First of all, I want to thank you for reviewing the work of me and my co-authors. You have spent a lot of effort and emphasized many details. Your criticism is important for improving this work and publication and your hints are useful for reaching this goal. In the following, I will go through and respond to your comments.

Specific comments:

1. Indeed, model development needs testing and novel model modules are only incorporated when there is an improvement or at least no deterioration in model performance in the variables that were chosen for evaluation. The choice of variables for model evaluation is really important. Since ECLand mostly is used in Climate Projections, the most important variable is land surface temperature which then is used for model evaluation. However, changes in vegetation representation barely affect energy balance calculations, especially not on a coarser temporal resolution that often is used for model evaluation. As a result, it is hardly surprising that model performance in our investigation diverges from published results during model testing. Thus, yes, we wanted to indicate that model testing needs a broader spectrum of target variables and different temporal resolutions. The work of Nogueira et al. (2020) is interesting but they focused more on updating land cover fractions and vegetation type clumping which had an important effect on land surface temperature. However, it still is worth including it to the discussion since it shows the importance of vegetation-related variables in ECLand.
2. The LAI from MODIS used for model input and model evaluation is not identical. Model input is a LAI climatology on monthly basis resulting from multi-year average MODIS values. Model evaluation is done with the daily MODIS values which are 8-day means. For the static runs, this comparison provides the information whether an incorporation of more site-specific climatology results in higher representativeness of local LAI evolution. For the dynamic simulations, comparing modeled LAI with daily MODIS values is used to examine whether the models are able to capture inter- and intra-annual LAI dynamics. However, we could show that even with the same source of the data the dynamic simulations are not fitting the observations. Since this concern arose, I need to provide more details on the MODIS LAI data and highlight the differences between data used for input and for evaluation and the reasons behind this differentiation. The evaluation would really benefit from using on-site LAI data from more than one site. We were very thankful for having an additional LAI data source at all. I tried reaching out to the FLUXNET community via their contact form several times but never got any responses.

1. We chose ECLand and Noah-MP because both models can be and are widely used for coupling them as LSMs with established climate projection models. Although Noah-MP provides no GPP and NEE output for the static runs, it still is interesting to look at the LAI-GPP relationship within the model that we did for Figure 8. Nonetheless, we need to be more careful with absolute statements that we did and will adjust the abstract and the discussion.
2. When looking at the global distribution of FLUXNET sites, a bunch of them is located in temperate climate conditions on the Northern Hemisphere. Including all sites with more than 5 years would create an overrepresentation of regions with high density in sites, resulting in an imbalance of PFT-aridity combinations for model evaluation with especially (semi-)arid short vegetation being underrepresented. Thus, we needed some sort of filter algorithm to avoid that overall model performance is either shifted towards better or worse performance due to this imbalance. Savannah types are additionally separated within IGBP PFT which is not done in the models. Accordingly, I did not when selecting the sites, meaning that SAV and WSA belongs to the same group within this selection process. I also merged PFT type MF with DBF since, after the selection via the aridity index, only two MF remained which is critically few. Other possible sites during the selection were thrown out mostly because of lacking or low-quality soil moisture data. Additionally, there is no possibility to create a second dataset with the same structure as ours because some aridity-PFT combinations are really rare. However, such an identical site selection would be helpful for strengthening and reproducing our findings. What we could do now, is to explain in more detail why and how site selection was done, and adapt Figure 1 in accordance with the model PFTs.

1. We will take care for mistakes in citations and linguistic deficits. Separating Results and Discussion part seems reasonable to me by considering the arguments you gave.

Technical Corrections

Thank you for these huge number of propositions. I will only respond to those that exceed language.

• 4: I will make it more general.
• 5: “Current implementation” meaning the model source code as is it published currently.
• 11: ok, I can change this.
• 12: I was not aware of the term “ensemble” being used in a specific manner. Can change this.
• 13: Benchmarking studies use relative metrics to create a rank order of the models but do not provide information on whether the best model in this ranking really achieves good fit with observations since the absolute metrics are not shown.
• 17: good to know
• 20: I can include 1-2 sentences on misrepresentation of water-sensitive regions.
• 21: Yes, vegetation parameters might impact transpiration within LSMs but processes are largely simplified or do not represent reality. As a result, surface exchange in LSMs can be insensitive to processes that involve vegetation.
• 26: This refers to the discrepancies in process representation related to vegetation.
• 28: While broader studies on model comparisons do not go into details of the model structure or processes, we aimed to investigate what are reasons for deteriorated model performance, which implies looking through the source code and conduct sensitivity tests. This deep analysis is time-consuming and, thus, limits the number of models that can be examined in that way.
• 31: I can replace the word “different” with “temporal” to avoid confusion in what the patterns are supposed to be different. However, “possible misrepresentation of the observations” states what it is supposed to say. Writing “possible misrepresentation of observed patterns” could make it more clearly.
• 32: ok, I can remove it.
• 34: ok
• 35: I will reformulate it. But since the site selection will be described in more details, that sentence might chance anyways.
• 36: I will add the reference. The longest time series within the selected sites was 1996-2014 which is 18 years. I refrained from using a continuous color scale because the observational time can only be full years and, thus, the sites have distinct duration classes. I merged two duration classes together since we do not have that many sites and some duration classes might be empty otherwise. Additionally, the differences between the colors can be seen better on this “coarser” scale.
• 38: We wanted to use as less abbreviations as possible to assure readability especially in Results and Discussions section. Additionally, the part where soil water content is referred to is limited which made it unnecessary to use an abbreviation.
• 39: ok
• 40: will do
• 41: I found this in my scripts when preparing for the GPP-LAI correlations. There, I deleted time slots where fillings (equal value until the 8^th digit) was longer than 90 days. For US-Var, I left out the year 2001 because GPP data was nonsense due to gapfilling. I need to explain more carefully which data were left out for model evaluation and why.
• 42: I can add a sentence on the quality flags of MODIS.
• 44: ok
• 45: I will rephrase that.
• 46: Yes, this refers to consecutive months. I will add that information.
• 48: Descriptive simulation settings are helpful and prevent confusion, so I will keep these.
• 49: Good point. I needed to keep data with QC 48 and 65 for creating the climatology to ensure that each month got a value for that site, which was a challenge especially for the single-year simulations. But since the trustability of these data points is low, they were left out for the temporal higher resolved evaluation.
• 53: ok
• 54: We initialized the model with the closest possible fit to the on-site conditions without changing any parameters. For ECLand, we had a global setup that we used. We didn’t adapt the parameters in the global setup. Additional tests were not conducted. I will incorporate the classes for ECLand into Table 2 to enhance clarity. However, you are correct with your concern that tile fractioning in ECLand into high and low vegetation in the default setup might bias the evaluation with point measurements that belong to only one of these vegetation types. I will conduct an additional test on that.
• 57: Yes, “respective cover” means the fraction of each vegetation type on the grid cell.
• 59: Here as well, we assured model setup to fit as closely as possible the on-site conditions.
• 62: Yes, as it is explained in Niu et al. (2011) section 4.2.
• 63: To be honest, the information that the tower ends in the vegetation canopy, is new to me especially since the aim of the network is to capture fluxes of the respective vegetation type. I checked the given measurement heights of the sites I chose, and two of them might look suspicious but I don’t know the vegetation on-site.
• 64: The uppermost soil layer for Noah-MP is 0.1 m and for ECLand 0.07m. I can add this information.
• 65: Steady state was not checked quantitatively but qualitatively.
• 66: The initial files contain information on soil, tile fractioning, LAI climatology, state variables at the time of the start of the simulation. For the latter, I could have replaced them by measured values from Fluxnet2015 but the values change during the spin-up anyways.
• 67: Because the position of the flux tower might be at the edge of a grid cell of soil grids, we decided to include the neighboring grid cells as well.
• 68: ok
• 69: This is about the minimum green vegetation fraction. I miscommunicated it in the manuscript. Setting the minimum green vegetation fraction to 1% assures that there is still a small amount of biomass after the winter, which is essential for the model to generate spring growth. Without any biomass (i.e. leaves) there would be no location for photosynthesis to take place (zero leaf area * high potential photosynthesis still is zero).
• 71: Yes, vegetation canopy surface temperature is meant.
• 72: Indeed, I can remove the shrublands from the table.
• 75: ok I can remove Eq. 1
• 78: You are right, but in case of very low variance, the numerator is also small, which results in reasonable values of the relative bias and not like 3000%.
• 82: ok, I can introduce a symbol for the normalized standard deviation.
• 83: ok
• 88: In Boussetta et al. (2013), they analyzed 32 Fluxnet sites. I will add this information.
• 89: I can reformulate the sentence.
• 90: I can reformulate this statement.
• 91: Good point. This might be even introduced in the methods section.
• 93: I can change the thickness of the arrows to be a bit higher, in hope that helps.
• 94: This statement refers to using MODIS climatology as LAI forcing. I will add this information again.
• 97: I can add sublabels.
• 99: Why should the resulting LAI with only one vegetation type be lower than with both? One grid cell in ECLand is split into high and low vegetation fraction with their LAI values. Meaning, one spot cannot have both vegetation types (there is no layering). The resulting LAI is then the weighted mean according the high and low vegetation fraction. Thus, if a grid cell in our setup is only a high vegetation type, resulting LAI is higher than for a grid cell that has also a low vegetation type fraction. This is the closest we can get to the footprint of flux tower observations.
• 101: Correlation coefficient for static Noah-MP LAI for US-Var was negative, so the arrow starts there. Regarding diverging relative bias and correlation coefficient for US-GLE, I will check the output of LAI.
• 102: Quantitative limits of the aridity classes based on Ashaolu & Ilorin (2018).
• 103: Yes, I can add colors.
• 104: The term model performance aims to include all the metrics that are discussed here. “Lower model performance” in general means that the majority of the metrics show deterioration. In other cases, the explicit metric is referred to. However, I can include an additional explanation in the methods section.
• 105: I can add further explanation on that.
• 107: ok
• 110: More insights and evidence for this are given in section 3.3.
• 112: Using the same metrics from other studies is not possible since they basically do not have them.
• 113: ok, I can delete this.
• 114: Since Results and Discussions will be separated in the revision (as stated before), the “new” results part will also include more quantities.
• 115: ok. I can do an additional t-test.
• 116: ok, I can extend that discussion on that.
• 117: ok
• 118: In the EF calculation, LE is in the numerator. Thus, lowering LE reduces EF. When LAI is modelled to be small, the transpiration can only be low (water balance) or, equally, LE is smaller (energy balance) because less energy is used to transpire water. With an underestimation of LAI also the EF representation deteriorates.
• 122: Yes, true, there were some exceptions. I can mention and discuss them.
• 123: Disagreement between statements from Ma et al. (2017) and our study is low. Only the bias values vary and I can look into that and maybe provide causes for that.
• 125: “Optimal values” refers to the values for soil characteristics in look-up tables. I can rephrase this to make it more clearly.
• 126: ok, I can reformulate it.
• 128: I agree with you that the phrasing is ambiguous and needs more explanation. But the finding, that model performance in LAI and in LE seems to be independent of each other although LE values depend on LAI values, does not affect the confidence in the results since it is one of the results.
• 132: Good point. I forgot to make the x axes consistent. For other figures, I could not relate that criticism. However, I can extent the figure caption.
• 133: LAI in Figure 8e-p is the model output.
• 135: Good suggestion!
• 136: I guess, this refers to the tropical site (GF-Guy) selected for the Figure 8. We were also concerned about this large range of LAI values. However, we handled data quality as careful as possible and used only days with high standard quality flags. I will additionally check the LAI seasonality of that site.
• 139: True, will rephrase.
• 145: Yes, the 11% of assimilation in the model goes into dark respiration. I did not check this ratio for the observations but definitely is a good idea.
• 148: I can rephrase this and be more precisely.
• 155: I will consider this suggestion after reformulating the discussion section.
• 156: ok
• 157: True, the headings need to be adapted.
• 158: There are some exceptions but for the majority of the sites, soil moisture did not respond to changes in LAI. We just mentioned the majority since the overall manuscript is long and provides much information anyways and we tried not to overload the Results section with small details.

Literature

Ashaolu, Eniola & Iroye, Kayode. (2018). Rainfall and potential evapotranspiration patterns and their effects on climatic water balance in the Western Lithoral Hydrological Zone of Nigeria. Ruhuna Journal of Science. 9. 92-116. 10.4038/rjs.v9i2.45.

Souhail Boussetta , Gianpaolo Balsamo , Anton Beljaars , Tomas Kral & Lionel Jarlan (2013) Impact of a satellite-derived leaf area index monthly climatology in a global numerical weather prediction model, International Journal of Remote Sensing, 34:9-10, 3520-3542, DOI: 10.1080/01431161.2012.716543

Ma, N., Niu, G.-Y., Xia, Y., Cai, X., Zhang, Y., Ma, Y., & Fang, Y. (2017). A systematic evaluation of Noah-MP in simulating land-atmosphere energy, water, and carbon exchanges over the continental United States. Journal of Geophysical Research: Atmospheres, 122, 12,245–12,268. https://doi.org/10.1002/ 2017JD027597

Niu, G.-Y., et al. (2011), The community Noah land surface model with multiparameterization options (Noah‐MP): 1. Model description and evaluation with local‐scale measurements, J. Geophys. Res., 116, D12109, doi:10.1029/2010JD015139.

Nogueira, M., Albergel, C., Boussetta, S., Johannsen, F., Trigo, I. F., Ermida, S. L., Martins, J. P. A., and Dutra, E.: Role of vegetation in representing land surface temperature in the CHTESSEL (CY45R1) and SURFEX-ISBA (v8.1) land surface models: a case study over Iberia, Geosci. Model Dev., 13, 3975–3993, https://doi.org/10.5194/gmd-13-3975-2020, 2020.

First of all, I want to thank you for reviewing the work of me and my co-authors. You have spent a lot of effort and emphasized many details. Your criticism is important for improving this work and publication and your hints are useful for reaching this goal. In the following, I will go through and respond to your comments.

General comments:

• Our main focus was testing whether switching on dynamic vegetation in the models enhance their performance regarding the target variables. We changed the LAI source in order to find out whether this more site-related information as initial input “helps” the model in their prediction of LAI and NEE. However, we did not aim for doing data assimilation since there are many investigations published on that. Prescribing the LAI is always for initializing the models independent of dynamic vegetation. In the model simulations themselves, this prescribed LAI is only used in the model runs with static vegetation. I can be more cautious with the terminology and the descriptions within the manuscript.
• The default climatology in the initial file (what I refer as LUT LAI) of ECLand is already based on MODIS values. A time span from 2000 to 2008 and disaggregation of the gridded values for LAI was used to create that climatology (Boussetta et al., 2013). LAI values in the look-up tables of Noah-MP are defined for the plant functional types (PFTs). These values are also based on MODIS observations which were disaggregated to the different PFTs on each observational grid cell (Oleson et al., 2010). I could not find any information from which time span these values were taken or how individual LAI climatology within one PFT were merged. In the default setup, this LUT LAI was used. For the other setups, those values in the LUT were replaced by “our” LAI values from MODIS.
• The LAI from MODIS used for model input and model evaluation is not identical. Model input is a LAI climatology on monthly basis resulting from multi-year average MODIS values. Model evaluation is done with the daily MODIS values which are 8-day means. For the static runs, this comparison provides the information whether an incorporation of more site-specific climatology results in higher representativeness of local LAI evolution. For the dynamic simulations, comparing modeled LAI with daily MODIS values is used to examine whether the models are able to capture inter- and intra-annual LAI dynamics. However, we could show that even with the same source of the data the dynamic simulations are not fitting the observations. Since this concern arose, I need to provide more details on the MODIS LAI data and highlight the differences between data used for input and for evaluation and the reasons behind this differentiation.
• I could spend a bit more space important processes in the model that are related to LAI but I would refrain from explaining the whole model.
• Thank you for the advice. I will do that.
• Splitting Results and Discussion section is planned.

• We chose ECLand and Noah-MP because both models can be and are widely used for coupling them as LSMs with established climate projection models. Although Noah-MP provides no GPP and NEE output for the static runs, it still is interesting to look at the LAI-GPP relationship within the model that we did for Figure 8. Nonetheless, we need to be more careful with absolute statements that we did and will adjust the abstract and the discussion.
• In principle, both models are initialized with the same values, fitting as close as possible to the on-site conditions. However, there are some technical differences in the model initialization that I described. I need to be more precise which differences result from model structure in order to avoid confusion on the model setup.
• For sure for Noah-MP, since there is only one vegetation type on the grid cell. For ECLand I would have needed to adapt vegetation to be either high or low vegetation in the initial file, which I didn’t because we used a global initial setup. I will conduct an additional test with ECLand whether setting the vegetation type to either low or high vegetation in the default static and dynamic runs change the model performance. Regarding the model performance of short vegetation types, I could interpret a bit more. One possible reason could be that forests have less dynamics in their productivity compared to crops, grasslands or shrubs. Surely, trees have dynamics in their leaf mass and photosynthesis rate dependent on environmental impacts but, in general, have access to deeper water resources and intrinsic carbon storages to at least partly overcome water scarcity. Shorter vegetation types cannot cope for limitations in this way, resulting in higher relative temporal variations.

Specific comments:

• L8: “More detailed information” refers to the on-site LAI. Either I could mention explicitly that or leave it out.
• L13-14: We didn’t aim to pinpoint poor model performance of the models themselves for single or all selected sites. The question of this investigation was whether model performance can be improved by dynamic vegetation. Since this is not the case, we provide possible explanations and misrepresentation of the relationship between LAI and GPP is the major one we figured here. However, I can reformulate the sentence to make this more clearly.
• L21: Good point, I can include that.
• L24-25: I could add “… features that aim for picturing reality more closely…”
• L26-36: No, we don’t want to come up with new evaluation schemes. Rather, we want to motivate why we did an analysis with only a few models and presenting absolute performance metrics, which seems like “a step back” in comparison with multi-model evaluations.
• L29-31: Multi-model evaluations like PLUMBER compare the performance of many models against each other and against alternative prediction schemes which might be simple regression.
• L34-35: I am not sure whether I understand your first question. Best et al. (2015) had statistical models within their multi-model ensemble. In the end, they presented a ranking of all models included. While doing so, surely, is important and provides information on which model is “more capable” in representing observations, normalized metrics do not show how close or “good” this representation actually is. The model that is best in a model ranking (and, thus, having high relative normalized metrics) could still have very poor individual model performance if all models of the ensemble fail in reproducing the observations.
• L40: Of course, there will be always uncertainty in measured data but I am sure that they accounted for that. Haughton et al. (2016) were investigating reasons for the outcomes of the PLUMBER study that simple empirical models outperformed most LSMs. They excluded systematic bias of flux tower data, time scaling effects and lack of energy conservation in the data as potential causes and stated that processes within or parameterization of the LSMs themselves need to cause poor performance.
• L41: What we were trying to say with that sentence was that benchmarking or ranking models alone is no suitable tool to identify specific causes for a mismatch between model predictions and observations. Achieving this, needs a deeper look into single models and their individual performance.
• L45: ok
• L46ff: Since one of the motivations to have dynamic vegetation in LSMs is to better predict impacts of water scarcity and drought events on the vegetation, we found it would be valid to argue that current implemented and used LSMs struggle in making prediction that fit observations in these conditions. However, I can think about shortening this paragraph.
• L66-67: We chose ECLand and Noah-MP because both models can be and are widely used for coupling them as LSMs with established climate projection models. Additionally, both models undergo constant developments and the modules with dynamic vegetation were introduced not long ago.
• L80: Aridity describes water deficit in long-term climate conditions. Following this, it is the ratio of annual potential evapotranspiration to annual precipitation, leading to larger values of this ratio meaning larger aridity of the site. However, the ratio in this dataset was calculated the other way around which is less intuitive. Also, since we planned to filter the sites on a logarithmic scale, inverting delivered the opportunity to include more semi-arid and arid sites which differ much between each other with respect to seasonality and vegetation dynamics while humid sites are more even.
• L83: It is not a common threshold but we needed to come up with one within our filter algorithm. The aridity indices of wetter sites are closer to each other than for drier sites. In order to not overrepresent dry sites within selection by using a threshold in absolute values of the aridity index, we transformed the aridity index to a logarithmic scale, creating almost linearity of the aridity index scale.
• L87: Haughton et al. (2018a) found out that, within the FLUXNET sites, drier sites (higher aridity index) and wetter sites with low temperature span tend to have higher predictability, meaning that it is easier to achieve good model performance. With our selection by aridity, we assured that we do not only include sites with high or low predictability.
• L97: Filling missing precipitation data with zeros is the only option that is possible. We don’t know whether it rained that hour or day. However, the model input cannot handle missing values.
• L97: I do not know how common the Kalman filter is. Gapfilling for the TERENO site “Hohes Holz” was done with it. FLUXNET usually uses Marginal Distribution Sampling which is a really complicated algorithm to implement and to run. Additionally, it cannot fill large gaps as well, which can be seen in time series data from some of the FLUXNET sites.
• L97-98: The ERA5 product I retrieved had 0.1° spatial and 1h temporal resolution and, thus, really helped with filling the gaps. The limit of 3h in using the Kalman filter evolved from the observation that the filter tends to overestimate the values when gaps are longer.
• L100-101: Longer periods where data is filled with Marginal Distribution Sampling within the FLUXNET dataset can be seen visually because variability is unnaturally low. “Longer” in this respect means at least a month.
• L104: Temporal resolution is 8 days. There are different MODIS datasets available. The one I used, MOD15A2H, has a spatial resolution of 500 m. I can add that information.
• L105: ok
• L106-107: Creating the LAI climatology means to calculate the average annual LAI cycle. For a 10-year time series of MODIS LAI, it might happen that some months have 30 values while other months have only 3 by selecting the same quality flags (i.e. 0 and 32). For example, a tropical site is covered by ITC cloudiness nearly at the same time of each year. Thus, all the values during that time have a lower quality flag and would be excluded. It happened that we were left with some months without any LAI information, so we included a larger set of flagged data points for the climatology.
• L112: See my explanation in point 2 of the general comments.
• Table 1: Yes, I can rephrase setup description and add information on the time span.
• L127: “Under-development” means that these models (and here especially the modules that incorporate dynamic vegetation to the models) are constantly extended and improved.
• L129: I do not understand your question. The prescribed LAI in ECLand has monthly resolution which we mimicked also with our MODIS climatology. Within model calculations of LAI are done on daily basis.
• L133-134: Yes, at maximum. It could be even one or none.
• L134-142 + L150-156: I tried to leave model description as short as possible. However, more details on LAI-related processes might help and I will include them.
• L164: Both models have two types of input: Initial files (with initial values for some variables to start with) for model setup and time series files with meteorological data for model runs. The initial files contain variables like vegetation type, deep soil temperature, soil layering, soil type, initial soil moisture, vegetation cover fraction and initial LAI value or LAI climatology which are not all present in the FLUXNET data. For ECLand, these initial data files were prepared for a global setup already and we could make use of that. For Noah-MP, no such setup existed and we created the initial files by ourselves by using the information we had. After model initialization followed the spin-up phase so that these initial values were not used any longer and became overwritten by actually modelled values.
• L166: Clustering the vegetation into high or low vegetation type does not depend on vegetation height but on the vegetation type on-site. Forests in any case are high vegetation no matter how big the trees actually are.
• L169: For sure.
• L172: The reason for the initial conditions of the two models being different is only because these initial files look different for both models and require slightly different set of variables. Apart from that, we kept initial conditions as close to each other and as close to on-site conditions as possible.
• L173: Good question, I will check that.
• Table 2: Yes, I will extend the table.
• L183: Yes, daily averages or sums (depend on variable).
• L188: In principle, the relationship between observed and modelled values of a target variable is expected to be linear.
• L189-190: I can introduce a symbol for the normalized standard deviation, likely will be s.
• L191-193: A “normal” relative bias was not applicable since our target variables (i.e. LE, H, NEE and GPP) have values that vary around zero. This results in relative biases that are not only really large partly but also difficult to interpret (e.g. reaching 3000% of relative bias but not because the model estimate is far away from observation but rather because the mean value is close to zero). By subtracting the minimum, the distribution is shifted to positive values only, with the minimum value being zero. As a result, the relative bias really contains an information on how much the estimates deviate from the mean since the reference system is the codomain of the variable. This works independently of the distance between x_min and x_mean.
• L199-200: Yes, exactly.
• L199-207: I can try to explain the elasticity in more detail. Unfortunately, I found no publication from environmental sciences that use the same metric, only from economics.
• L213: All symbols that are in the Taylor plots.
• L217-218: I can add information on that.
• L222: Here, I refer only to literature because Stevens et al. (2020) also replaced LUT LAI by MODIS LAI and compared model results.
• L224: For the dynamic simulations, LAI is not prescribed but still part of the initial files. It is expectable that LAI predictions for the dynamic simulations are independent of the initial input. However, dynamic ECLand still incorporates prescribed LAI to 5% (RLAIINT=0.95 was defined by the developers’ team to be fully dynamic).
• L225: ok
• L226: ok
• L226: Increased variance in comparison with static ECLand simulations.
• L227: We could not find any tendencies regarding aridity or vegetation type to have positive or negative shift in relative bias.
• L231: No, I did not. Sparse vegetation are savannas and shrublands because they have no closed canopy surface.
• L243-244: ok
• L255: ok
• Fig 2: I will add the aridity index classes.
• Fig 3: I will change colors. Static Noah-MP produces no output for NEE and GPP which is according to model structure. Thus, only the values for the dynamic runs can be presented here.
• L265: Yes, will be added to model description.
• L281-282: Although Noah-MP provides no GPP and NEE output for the static runs, it still is interesting to look at the LAI-GPP relationship within the model that we did for Figure 8. Apart from that, we still can look at LAI, latent heat flux and soil moisture.
• L284: For most of the sites, GPP was overestimated with dynamic Noah-MP (Tab. A3). Yes, you are absolutely right, I will rethink this statement.
• L289: yes
• L291: On-site LAI and MODIS LAI were linearly correlated. MODIS LAI might be biased for some sites, but so might be on-site measured LAI due to technical limitations (scatter correction, saturation effect…). During development of the dynamic vegetation modules, a tuning of the parameter sets was done but not to MODIS LAI as target variable. However, mismatch between MODIS and on-site LAI is reflected in lower performance of NEE and GPP of the static ECLand simulations. The reason is unclear: It could be that on-site LAI does not reflect actual LAI but it could also be that calculations of GPP in relation to LAI do not match reality (similar to what we have shown in Fig. 8). For the dynamic ECLand runs, differences between MODIS and on-site LAI play only a minimal role since 95% of the LAI calculations come from dynamically predicted LAI and NEE and GPP predictions are even fully dynamically predicted. I will rephrase this paragraph to make it more clearly.
• L306-307: The performance is not different for the dynamic simulations. But for the static runs, it is. Thus, we recommended here to use static simulations with MODIS climatology forcing. However, I just recognized by reading your comment that we can only recommend this for ECLand since for static Noah-MP we don’t know the actual performance regarding NEE and GPP.
• L329-330: It might be that carbon and water transport processes are coupled not tightly enough. With NEE estimates fitting well, the photosynthetic activity also is good captured by the model. The demand of water by the photosynthesis might be underestimated by the model and, leading to less transpiration and, thus, also to a lower fraction of energy that is used for latent heat transport. Additionally, downward CO2 transport and upward water transport through turbulent fluxes occurs in the same eddies which is not captured by the model. These are just some ideas on that so far.
• L335: ok
• L351: Vegetation needs water for photosynthesis which stems from the soil. Thus, more photosynthetically active biomass extracts more water from the soil and, otherwise, less soil water restricts maximum plant productivity and biomass build-up.
• L354-359: Yes, you are right. The reason for unaffected soil moisture to vegetation dynamics still remains unclear. Referring to the point before it could be due to the implemented interaction of carbon and water processes. First, the potential photosynthetic activity in dependence of leaf area and radiative conditions is calculated. Then, the limitation factor of extractable water is estimated according to available soil water and roots. Lastly, the photosynthetic activity is adapted to that restriction and transpiration rate adapted to conductivity and atmospheric conditions. As a result, the only included path is that soil moisture impacts photosynthetic activity and biomass build-up. But there is no feedback that more biomass needs/loses more water that will be taken from the soil because photosynthetic activity relates only to the carbon fluxes but not to the water fluxes.
• Fig. 7: Sorry that footnote was there by accident.
• L389-396: I can add LAI-GPP elasticity for ECLand in Figure 7. Other studies also found a linear relationship between LAI and GPP but with large variability. We could add more than the one cited from Hu et al. (2022). Some sites might be exceptions from the linearity (IT-Ren) where LAI-GPP relationship appears to be a non-linear saturation function.
• Fig. 8: I will add description of the arrows to the figure caption. Since the most probable in the observations LAI-GPP relation is a linear one, Pearson correlation coefficient is the statistical basis of this linear regression and also the measure for the relationships from the model output. I cannot compare different kinds of correlation coefficients.
• L403: ok
• L409-410: ok
• L454-455: I cannot replace “real” by “observed” because I am not referring to any measured observations here. The reality this sentence is referring to is the fact that trees do not immediately lose their leaves when they are faced to a few days of suboptimal conditions for photosynthesis. However, I can think of alternative terminology.
• L461-462: I will rephrase this.
• L467: Compared to forests that are more resistant and resilient for e.g. water scarcity, short vegetation more dynamically and more instantly responds to environmental limitations for its growth. Thus, firstly, assuming the same LAI cycle for each year and, secondly, assuming a constant LAI values over a whole month as in the static model simulations, do not represent reality. Our expectation was that modelling vegetation dynamically would cope for that variability and, as a result, yield in better performance of observed ecosystem fluxes.
• L480-481: Other models have processes implemented differently. So, there is no chance in directly transferring results and conclusions from these two models to others. I can check the new PLUMBER study which I didn’t recognize yet.

Literature

Best, M. J., Abramowitz, G., Johnson, H. R., Pitman, A. J., Balsamo, G., Boone, A., Cuntz, M., Decharme, B., Dirmeyer, P. A., Dong, J., Ek, M., Guo, Z., Haverd, V., van den Hurk, B. J. J., Nearing, G. S., Pak, B., Peters-Lidard, C., Santanello, J. A., Stevens, L., and Vuichard, N.: The Plumbing of Land Surface Models: Benchmarking Model Performance, Journal of Hydrometeorology, 16, 1425–1442, https://doi.org/10.1175/jhm-d-14-0158.1, 2015

Souhail Boussetta , Gianpaolo Balsamo , Anton Beljaars , Tomas Kral & Lionel Jarlan (2013) Impact of a satellite-derived leaf area index monthly climatology in a global numerical weather prediction model, International Journal of Remote Sensing, 34:9-10, 3520-3542, DOI: 10.1080/01431161.2012.716543

Haughton, N., Abramowitz, G., Pitman, A. J., Or, D., Best, M. J., Johnson, H. R., Balsamo, G., Boone, A., Cuntz, M., Decharme, B., Dirmeyer, P. A., Dong, J., Ek, M., Guo, Z., Haverd, V., van den Hurk, B. J. J., Nearing, G. S., Pak, B., Santanello, J. A., J., Stevens, L. E., and Vuichard, N.: The plumbing of land surface models: is poor performance a result of methodology or data quality?, J Hydrometeorol, 17, 1705–1723, https://doi.org/10.1175/JHM-D-15-0171.1, 2016.

Haughton, N., Abramowitz, G., De Kauwe, M. G., and Pitman, A. J.: Does predictability of fluxes vary between FLUXNET sites?, Biogeosciences, 15, 4495–4513, https://doi.org/10.5194/bg-15-4495-2018, 2018a.

Oleson, K. W., Lawrence, D. M., Bonan, G. B., Flanner, M. G., Kluzek, E., Lawrence, P. J., … Zeng, X. (2010). Technical Description of version 4.0 of the Community Land Model (CLM) (No. NCAR/TN-478+STR). University Corporation for Atmospheric Research. doi:10.5065/D6FB50WZ

Stevens, D., Miranda, P. M. A., Orth, R., Boussetta, S., Balsamo, G., and Dutra, E.: Sensitivity of Surface Fluxes in the ECMWF Land Surface Model to the Remotely Sensed Leaf Area Index and Root Distribution: Evaluation with Tower Flux Data, Atmosphere, 11, https://doi.org/10.3390/atmos11121362, 2020.

First of all, I want to thank you for reviewing the work of me and my co-authors. You have spent a lot of effort and emphasized many details. Your criticism is important for improving this work and publication and your hints are useful for reaching this goal. In the following, I will go through and respond to your comments.

• Indeed, the footprint of a flux tower observation has a smaller area than the grid cell of LSMs. Nonetheless, comparison of model output against point-level observations such as those from FLUXNET is a common way to perform model evaluation, especially, since most LSMs are able to be used on a wide range of spatial scales. FLUXNET delivers the basis for such a model evaluation on smaller scales. PFTs are a concept to simplify the parameterization of vegetation that is expected to respond in a similar way to its environment. As a result, they should be transferable and representative for all subtypes of vegetation that are merged into one PFT. If not, they would have been separate groups. We set the vegetation of the considered grid cell within the model to the PFT that fit closest to the on-site conditions to minimize potential mismatches in parameterization.

• Firstly, the dynamic vegetation modules are not unparameterized, only not additionally calibrated for that specific site. Switching on dynamic vegetation introduces, as you said, environmental dependency of LAI to the model. If the model is allowed to adapt the vegetation (and its productivity and LAI) to environmental conditions, it can be expected that model predictions are closer to the observations compared to simulations with static vegetation. The climatology contains long-term seasonality of LAI. It represents the average temporal pattern of LAI that is adapted to the long-term mean environmental conditions. Intra- and interannual variability as a result of environmental conditions cannot be included into the LAI climatology. To cope for this, dynamic vegetation modules were implemented.
• Results and Discussion parts will be separated.
• By that point of the manuscript, this conclusion is curious. The answer is within section 3.3. This independency does not mean that the predicted values of latent heat flux do not change when LAI is changing. Rather, the model performance in latent heat flux does not change. Together, this means that the mismatch between modeled and observed latent heat flux might be small or large (depending on the site) but is in almost the same extent small or large with a different LAI representation, resulting in the same model performance.
• Yes, model description will be extended by all processes related to LAI.
• Good point. I can add this to the appendix.
• When looking at the global distribution of FLUXNET sites, a bunch of them is located in temperate climate conditions on the Northern Hemisphere. Including all sites with more than 5 years would create an overrepresentation of regions with high density in sites, resulting in an imbalance of PFT-aridity combinations for model evaluation with especially (semi-)arid short vegetation being underrepresented. Thus, we needed some sort of filter algorithm to avoid that overall model performance is either shifted towards better or worse performance due to this imbalance. Unfortunately, there is no possibility to create a second dataset with the same structure as ours because some aridity-PFT combinations are really rare. However, such an identical site selection would be helpful for strengthening and reproducing our findings.
• Even when the models run with vegetation dynamics, the LAI climatology is still part of the initial files. For Noah-MP, these climatological values are not used for the dynamic setup. This is why we end up with the same model performance for all dynamic Noah-MP runs. In ECLand, the vegetation is not totally dynamic. For instance, with a fraction of 5% the prescribed LAI still merges into the LAI estimate for that simulation day (defined by Souhail Boussetta, what they use as a dynamic ECLand setup). Thus, also model performance of ECLand differs a bit depending on LAI climatology source. The two columns in the Figure should show in which direction and how much model performance shifts for the dynamic simulations compared to only relying on the prescribed LAI climatology of the simulations with static vegetation.
• Static Noah-MP does not calculate GPP and NEE (missing values in the output file). This relates to model structure. LAI for the next time step is already known, so there is no need to estimate assimilation by photosynthesis or allocation to plant tissues.
• The default initial files were based on ERA5 dataset.
• Yes, I will add the figures or at least refer to the tables with the performance measures in the appendix.