The diagnostic model used in this study is based on the algorithms applied in the LSA SAF products. The LSA SAF algorithm to simulate surface turbulent energy fluxes was developed in the framework of the EUMETSAT deployment of “Satellite Applications Facilities” (SAFs; https://www.eumetsat.int/about-us/satellite-application-facilities-safs, last access: 15 July 2022) and is used to generate LSA SAF ET, LE and H products operationally (i.e in near-real time). It is a soil vegetation atmosphere transfer (SVAT) model, largely driven by remote-sensing-based observations of downwelling long- and shortwave radiation, LAI, and albedo. It relies on the approach to calculate the stomatal response to environmental factors.

In operational mode, the observations of the Spinning Enhanced Visible and Infrared Imager (SEVIRI) on board the Meteosat Second Generation Satellite (MSG) are the primary source of the forcing variables. A more in-depth outline of the algorithm is given in . Consequently, it was designed to run at the resolution of MSG observations, but its capabilities at the sub-kilometre scale were recently demonstrated . For this study, the model was configured to run at the kilometre scale (i.e. the local scale corresponding to the footprint of eddy covariance measurements), using LAI from the European Copernicus Global Land Service (CGLS) and soil moisture from ERA5.

More recently, a LSA SAF GPP product was developed, based on the Monteith light use efficiency (LUE) concept . This product is calculated at the end of the LSA SAF pipeline, as it relies on several other LSA SAF products, such as ET, reference ET, LAI and FAPAR. The same formulation was adopted in the diagnostic model in our study, resulting in coherent surface fluxes.

Contrary to ISBA and ORCHIDEE, the calculations for LE and GPP in the diagnostic model do not share parameters like stomatal resistance. Instead, the GPP calculations are coupled to LE by using the actual evapotranspiration as an input variable.

Within the Surfex (SURFace Externalisée) land surface model, ISBA (Interactions between Soil, Biosphere and Atmosphere) is the component dedicated to modelling the exchange of water, energy and carbon fluxes between the soil–vegetation–snow continuum and the atmosphere . In this case, a configuration of ISBA with interactive carbon cycling is used, i.e. ISBA-CC . The fluxes of water and carbon from the vegetation are coupled through the stomatal resistance. This shared parameter is calculated through the A-gs surface scheme and largely depends on soil moisture stress and air temperature . The parametrization for this scheme is based on plant traits derived from the TRY database .

The assimilation of carbon results in the evolution of LAI through a biomass allocation scheme. The growth and senescence of leaves is purely photosynthesis-driven. The biomass reservoirs are coupled to a soil organic matter module to calculate the respiration terms.

The simulations with ISBA were performed on the Surfex v8.1 platform (https://www.umr-cnrm.fr/surfex/, last access: 15 July 2022). The soil profile was discretized in 14 layers (up to 12 m depth), using a diffusion scheme for soil heat and water transfer and an exponential decrease in hydraulic conductivity through the profile. The nitrogen dilution scheme and canopy radiation transfer scheme were enabled. In the forest patches, the energy fluxes were calculated with the recently developed multi-energy balance scheme MEB;. Contrary to the standard soil–vegetation composite version of ISBA (which was used for the non-forest patches), MEB explicitly solves the transfer of mass and energy between the soil surface, the snowpack, the canopy and the atmosphere. At the time of this study, the combination of MEB and prognostic LAI modelling is still considered experimental . A spin-up period of 3 years was sufficient to eliminate effects from the initial model state on the surface fluxes (respiration is not analysed in this study). ISBA was not coupled to a hydrological model (e.g. CTRIP; ). Consequently, there was no lateral groundwater flow or a water table, only free drainage at the bottom of the soil profile.

ORCHIDEE is the land surface model of the Institut Pierre Simon Laplace (IPSL) earth system model and was initially described in . We used the version prepared for the 6th Coupled Model Inter-comparison Project (CMIP6) .

The LAI is prognostic, and the phenology models used for the various plant functional types (PFTs) are described in and . The canopy is discretized in layers of increasing thickness from the top to the bottom of the canopy. The incoming light is attenuated through the canopy following a Beer–Lambert extinction law. The photosynthesis is modelled at the leaf level following for C3 species and for C4 species. The maximum carboxylation rate at 25 ∘C is a PFT-dependent parameter. The maximum carboxylation rate varies with the temperature following and . A water stress function depending on soil moisture and root profile is applied to the maximum carboxylation rate and the stomatal and mesophyll conductances. An analytical solution to the three equations linking CO2 assimilation, stomatal conductance and CO2 leaf intercellular concentration is computed following . The assimilation is then upscaled over the layers to calculate the GPP.

A single-layer energy balance is computed per grid cell. LE is the weighted average of the snow sublimation, the soil evaporation, the canopy transpiration and the evaporation of foliage water; all these terms were initially computed following . The soil is now discretized over 2 m into 11 layers of increasing thickness, and the hydrology scheme follows Richard's equation . There is free drainage at the bottom. The soil thermodynamics are described in , and the snow scheme is detailed in .

To initialize the simulations, a first spin-up phase was performed, where we cycled over the available FLUXNET years for at least 45 years. This enables an equilibrium to be reached for the aboveground biomass and the water stocks and fluxes, as an initial state for the transient simulation.

The capability of the models to reproduce the timing of the seasonal cycle of LE, GPP and LAI was evaluated. The detection of the start, maximum and end of the seasonal cycle (SOS, MOS and EOS) was achieved by applying a smoothing operation (20 d rolling mean), followed by a threshold procedure . In this threshold procedure, the minima and maxima were used to delineate the growing and senescent phase of the season. MOS was defined as the date when the maximum of the season is reached; SOS and EOS were defined as the date where the growing or senescent phase crosses the threshold value T. T was calculated for each growing or senescent phase as T=P5+0.2(P95-P5), where P5 and P95 are the 5th and 95th percentile.

To compare the model dynamics, the simulated LE flux partitioning, water balance and water use efficiency (WUE) were evaluated as well. Direct observations of the LE flux partitioning were not available, but it is possible to extract the transpiration component from the total LE flux, using the underlying water use efficiency (uWUE) method . From the GPP and transpiration (Tr), the WUE was derived: 6WUE=GPPTr.

To assess the sensitivity of the fluxes to the state variables (Se and LAI), the slope of the seasonal anomalies of the fluxes against the anomalies of the state variables was determined. This analysis was performed for the observations and the simulations and compared. Note that the linear slope was used here, though a linear response is not necessarily expected (e.g. the response to soil moisture anomalies depends on a wet/dry regime). The goal of this analysis was to investigate whether LSMs are capable of reproducing a similar relationship as found in the observations. Significant differences between the models were evaluated with the Wilcoxon signed-rank test.

To evaluate whether errors in the state variables result in errors in the surface fluxes (or vice versa), the Spearman rank correlation between both was calculated. Since Copernicus LAI was the reference LAI, this analysis was not possible for LAI in DiagMod.

The diagnostic model is a suitable vehicle to test the impact of the prognostic state variables from ISBA and ORCHIDEE on the surface fluxes. Given its architecture to easily ingest state variables, it can serve as an independent model platform to evaluate the quality of the soil moisture and LAI. DiagMod simulations were performed using soil moisture and/or LAI from ISBA and ORCHIDEE and compared to simulations with soil moisture from ERA5 and CGLS LAI (resulting in seven runs per site; see Table ).

The fraction of absorbed photosynthetically active radiation (FAPAR) is an important variable in DiagMod to produce GPP, but it is no output of the prognostic models. In order to be consistent with the prognostic LAI, FAPAR was estimated using a simple Beer law with a general-purpose extinction coefficient value of 0.5 (Eq. ; ).

7FAPAR=1-exp⁡(-0.5LAI) The soil moisture of the soil profiles in the prognostic models was integrated to match the four layers in DiagMod (0–7, 7–21, 21–72, 72–189 cm). Furthermore, the soil moisture was rescaled using the wilting point and field capacity parameters of the models.

Prior to the evaluation of the prognostic state, the reproducibility of the prognostic models by the DiagMod was tested. The detailed results are shown in the Supplement. It was found that the surface fluxes produced by DiagMod, forced by the same atmospheric conditions, soil moisture and LAI, were more closely correlated to those from ISBA compared to ORCHIDEE. Differences can be caused by different parametrization of the plant physiology, as well as the representation of processes (or lack thereof), such as rainfall interception, snow cover or canopy radiation transfer.

The bias (ME) and accuracy (RMSE) of the simulated LE and GPP are shown in Fig. , together with Taylor diagrams of the simulated fluxes and their seasonal anomalies. It was evident that the inter-site variability in the model performance is much larger than the inter-model variability. In terms of bias and accuracy, the differences between the models were relatively limited. All models suffered a substantial underestimation of LE, whereas the overall bias in GPP was relatively small. Significant differences (Wilcoxon p<0.05) were found in the bias of GPP between DiagMod (overestimation) and ISBA (underestimation), and the simulated LE was significantly more accurate in ISBA compared to ORCHIDEE.

Notably, no substantial bias was found in the simulated H of any model to compensate for the consistent bias in LE (results shown in the Supplement). In this study, the corrected fluxes from the FLUXNET/ICOS dataset were used as a reference. If the non-corrected fluxes were used instead, the bias in LE was reduced, but the simulated H was overestimated (not shown here). This points at the significant uncertainty associated with the observed fluxes from eddy covariance measurements. The estimated observation uncertainty in the turbulent fluxes (associated with random measurement errors and energy balance correction) had the same order of magnitude as the model errors.

The Taylor diagrams in Fig.  show that the average variability in the simulated LE and GPP was in fair agreement with the observations. After removal of the mean seasonal cycle, the performance of the models decreased (rANOM, NSANOM), but the mean variability in the anomalies is reasonably accurate. In terms of r and rANOM of LE and GPP, ORCHIDEE was significantly outperformed by ISBA and DiagMod (Wilcoxon p<0.05). No significant differences were found between ISBA and DiagMod.

The impact of the land cover type of the test site on the model performance is illustrated in Fig. . Here, the test sites are classified by the dominant vegetation type. The NS and NSANOM of the simulated LE were not significantly impacted in any of the models, whereas a significant influence (Kruskal p<0.05) was found on the quality of the simulated GPP in DiagMod and ORCHIDEE. The NS and NSANOM of the simulated GPP in ORCHIDEE were significantly better (Mann–Whitney U p<0.05) for forest sites compared to sites that were dominated by herbaceous vegetation. Inversely, the simulation of the seasonal GPP anomalies in DiagMod were significantly better at herbaceous test sites (Mann–Whitney U p<0.05). No significant impact was found in the ISBA simulations. Notably, the differences between the models were most pronounced at the herbaceous sites (see also Table ). Yet, despite its poorer performance at the herbaceous sites, ORCHIDEE simulated GPP at forest sites most accurately compared to the other models.

Similar results were found with the other validation indices. A more detailed breakdown of the results per PFT and HCB is given in the Supplement. A significant impact of PFT and HCB on the NS of the simulated GPP (Kruskal p<0.05) was found in all models. This was contrasted by LE, where a significant impact of HCB (Kruskal p<0.05) was found only for ORCHIDEE.

The validation results of Se and LAI are shown in Fig. . The soil moisture from ERA5 (used in DiagMod) tended to be overestimated compared to in situ observations, whereas an overall negative bias was found in ISBA and ORCHIDEE. The simulated variability in soil moisture was too low in all models, in particular for ORCHIDEE. Notably, ERA5 outperformed ISBA (p>0.05) and ORCHIDEE (p<0.05) in terms of accuracy, despite their use of in situ meteorological forcings (e.g. precipitation). ORCHIDEE performed significantly worse than the other two models for all validation metrics (Wilcoxon p<0.05). The highest correlation in the anomalies was simulated by ISBA.

Compared to the surface fluxes, the accuracy of the simulated soil moisture was substantially lower. The validation scores of Se are given in Table , separated per dominant land cover type. In all models, the simulated Se was significantly better for herbaceous sites compared to forest sites. The herbaceous sites are generally found in a water-driven dryland climate, with strong precipitation-driven anomalies.

Similarly, the prognostic LAI was also of poorer quality than the simulated surface fluxes. ISBA had a significantly better ME and RMSE than ORCHIDEE, but both models overestimated LAI and strongly underestimated its variability. In particular, the variability in LAI in the evergreen needleleaf forests was strongly underestimated in both models, as well as the variability in LAI in evergreen broadleaf forests in ORCHIDEE. Furthermore, both models obtained only a poor correlation and achieved a very poor correlation of the seasonal anomalies. In both models, the simulated LAI for forest sites was better than for the herbaceous sites, though not significantly (p>0.05). The simulated anomalies were modelled significantly better (p<0.05) at forest sites than herbaceous sites (Table ).

The timing of the start, maximum and end of the seasonal cycle was validated for LE, GPP and LAI. Figure  shows the boxplots of the mean errors at all sites. In all models, the bias and accuracy of the seasonality of LE and GPP were comparable, whereas the leaf phenology (i.e. LAI) was poorer. The simulated phenology of LAI was delayed substantially, in particular in ISBA. This bias was most pronounced by the MOS, and to a lesser extent in EOS.

ISBA performed significantly worse than ORCHIDEE (Wilcoxon p<0.05) for ME of MOS GPP and MOS LAI and RMSE of MOS LAI. The prognostic LAI in both models tended to peak towards the end of the growing season, whereas the maximum LAI was reached in the beginning of the season according to the observations. This is illustrated in Fig. , where the mean annual LE, GPP and LAI cycles of ENF and DBF sites are shown. The delayed GPP phenology in ISBA is a feedback effect of the delayed prognostic LAI. However, the effect is dampened since GPP is largely driven by atmospheric forcings as well.

At forest sites, EOS of LAI tended to be simulated with the highest accuracy. The phenology of herbaceous sites had a higher variability (median standard deviation of EOS LAI at forest sites was 7.7 d compared to 20.6 d at herbaceous sites), which turned out to be challenging to capture for ISBA and ORCHIDEE. An example is shown for the savanna sites in Fig. . DiagMod relied on the remote-sensing-based LAI and was significantly more accurate than the prognostic models in capturing EOS of GPP (Wilcoxon p<0.05).

As the models were configured to run without dedicated management practices for the crop sites, EOS was estimated too late due to the harvest practice (Fig. ). Even in DiagMod, EOS of GPP was delayed.

The water balance partitioning in ISBA and ORCHIDEE is shown in Fig. . In both models, the evapotranspiration fraction across PFTs was similar, but substantial differences were found in the drainage and runoff in both models. Whereas nearly no water was lost through runoff in the ISBA simulations, a substantial amount of runoff was simulated with ORCHIDEE. On the other hand, the drainage in ISBA was consistently larger than in ORCHIDEE. DiagMod does not compute a water balance, so it could not be included in this comparison.

Both models agreed that the largest fraction of LE is through transpiration of the vegetation (Fig. ). Aside from a few exceptions, T / ET in ORCHIDEE was larger than in ISBA. The median T / ET in ISBA (0.53) is lower than in ORCHIDEE (0.68) and is closer to the values derived from the tower observations with the uWUE method (0.54). However, measurements by indicate that this is an underestimation and suggest 0.62±0.06 as a global estimate.

When the observed average water use efficiency is plotted versus the average LE flux, a pattern emerges in which the sites are grouped per PFT (Fig. ). A similar pattern was found in the ISBA simulations, but not in the ORCHIDEE simulations. The range in WUE across the test sites was much smaller in ORCHIDEE than in the observations.

The difference in water use efficiency can be attributed to differences in the modelled plant physiology or the amount of drought stress experienced by the vegetation. As mentioned above, the root zone soil moisture dropped significantly more frequently below field capacity in ISBA compared to ORCHIDEE.

The sensitivity of the surface fluxes to soil moisture and LAI was quantified with a simple linear regression between their anomalies. The slope of these regressions indicates the strength of the response to the state variables.

It was found that the sensitivity of the fluxes to the soil moisture was strongly dependent on the land cover type, in both the observations and the models (Fig. ). A stronger response was found at the herbaceous sites compared to the forest sites. ISBA and ORCHIDEE have too high a sensitivity to soil moisture, whereas the response in the diagnostic model was closer to that in the observations. In Fig. , the same data are plotted but classified per aridity class. This illustrates the oversensitivity of ISBA and ORCHIDEE to drought stress. Despite their differences in implementation and parametrization, a striking similarity in their sensitivity to drought was found, for both LE and GPP. The observations did not show an increase in sensitivity of GPP to Se at dryer sites.

At the forest sites, the response of GPP to Se anomalies is counterintuitively negative. This might indicate that soil moisture anomalies at forest sites were more dominated by wet anomalies, associated with rainfall events. These events coincide with a reduction in solar radiation, hence resulting in a negative GPP response. At herbaceous sites soils were generally drier, so the positive impact of the reduced drought stress after the rainfall event was more dominant, resulting in a positive response. This behaviour was mimicked well in the models.

The sensitivity of LE and GPP to LAI was generally higher at the herbaceous sites. Here, the models tended to underestimate the sensitivity to LAI. At the forest sites, the sensitivity was lower according to the observations. The modelled sensitivity of LE to LAI was reasonably accurate, whereas the sensitivity of GPP to LAI was too strong.

To evaluate the impact of the quality of Se and LAI on the simulated surface fluxes, the Spearman correlation of the errors in the state variables and the fluxes was calculated (Fig. ). It was found in both ISBA and ORCHIDEE that LAI had a stronger error correlation to LE and GPP compared to Se. Grouped per dominant land cover type (Fig. ), both models agree that the error correlation between LAI and GPP was higher at the herbaceous sites compared to the forest sites. Notably, this was not the case for LAI–LE.

Furthermore, the errors in LE were most strongly correlated to those in Se for all models. The highest error correlation was found in DiagMod, where this was most pronounced for the herbaceous sites. At these sites, the Se–GPP error correlation was also the strongest for DiagMod, whereas no strong Se–GPP error correlation was found in the other models.

The simulated LE and GPP from the DiagMod runs with soil moisture and/or LAI from the prognostic models were validated with tower observations. The resulting NS is shown in Table  and Fig. . Similar tendencies were found in RMSE, Pearson r and validation of the seasonal anomalies (not shown here). The DiagMod run with CGLS LAI and ERA5 soil moisture serves as a reference to evaluate the prognostic state variables.

Soil moisture had a stronger impact on LE compared to LAI. A significant (Wilcoxon p<0.05) reduction in NS was found when using soil moisture from ORCHIDEE. This effect is most pronounced at the herbaceous (more water-limited) sites. This is in contrast with the runs using soil moisture from ISBA, which seemed to improve the simulated LE at herbaceous sites (though not significantly). Similar but smaller effects were found at the forest (less water-limited) sites. On the other hand, the opposite was found for the crop sites (n=4) where simulations with soil moisture from ISBA reduced the NS of the simulated LE significantly.

Despite strong differences in LAI, no significant impact was found on LE (with the exception of crop sites with ISBA LAI). A stronger sensitivity to LAI was found in the DiagMod simulations of GPP. A significant reduction in NS was found in all DiagMod runs, but most explicitly in the runs using the prognostic LAI. As in the simulations of LE, this was most pronounced for the herbaceous sites. The use of LAI from ISBA and ORCHIDEE strongly degraded the simulated GPP at these sites, whereas it was unaffected by injecting the prognostic soil moisture.

Overall these results are in line with the error correlations in Fig. . The higher error correlation of LAI to LE and GPP compared to the error correlations of soil moisture was confirmed. Additionally, the stronger impact of prognostic LAI on errors in GPP and of soil moisture on LE was found in both analyses.

With the emergence of freely available data from eddy covariance networks, the use of local datasets is an increasingly standardized approach to evaluate the performance of land surface models . However, the eddy covariance observations notoriously suffer from substantial biases and non-closure of the energy balance . The non-closure of the energy balance is attributed to (1) large advective fluxes caused by surface heterogeneities; (2) systematic measurement errors due to mismatch in observation footprint or inadequate sample rate; or (3) thermal processes, such as heat storage or vegetation metabolism . The test sites in this study were selected to have a relatively homogeneous land cover. Regardless, the resulting uncertainty in the observations was of the same order of magnitude as the model errors. The turbulent fluxes are typically underestimated, as is the GPP . Note that GPP is not corrected for this possible bias in the ONEFLUX processing pipeline . Furthermore, some studies have indicated that the eddy covariance observations are closer to lysimeter data if the energy balance is closed by correcting LE only . Considering this, the negative bias of the simulated LE (and GPP) in this study could be even underestimated. Conversely, others suggest that most or all of the deficit might be related to H or found a good match with independent reference data without LE correction . Validation results of the turbulent fluxes without energy balance closure correction are given in the Supplement.

Generally, the differences between the accuracy of the simulated surface fluxes was most distinct at the sites dominated by herbaceous vegetation (excluding crop sites). These sites have the most pronounced inter-annual variability, and seasonal anomalies are strongly driven by precipitation events . This can be largely attributed to their natural occurrence in dryer climates and shallower root system compared to forests. The seasonal cycle of LAI at the herbaceous sites and its variability were simulated poorly with the prognostic models. The error correlation analysis indicated that these errors were strongly related to errors in the surface fluxes.

At the crop sites, management practices were missing in the prognostic models. In the mean annual cycle of LAI (Fig. ), it is evident that no harvest occurs. Despite this, the simulations of LE were not significantly less accurate compared to other land cover types. After harvest, LE consists largely of bare soil evaporation. Though vegetation was still present in the models, the bulk LE was still reasonably accurate. More evident degradation of the results was found in GPP after harvest, which was overestimated. Even in the diagnostic model, where management practices were incorporated implicitly in the forcing variables, GPP was overestimated. Notably, despite the missing management practices in the prognostic models, the quality of the simulated LE and GPP (and their anomalies) was not significantly different from that at natural herbaceous sites.

Still, the diagnostic model performed consistently well for all types of land cover, contrary to the prognostic models. Only the seasonal variability in GPP at forest sites was simulated less accurately than with the prognostic models. Whereas the remote-sensing-based observations adequately captured this variability for the herbaceous and crop sites, they seemed to fall short for the forest sites.

The errors in the surface fluxes were strongly correlated to errors in LAI for both prognostic models, even though their sensitivity to LAI reflects the observed sensitivity reasonably well (compared to the sensitivity to soil moisture). This seemed to indicate that the source of the errors in the fluxes lies in the feedback mechanism between GPP and LAI (i.e. biomass allocation and phenology), rather than in the forward link between GPP and LAI (i.e. photosynthesis and leaf to canopy upscaling).

The prognostic simulation of LAI in ISBA was introduced by and uses a fairly simple scheme. The latest update was the revision of plant trait parameters according to the TRY database . It has frequently been reported that the seasonal cycle of the simulated LAI in ISBA is delayed by a month or more . attributes this to the leaf longevity parameter, and mention the vegetation undergrowth dynamics as a possible cause for the mismatch between remote-sensing-based LAI and the prognostic LAI in LSMs. However, the issue seems to be related to the architecture of the biomass allocation scheme as well. The assimilated carbon is attributed to the leaf biomass pool first, from where it trickles down to the other pools. No carbon reserve dynamics are implemented. The consequence is that the simulated LAI in ISBA starts slow during spring, as GPP is underestimated due to a low LAI. It continues to build up LAI until late in the second half of the season, when photosynthetic conditions become suboptimal, and leaf senescence is triggered. In contrast, the observed seasonal LAI cycles reach a maximum in the first half of the growing season.

The functional evaluation with the diagnostic model demonstrated that a fairly simple model is capable of simulating the surface fluxes accurately, given accurate observations of LAI. The prognostic LAI generally degraded the results compared to simulations with remotely sensed LAI. Data assimilation experiments have demonstrated the potential of remotely sensed LAI to improve the surface fluxes . Improvements to prognostic LAI schemes are required to increase the skill of the LSMs to simulate surface fluxes.

In that context, processes from ORCHIDEE and other LSMs could be adopted to improve the fairly simple biomass allocation scheme in ISBA. The importance of non-structural carbohydrates to capture the leaf phenology in LSMs is well known, though rarely implemented . indicates that a full-grown canopy of a deciduous broadleaf forest contains approximately 30 % of the total yearly assimilated carbon, yet it is grown in 1 month (1/5–1/6 of the growing season). This rough simplification illustrates that reserve dynamics are essential to simulate the seasonal cycle of the vegetation accurately. Such dynamics are implemented in ORCHIDEE: once certain phenological criteria are fulfilled, the carbon in a reserve pool is allocated to leaf biomass to kick-start the phenological cycle. Still, despite the dedicated phenology modules, non-structural carbohydrate reserve dynamics and a more advanced leaf demography, simulating LAI remained challenging in ORCHIDEE. The timing of the phenological cycle was more accurate in ORCHIDEE, though the accuracy of the simulated LAI was significantly poorer than ISBA. This was the case in particular for herbaceous vegetation. This tendency towards delayed phenology (and in particular a delayed leaf senescence) is found in most earth system models in CMIP5 and CMIP6 .

The discrepancy in complexity between the modelling of photosynthesis and that of the biomass allocation has been highlighted by several authors , though the main challenge lies in the parametrization of those processes. The allocation of carbon in terrestrial vegetation is an important knowledge gap, hindering the advancement of earth system models.

Finally, there are several important differences between the remote-sensing-based vegetation and the idealized vegetation in the models which need to be recognized when comparing both. Firstly, the role of the understorey has a well-known impact on the remote-sensing-based LAI , whereas the LSMs do not consider the separate evolution of an understorey. This can result in substantial differences in the seasonal cycle of LAI. This was illustrated by the differences in the simulations and observations of the LAI cycle at ENF sites. Continuous in situ LAI observations with hemispherical photography at ENF sites are rare, but reported that the effective canopy LAI (including non-green foliage) at FI-Hyy (boreal ENF site) remained constant from June till mid-September. This is in agreement with the flat LAI cycle for ENF in ORCHIDEE but is in contrast with the remote-sensing-based LAI and the prognostic LAI in ISBA. In an empirical model based on in situ observations for the FR-LBr site, LAI demonstrated a seasonal cycle. The understorey was responsible for most of the seasonal variation, and 30 % of the LAI was attributed to the understorey during the summer . The seasonal cycle in the remote-sensing-based LAI seems exaggerated (ranging between 1 m2 m-2 in winter and 4 m2 m-2 in summer). However, considering the understorey and seasonal variation in needleleaf greenness , assuming a flat LAI does not seem accurate either, in the context of simulating GPP.

This brings up a second issue: the remote-sensing-based LAI is the “green” LAI, i.e. photosynthetically active leaves , whereas LAI in LSMs is a key variable which wears many hats. A single LAI variable is used to represent the role of leaves in several processes (photosynthesis, interception, canopy radiative transfer, surface roughness, etc.), in which the greenness of the canopy is not always important. These discrepancies contribute to the mismatch between LAI in the observations and the models. Addressing them might further advance the representation of vegetation in LSMs.

A significant difference between ISBA and ORCHIDEE is found in the simulated soil moisture dynamics, the water partitioning and the water use efficiency. The simulated WUE in ISBA was in fair agreement with what is deduced from the eddy covariance observations. In contrast, the WUE in ORCHIDEE had a much narrower range. The comparison of the LE partitioning shows also that a larger fraction of the water was transpired in ORCHIDEE compared to ISBA. The differences in WUE and flux partitioning could be attributed to differences in the simulated plant physiology or to the quality of the simulated soil water content. The variability in the simulated water content in ORCHIDEE was strongly underestimated, and the vegetation experienced significantly less drought stress in ORCHIDEE. It is likely that this translated to a low variability in WUE as well. Furthermore, a substantial part of the precipitation was lost as surface runoff compared to ISBA. Though we did not have validation data to evaluate the water partitioning, it seems that the simulations of ORCHIDEE could be improved significantly by addressing the soil moisture dynamics. The superior simulation of soil moisture in ISBA contributes to the good performance in simulating the surface fluxes, in particular for sites with herbaceous vegetation and water-driven climate. The functional evaluation demonstrated that the prognostic soil moisture from ISBA even resulted in an improvement in the simulated LE for these sites compared to simulations with ERA5 soil moisture.

Overall, the accurate simulation of soil moisture and water infiltration is a challenge (perhaps one of the main challenges) in land surface models . The poor quality of the simulated soil moisture compared to in situ observations is also evident in this study, despite the use of the multi-layer diffuse water transport scheme. The soil physical parameters are determined using a global pedotransfer function (PTF), using only texture as input. New, advanced PTFs with global coverage have emerged in recent years, using not only texture, but also climatology and land use as predictors . As soil moisture is the basis of many processes in LSMs, incorporating these PTFs seems to be the logical new step forward in LSMs .

The local-scale simulations in this study were not coupled to a hydrological model; thus groundwater dynamics were lacking. Though only a limited effect of capillary rise was found in studies with a coupled groundwater hydrology, the impact can be non-negligible for forest ecosystems with a deep root system . The further development of groundwater dynamics in LSMs is indispensable for the accurate coupling of energy, water and carbon in forest vegetation and its response to severe drought events.

Several efforts have already explored the potential of improving soil moisture dynamics in LSMs. Substantial improvements to soil moisture have indeed been obtained by calibrating the pedotransfer functions or soil physical parameters. Yet, the impact thereof on the surface fluxes has been found to be relatively limited , or in some cases even negative . Though many parameters in ISBA and ORCHIDEE are derived from databases , the LSMs have been calibrated to produce accurate surface fluxes using (amongst others) eddy covariance observations. The limited accuracy of the soil moisture dynamics might have been overcompensated in the resulting parametrization . The oversensitivity to drought stress in ISBA and ORCHIDEE is possibly an illustration of this. Improvements to the intricate network of gears under the hood of LSMs are a delicate matter. Addressing the soil moisture dynamics should go hand in hand with corrections to the oversensitivity to drought stress.