Summary: The 2m temperature (T2m) in weather prediction models is a diagnostic variable, derived from the variables of the model’s surface and near surface layer. By construction, it reflects the assumptions of the surface scheme and therefore it is subject to representation errors when compared to other models or observations. In the Copernicus pan-Arctic Regional Reanalysis (CARRA2), this problem is more complex because of the sophisticated surface scheme, which explicitly accounts for vegetation canopies. In grid boxes dominated by high vegetation (like forests), the diagnostic T2m represents conditions above the explicit canopy rather than near the ground. This can lead to unusually warm values in forested regions such as Siberia or the Sahtu area in Canada. As a result, intercomparisons with other datasets may give the impression of poor CARRA2 performance, when the differences are primarily due to the definitions of the T2m diagnostics and not the poor performance of the model. This guide documents this issue.

1 Introduction

The 2m temperature (T2m) provided by Numerical Weather Prediction (NWP) and reanalysis systems is not a model prognostic variable but a diagnostic field. It is generally obtained by interpolating between the model’s lowest atmospheric level and the surface using stability functions from the surface-layer scheme of the model. By design, such diagnostics are sensitive to model-specific assumptions and tuning, such as artificially enhanced roughness lengths or atmospheric stability introduced to improve flux calculations (Kähnert et al. 2025). This means that even if two models have identical surface and atmospheric air temperatures, they may still yield different T2m values simply because of differences in interpolation methods or surface parameterizations. We call this a representation error. Any intercomparison of T2m between different models or between a model and an observation is therefore subject to such an error.

Due to the “relative simplicity” of surface schemes in most models, these representation errors are usually minor and overshadowed by more fundamental deficiencies, such as for instance biases in the snow scheme (Day et al. 2020), excessive turbulent mixing (Holtslag et al. 2013), or flaws in radiation parameterizations (Edwards 2009). However, for CARRA2 with a sophisticated surface parameterization scheme, this representation error can become the largest error source over forest dominated areas due to CARRA2’s explicit representation of the canopy (Le Moigne 2009, Boone et al. 2017, Napoly et al. 2017). Please note that in the original Copernicus Arctic Regional Reanalysis (CARRA) this problem was not present due to its much simpler surface scheme.

In order to demonstrate the problem Figure 1, shows a domain-wide intercomparison between CARRA2 and ERA5. In certain areas CARRA2 exhibits pronouncedly warmer T2m compared to ERA5, with differences between ERA5 and CARRA2 well exceeding 10K in some locations.

Figure 1: T2m differences between CARRA2 and ERA5 for the months of December, January, and February averaged over 7 winters: 1990/91, 1995/96, 2000/2001, 2005/06, 2010/11, 2015/16 and 2020/21. a ) Over the entire domain b),c) over specific sub-areas in Siberia. (These small areas are on the upper right side of the full domain in panel a.)

An inspection of the model physiography reveals that these warm anomalies coincide with grid boxes dominated by forests (compare Fig. 1b,c to Fig. 2a,b respectively), emphasizing that they arise from differences in how T2m is represented over high vegetation in ERA5 versus CARRA2.

Figure 2: Area fraction of the boreal needleleaf deciduous cover type. a) the region highlighted in Fig. 1b, b) the region highlighted in Fig. 1c. 

2. The “warm” T2m

As mentioned in the introduction T2m in CARRA2 is a diagnostic variable. That means that T2m is not evolved in time by prognostic model equations, like e.g. surface or atmospheric temperatures at model levels (prognostic variables), but is diagnosed at every time step based on a relation to the prognostic variables. In CARRA2 we use a formulation following Geleyn (1988):

where z is the requested height (here 2m), Ts denotes surface temperature, Ta atmospheric temperature (available as prognostic variable at lowest model level), and si is an interpolation factor dependent on z, the roughness length for momentum z0, the roughness length for heat z0H, and Ri, the bulk Richardson number:

Thus, T2m is obtained by a stability-dependent interpolation between the prognostic temperature at the lowermost model level and the surface temperature. It is this surface temperature, or better to say, the choice of what is used as surface temperature in Equation 1 that causes the warm temperatures over forest dominated areas. 

Figure 3: Schematic representation of the nature tile in CARRA2. Graphic adapted from Kähnert et al. (2023). 

SURFEX (Le Moigne, 2009) is the surface model applied in the CARRA2 system. In total four tiles (representing fractions of a grid box area) are employed by SURFEX: sea, inland water, nature and town. Forests, or high vegetation, are calculated as part of the nature tile in SURFEX. The nature tile represents bare-soil processes, as well as low and high (forest) vegetation processes, while also accounting for partial or full snow cover (Figure 3). Individual state variables are calculated for each of these cover types, among them the canopy air temperature (Tc) for high vegetation. The canopy air temperature represents a bulk temperature for the entire high vegetation canopy layer, which is explicitly separated from the soil surface as illustrated in Figure 3 (Boone et al. 2017). 

For every patch or tile within a grid box, an individual T2m​ is computed using its corresponding temperature as Ts in Equation 1, with the high-vegetation patch specifically relying on Tc​. The resulting estimates are subsequently combined into a grid-averaged T2m​ by means of area-weighting. In grid boxes dominated by forest, the mean T2m​ is strongly influenced by the high-vegetation contribution based on Tc​. Often, the Tc​ values do not deviate substantially from the lowest model level atmospheric temperature, Ta​. As an illustration, Figure 4 shows the evolution of the tile-specific Ts and corresponding T2m during a stable weather period. The surface temperatures for the bare-soil and low-vegetation tiles (black lines) exhibit pronounced cooling, resulting in strong surface-layer inversions exceeding 15 K (Fig. 4b,c). According to Equation 1, the associated T2m values (red lines) are similarly cold and closely follow the evolution of their respective surface temperatures. In contrast, the surface temperature of the high-vegetation tile remains considerably warmer, producing a much weaker inversion of about 5 K (Fig. 4d). Consequently, the corresponding T2m is notably warmer and more consistent with the evolution of the atmospheric temperature (blue line), owing to the reduced stability (Ri, see Eq. 2). Depending on the fractional coverage of each tile within a grid box (Fig. 4a), the grid-averaged T2m may therefore be strongly influenced by the high-vegetation component, resulting in comparatively warm overall values.

Figure 4: a) Schematic representation of tile fractions in a grid box, with red shading indicating tendencies for warmer or colder tile-specific surface temperatures. b)-d) Illustrative, tile-specific (bare soil, low vegetation and high vegetation) temperature evolutions during a stable boundary layer period for surface temperatures (black), 2m-temperatures (red), and atmospheric temperatures (blue). 

It becomes clear that the representation of “2m temperature” is shaped by these formulations. If we picture a grid box fully covered by forest, T2m always describes “2m temperature above the canopy”. However, in a model without an explicit representation of the canopy the corresponding canopy air temperature Tc does not substantially differ from the surface temperature. In the operational NWP configuration of the HARMONIE-AROME model (using a surface scheme without explicit canopy representation), for example, the temperature differences between high and low vegetation during cloud free and very stable nights are between 1-3 K. Thus any representation error is small compared to the differences of more than 15 K that can appear in CARRA2. So while the more sophisticated representation of the surface in CARRA2 yields us with better surface-air interactions in form of fluxes, a proper representation of snow and an explicit handling of the energy budget within the surface layer, it simultaneously emphasizes the representation error in T2m.

Hence, any comparison between CARRA2 and another model such as ERA5 will yield substantial differences in T2m over forest dominated areas. Similar differences will occur when one would compare an independent observation typically taken 2m above ground with CARRA2 in such areas (the practice for observing T2m is to avoid forest and high vegetation areas). Yet, it is clear that many users require a T2m that is more similar to an observation taken 2m above ground. Ideally, users need to work directly with tile-specific near-surface temperatures, as these represent the physically consistent outputs of the land-surface scheme. For example, the aforementioned T2m similar to an observation would be represented by the T2m of the low vegetation patch. Unfortunately, such variables are not routinely available for the initial release of the CARRA2 data stream but will be published at a later stage (around mid-2026).

3. Conclusion and general considerations

This user guide highlights that differences in T2m diagnostic formulations play a decisive role in shaping its values over forested regions. In CARRA2, the explicit treatment of high vegetation introduces a more realistic surface–atmosphere coupling, but it also changes the meaning of T2m compared to other datasets or observations that lack a canopy representation. Over forest-dominated areas, T2m effectively reflects conditions above the canopy rather than near the ground, leading to systematically warmer values under stable and cloud-free conditions.

These differences should not be interpreted as a deficiency in CARRA2, but rather as a reflection of its sophisticated surface parameterisation. When compared against models such as ERA5 or against near-surface observations, the apparent warm bias primarily arises from this representational mismatch.

Until tile-specific T2m outputs become available in future CARRA2 releases, users are urged to treat areas dominated by high-vegetation with care. Doing so will help ensuring that comparisons across datasets and with in-situ measurements remain physically consistent and scientifically meaningful.

References 

Boone A, Samuelsson P, Gollvik S, Napoly A, Jarlan L, Brun E, Decharme B (2017) The interactions between soil–biosphere–atmosphere land surface model with a multi-energy balance (ISBA-MEB) option in SURFEXv8—Part 1: model description. Geosci Model Dev 10(2):843–872 

Day JJ, Arduini G, Sandu I, Magnusson L, Beljaars A, Balsamo G, Rodwell M, Richardson D (2020) Measuring the impact of a new snow model using surface energy budget process relationships. J Adv Model Earth Syst 12(12):e2020MS002,144

Edwards JM (2009) Radiative processes in the stable boundary layer: Part I. Radiative aspects. Boundary-Layer Meteorol 131(2):105

Geleyn J.-F. (1988) Interpolation of wind, temperature and humidity values from model levels to the height of measurement. Tellus, 40A(4):347–351

Holtslag AAM, Svensson G, Baas P, Basu S, Beare B, Beljaars ACM, Bosveld FC, Cuxart J, Lindvall J, Steeneveld GJ, Tjernström M, Wiel BJHVD (2013) Stable atmospheric boundary layers and diurnal cycles: challenges for weather and climate models. Bull Am Meteorol Soc 94(11):1691–170

Kähnert, M., Sodemann, H., Remes, T.M. et al. (2023) Spatial Variability of Nocturnal Stability Regimes in an Operational Weather Prediction Model. Boundary-Layer Meteorol 186, 373–397, https://doi.org/10.1007/s10546-022-00762-1

Le Moigne P (2009) SURFEX scientific documentation. CNRM Tech Rep p 211 pp

Napoly A, Boone A, Samuelsson P, Gollvik S, Martin E, Seferian R, Carrer D, Decharme B, Jarlan L (2017) The interactions between soil-biosphere-atmosphere (ISBA) land surface model multi-energy balance (MEB) option in SURFEXv8—Part 2: introduction of a litter formulation and model evaluation for local-scale forest sites. Geosci Model Dev 10(4):1621–1644