). As in Davin and Seneviratne (2012), a root-mean-square error (RMSE) is calculated at all grid cells based on monthly values over a multi-year period (period depending on reference product, see table 2) thus integrating both temporal and spatial performance. When possible, different reference products are used for a given variable, because the choice of reference product is likely to have a major influence on the inferred scores beyond all other methodological choices (Schwalm et al 2013).

Zoom In Zoom Out Reset image size

Considering first surface fluxes, the coupling with CLM4.0 dramatically improves the performance of the standard COSMO-CLM (figure 1(a)). COSMO-CLM² outperforms not only COSMO-CLM but also most other EURO-CORDEX RCMs for both radiative and turbulent fluxes. In some cases (e.g. for evapotranspiration) COSMO-CLM² also outperforms the EURO-CORDEX multi-model mean (MMM). This result highlights the added value of using a more advanced LSM compared to the simpler schemes commonly used in current RCMs. In line with Best et al (2015), we also note that most models typically have larger errors for sensible heat flux than for latent heat flux.

As a consequence of the better representation of surface fluxes, 2-meter temperature is also better simulated in COSMO-CLM² (figure 1(b)), which outperforms the standard COSMO-CLM as well as most other RCMs. The improvement is particularly substantial for maximum temperature (monthly average of daily maximum temperature) indicating that the representation of land processes influences daytime more than nighttime conditions. While the Diurnal Temperature Range (DTR) is also improved in COSMO-CLM², it is interesting to note that absolute errors are typically higher for DTR than for other metrics (e.g. Tmax) indicating that the representation of the diurnal cycle is a critical remaining deficiency in current RCMs.

No improvement, however, is seen for precipitation which is even slightly degraded (figure 1(c)). In view of the multi-model spread this degradation remains relatively minor as COSMO-CLM and COSMO-CLM² still cluster together in terms of ranking. This degradation for precipitation might seem counterintuitive given that both surface temperature and surface fluxes are generally improved in the model. In this respect, we note that atmospheric parameters in the model have been tuned in the context of TERRA_ML and not of CLM4.0. We therefore expect that a retuning would be necessary to obtain optimal performances in particular in terms of precipitation. This is, however, not the scope of this study as a different atmospheric setup would make the attribution of assessed differences to land processes more difficult.

For comparison purpose, we also performed the same multi-variate ranking procedure including in addition an earlier version of COSMO-CLM² evaluated in Davin and Seneviratne (2012) and based on CLM3.5 instead of CLM4.0 (figure S1). The coupling with CLM3.5 already improves overall model performance compared to COSMO-CLM, while switching to CLM4.0 provides further improvements compared to CLM3.5 in line with global offline evaluation results (Lawrence et al 2011).

The only difference between COSMO-CLM² and COSMO-CLM being the LSM used, the overall better performance of COSMO-CLM² can be attributed to the representation of land processes. This interpretation is further supported by the more realistic surface fluxes simulated in COSMO-CLM² as shown in this section. In the next section, we examine more specifically the nature of the model biases to better characterize the mechanisms at play.

The sign and magnitude of model biases is generally not uniform across Europe and in particular a North-South contrast is visible for temperature (figure 2), radiation (figure 3) and other variables (Supplementary figures). For this reason, we focus our analyses on two regions representative of Southern and Northern Europe (boundaries displayed in figure 2) in addition to the bias maps provided in figures 2, 3 and in the Supplementary Information.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Over Northern Europe, most RCMs underestimate temperature in particular in spring and summer (figure 4(a)), a tendency mostly affecting daytime temperatures (figures S2–4). This bias reflects a systematic underestimation of surface shortwave radiation (figure 3; figure S6) in turn linked to a tendency of the RCMs to overestimate cloud cover (figure S8). We note an inter-model correlation between summer temperature and net shortwave radiation of 0.7 over Scandinavia, thus confirming that the cause of temperature biases in this region is essentially of radiative origin. While atmospheric processes are obviously critical for cloud cover biases, land processes may also play a role by indirectly modulating cloud formation through energy partitioning at the surface (Davin et al 2011). All the EURO-CORDEX models in fact overestimate summer evaporative fraction over the northern half of Europe (figure S11). COSMO-CLM² alleviates this problem compared to COSMO-CLM which implies reduced water input to the atmosphere with beneficial effects on simulated cloud cover and net shortwave radiation (figure 3), as previously found in earlier model versions (Davin et al 2011, Davin and Seneviratne 2012).

Zoom In Zoom Out Reset image size

In contrast, warm biases dominate in summer over Southern Europe (figure 2; figure 4(b)) in conjunction with an overestimation of interannual summer temperature variability (figure 5(c)). This persistent deficiency has been reported in previous RCM intercomparisons and attributed to excessive summer drying (Hagemann et al 2004, Christensen et al 2007, Hirschi et al 2007, Vautard et al 2013, Kotlarski et al 2014). Most EURO-CORDEX RCMs indeed strongly underestimate summer evapotranspiration over the Mediterranean region (figure 5(a)) and the magnitude of this underestimation correlates well with temperature biases across models (figure 5(c)). Both the evapotranspiration bias and the resulting temperature bias are largely alleviated in COSMO-CLM² compared to COSMO-CLM and other RCMs confirming that land processes representation plays a major role in this long-standing deficiency. One possible factor in this improvement is that CLM4.0, unlike the other models considered here, includes a representation of groundwater which can limit the excessive summer drying. Another hypothesis is that COSMO-CLM, as most other RCMs, generally overestimates evapotranspiration when water is not limited (this is for instance the case in the spring for Southern Europe but this happens also more generally over Northern Europe as mentioned previously), thus leading to depleted water conditions later in summer. COSMO-CLM² exhibits a more conservative water use behaviour in the spring letting more water available for transpiration in the following summer. This might have a physical cause (e.g. higher aerodynamic resistance) or a physiological cause linked to the explicit link between transpiration and photosynthesis represented in CLM4.0.

Zoom In Zoom Out Reset image size

Another aspect playing a role in seasonal temperature variations over southern Europe is the representation of ground heat flux (GHF). Most models tend to overestimate the annual temperature range, with too low temperatures in winter and the opposite in summer (figure 4(b)). This problem is notably reduced in COSMO-CLM², which strikingly also exhibits a larger GHF annual amplitude (figure 5(b)). A larger amplitude means that more energy is stored into the ground during summer and subsequently more energy can be released to the atmosphere the following winter. This results in a dampened annual cycle of temperature as the additional energy stored into the ground cannot be used to warm near-surface air in summer but can be released in winter and limits then the winter cooling. The deeper bottom boundary condition (42 m) for thermal calculations in CLM4.0 compared to other models (usually not more than 10 meters) results in a larger soil volume and heat storage capacity that can explain the larger annual range in GHF. Supporting this interpretation of the important role of GHF, we also find a relatively good inter-model relationship between the simulated annual temperature range and GHF (figure 5(d)). In other words, models with low GHF annual amplitude tend to overestimate more the annual temperature range. Previous studies already highlighted the importance of placing the bottom boundary condition for thermal calculations much deeper than 10 metre to adequately represent GHF and soil temperature dynamics over seasonal and decadal time scales (Smerdon and Stieglitz 2006, MacDougall et al 2008, 2010).