Scientific/Detailed Comments: Referee comments in bold and author answer non-bold.

In this manuscript the authors described assimilation of soil moisture observations into the land surface model CLM5. In general, the paper is well structured, described and written. I have several questions which are still unclear for me.

1) as you mentioned in the end of section 2.1, you leave the data gaps. I'm wondering your observation frequency. 

The observation frequency varies among data sources, e.g. FLUXNET provides data in 30-minute intervals. However, as mentioned earlier in section 2.1, we use daily averages of the soil water content, evapotranspiration, and sensible heat for the assimilation and analysis. We will add more information on the observation frequency to the section.

2) for parameter updating in 2.3.2, you update the solid fractions instead of hydraulic conductivity. Please explain the reason. Does it work better, or what's the particular reason for this?

We performed data assimilation using a direct approach to update combinations of soil hydraulic conductivity, porosity, and soil matrix potential but found that the results from indirectly updating sand, clay, and organic matter fractions using pedotransfer function to provide better results. Specifically, we encountered issues with the direct update that caused parameters to approach their limits instead of reasonable values that did not occur with the indirect approach. In contrast, updating the solid fractions provided in general more stable parameter estimates and thus better simulation results. We will add more information on the rational for updating the solid fractions instead of hydraulic conductivity.

3) line 175: how about the observation errors vs. perturbations?

(Line 175: “Only the PFT were manually assigned for each site. For the ensemble creation, the fractions ofsand, clay, and organic matter are modified for each ensemble member. The perturbations are normally distributed with mean zero and a standard deviation of 10%.”)

For the data assimilation the observation error is assumed be constant and set to a RMS of 2%. We will specify this in the revised version.

4) line 190: parameter updating refers to fractions of sand, clay and organic? Are there any other parameters included, e.g. the vG parameters and porsity? If not, why do you only update these parameters?

Yes, parameter updating refers to the sand, clay, and organic matter fractions. We use the indirect approach in which the soil hydraulic parameters are calculated from these characteristics using the Clapp and Hornberger (1978) pedotransfer function as mentioned in 2.3.2. CLM5 uses the Brooks-Corey parameters and not the vG parameters as mentioned in section 2.2. As already mentioned in comment 2), we found that the indirect approach via pedotransfer functions provides better results than direct updating of hydraulic parameters.

5) On the DA effect on ET: I guess that in your experiment DA will also indirectly correct ET through assimilation of soil moisture. However, ET is more directly influenced by e.g. LAI. Do you try assimilating other observations and update other state variables through DA? This might help.

We agree that assimilating LAI will most likely improve the ET estimates more than the assimilation of soil moisture, as mentioned in section 4.4 and 5. We want to extend our assimilation routine to be able to also use LAI observations in the future to update parameters of the vegetation module in CLM5. However, in this particular study, we wanted to show that soil moisture assimilation does not improve ET characterization, even not if local high-quality in-situ soil moisture observations are available, in contrast to the studies mentioned in the introduction that mostly used satellite-based observations of soil water content. We will improve our motivation to make this clearer.

Scientific/Detailed Comments: Referee comments in bold and author answer non-bold.

This reviewer had two critical concerns about this manuscript related to the simulation of leaf area, and the design and implementation of the parameter estimation:

First, I don’t see how the author’s came to the conclusion “[the] results suggest that state-of-the-art LSM such as CLM5 still suffer from uncertainties in the representation of soil hydrological processes in forests, e.g. deep root water uptake, uncertainties in the representation of biological processes of tree transpiration”, without accoutning for biases of the simulated leaf area for the sites. Running CLM5-BGC can lead to erroneous leaf area values, and is why CLM users often use CLM-SP (prescribed leaf area) to diagnose to what extent the simulation of leaf area is impacting your soil moisture and ET relationship. Have a look at Li et al., (2022) or Fox et al., (2022) as an example of how prescribing or assimilating observations of leaf area can effect the representation of carbon and water cycling. https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021MS002747

If the simulated leaf area is incorrect in magnitude and timing it is unlikely that improving the soil water content through DA will improve ET. The authors do not account for this in the manuscript.

We agree that running CLM5-BGC can lead to erroneous LAI simulation results and that CLM-SP can be used to diagnose to what extent this uncertainty affects the simulation of soil moisture and ET. Therefore, we will also explore the difference between CLM-SP and CLM-BGC simulations and add the results in the revised manuscript. Depending on the results, we will also rephrase the statement criticized in the reviewer comment.

Second, in reference to the parameter estimation, the authors update the soil layer fractions sand/clay/organic which then impact to what extent the true soil hydraulic properties are weighted in the soil layer. It is a strange approach to access the ‘true’ global hydraulic parameters through a fixed land surface property. This would be equivalent to treating the PFT for each site as a ‘parameter’ rather than a fixed/prescribed property, and varying the PFT fraction across MF/ENF/DBF/ENF/WSA for each site, which would then weight the parameter set associated with each PFT. This is not done, and the authors don’t do that here. Instead the author’s prescribe the PFT for each site and keep that fixed. Similarly, it is unclear why the author’s would then not prescribe the sand/clay/organic layers either from the default CLM files, or prescribe the sand/clay/organic fractions from the FLUXNET site data.

Soil texture data are always subject to uncertainties and therefore it is reasonable not to assume them to be fixed. Therefore, Tthe indirect approach of updating soil characteristics, i.e. sand, clay, and organic matter fraction, to update the soil hydraulic parameters using a pedotransfer function has been used already in various studies (Naz et al., 2019; Han et al., 2014; Baatz et al. 2017).

Secondly, we also performed data assimilation using a direct approach to update combinations of saturated hydraulic conductivity and porosity, but found updating sand, clay, and organic matter fractions provided in general more stable parameter estimates and thus better simulation results. One reason for this finding is that the pedotransfer function keeps the soil hydraulic parameters reasonably correlated to each other. We will add more information on the rational for updating the solid fractions instead of soil hydraulic properties.

Furthermore, the choice of parameters to be estimated should more closely relate to processes controlling the SWC-ET relationship, such as stomatal conductance, and the *vegetation* hydraulic parameters specifically involved in the Plant Hydraulic Stress formulation in CLM5. The soil hydraulic parameters estimated in the manuscript, seem to control the dynamics of moisture within the soil layer, but not the transfer of water from root to leaf, which is controlled by the stomatal conductance and PHS (Kennedy et al., 2019). Since the author’s are already adjusting for SWC, adjusting parameters inherent to the soil hydraulics seems somewhat redundant.

We agree that we could improve the SWC-ET relationship better by updating vegetation hydraulic parameters and will do so in further studies. However, for this particular study we chose to use the parameter updating to improve the soil water content estimation as much as possible and analyze the effect of this improvement to the related evapotranspiration estimation. This is related to a central objective of this study: investigate whether precise in situ soil moisture measurements can improve ET characterization, as it was found in several studies that more uncertain remotely sensed soil moisture data were not able to do so. Notice that assimilating soil moisture data with the aim to improve modelling land surface processes is a common strategy.

Finally, the authors never provide figures/tables of the influence of the DA on the parameter values. Do the parameter values converge to a new value, or are they temporally changing and random? Without this information it is impossible to assess why the parameter estimation had little influence on ET.

We agree that we did not provide enough information about the parameter updates, since we did not see it as a focus of this study, and will include more information in the revision.

Detailed comments:

Line 55-60: The literature review seems a bit incomplete. Site level studies where parameters were hand tuned for the site of interest for forested sites also seems relevant: Duarte at al, 2017; https://doi.org/10.5194/bg-14-4315-2017 and Raczka et al., 2016; https://doi.org/10.5194/bg-13-5183-2016.

We focused in our literature review more on studies using data assimilation and agree that we can extend the literature review to include references to studies using manual tuning of parameters. We will update the literature review section accordingly in the revision.

Line 61: crop cover, not cover crop

The referenced study is about the inclusion of the management practice of cover crops, i.e. crops specifically planted to cover the soil during winter to reduce soil erosion, soil compaction, and nitrogen leaching and to increase agricultural productivity by nitrogen fixation. Therefore, we think the term “cover crop” is correct here.

Line 69: “The point scale measurements use invasive equipment and the specific measurement volume, exact depth of the sensors, number of sensors, and number of stations varies from site to site. For a few sites we use soil water content measurements from Cosmic Ray Neutron Sensing (CRNS) from the COSMOS-Europe data set (Bogena et al., 2022).”

Seems worthwhile to diagnose the in-situ observations with the CRNS data for the same sites. This could help quantify representation of point level measurements for a flux tower.

We agree that a study comparing the in-situ observation with the CRNS observations for sites where both data exist would be useful. However, in this study we focus on the data assimilation and effect on modeled evapotranspiration and not the direct comparison between observational methods. This is already investigated and documented in other papers.

Line 73: “The CRNS provides continuous and non-invasive soil water content measurements over a spatial footprint of hundreds-of-meters and integrates from the surface to a depth of 10-70 cm vertically in the soil (Zreda et al., 2008; Köhli et al., 2015).”

If the COSMIC measurements are non-invasive, how are they integrating the soil water content from 10-70 cm vertical depth? Clearly these are not just measurements, but modeled output that is likely constrained from the surface measurements (1-5 cm). What are the uncertainty values for the CRNS values? More details are needed.

The CRNS measure neutrons as proxy for soil water content and use conversion functions and weighting to provide soil water content in 10-70 cm vertical depth. The uncertainty of CRNS-derived soil moisture varies not only with the different neutron detectors but also with the number of counts in a time period and thus results under lower soil moisture conditions are more accurate. Detailed description of the methodology and uncertainty of CRNS is provided by the referenced studies and we will add more clarification in the revision.

Lines 65-83: Seems odd the authors would not mention their previous manuscript Strebel at al., (2022) in this review section (Wustebach catchment), which essentially uses the same exact DA setup, but expands the analysis to more sites in this analysis. Seems the authors should state their previous findings here, and use that to motivate and distinguish this analysis from the previous one.

The previous study mentioned here focused more on the implementation details of the software coupling and therefore we reference it in the method section. However, we agree that we can also mention the study in the introduction to motivate this study and will change this in the revision.

Line 97: ‘daily average soil water content data are assimilated’. Need more detail of what depth these observations are taken from. This needs to be explained either here or more clearly in the methods.

We describe the vertical layout in section 2.4.1. However, we will add more information in the revision.

Table 1: I think elevation would also be relevant to report here.

We will add elevation to table 1 in the revision.

Figure 1: Would also be helpful to include information about data source in this figure too, similar to Table 1.

We will clarify the source of the data in Figure 1.

Line 116: ‘partitions’ is strange in this context. Suggest ‘CLM5 simulates sensible and latent heat flux for both vegetated and ground fluxes’

We agree and will use the suggested text in the revision.

Line 120: same comment as above, suggest: ‘canopy evaporation is represented as the sum of steam and leaf evaporation as a function of temperature.’

We will use the suggested text in the revision.

Line 127: How many ensemble members were used to sample the model uncertainty?

We used an ensemble of size 96 for all our simulations. We will add this clarification in the revision.

Line 144: “For simulations assimilating CRNS, H assigns the mean observed SWC to all the layers down to the measurement depth.”

It is unclear what this means. Guessing from previous statements that CRNS is the integrated water content from 10-70 cm, so are you comparing against the CLM modeled soil layers that coincide with 10-70 cm? You need to be more explicit of how you calculate your observation operator in general. What CLM soil layer variables are you using to calculate the expected observation? Also for the point measurement of soil water content, where are the depth measurements? I don’t see this mentioned in this section where it would be appropriate.

I see you describe the FLUXNET soil water content in lines 176 through 179, but it seems out of place. This description should come earlier and discussion of the observation operator should be in one place. The statement ‘For the CRNS sites, the measurement depth for each individual measurement is calculated following Schrön et al. (2017) and is included in the dataset from Bogena et al. (2022).’ Seems to conflict with an earlier statement that the CRNS integrates between 10-70 cm. depth.

For CRNS the footprint depth varies with soil water content. The cited “10-70” cm is the usual range of that footprint depth. However, since the reference data set provides the measurement depth for each observation the depth for H also varies. We will move the description of the vertical discretization and add more information to an earlier section in the revision.

Line 159: “By default, CLM5-PDAF updates soil hydraulic parameters through changes to fractions of sand, clay, and organic matter and the pedotransfer function of Clapp and Hornberger (1978).”

This statement and methodology is problematic. See my comment in summary section above.

See our comment above.

Line 162: It seems you are allowing the % sand, clay and organic to vary by layer? Does this make physical sense? This would allow for each layer to vary independently with no fixed bounds other than 0 or 100%? I cannot find any results that show the adjusted parameter values.

The sand, clay, and organic matter fraction already vary by layer even in the default CLM5 surface files. This makes physical sense, especially for organic matter that is usually very high in the first few layers and rapidly decreases for any deeper layers. However, during parameter updates the layers are updated using the Kalman gain which ensures that spatial correlation in soil texture between soil layers is taking into account. We will add more details on the updated parameters in the revision.

Line 174: You are perturbing the % sand, clay and organic to generate ensemble spread, but you are also estimating these parameters at the same time. How do you maintain ensemble spread, or prevent the system from collapsing on a solution?

It seems you address this in Line 192, but seems out of place to mention it here. Should come earlier.

Yes, the sand, clay, and organic matter fractions are both perturbed to create the ensemble spread and updated. However, the atmospheric forcing variables are also perturbed and create ensemble spread and prevent a complete collapsing on a solution. We will move the description of the damping to an earlier section in the revision to make this clearer.

Line 226: “This indicates that the SWC-ET relation is incorrect for these sites. A possible explanation is the water uptake by roots in the deeper layers is underestimated for forest sites, as also suggested by Shrestha et al. (2018).”

I don’t see how you can come to this conclusion without diagnosing your leaf area (LAI) mismatch. See my general comment/concern above.

We will rephrase the statement in the revision.

Line 315-325: Exactly. Prescribing LAI through CLM5-SP or assimilating LAI observations is necessary to demonstrate that soil water content does not improve ET, and that it is a parameter problem.

We agree that assimilating LAI will improve ET. However, we still think that it is noteworthy that data assimilation of high-quality in-situ soil water content provides no improvement to ET estimation for forested sites with CLM5-BGC.

Line 337: “These results suggest that state-of-the-art LSM such as CLM5 still suffer from uncertainties in the representation of soil hydrological processes in forests, e.g. deep root water uptake, uncertainties in the representation of biological processes of tree transpiration, partly related to uncertain vegetation parameters.”

I disagree, at least I don’t think your results demonstrate this. The inability to prescribe leaf area, or account for the assimilation of leaf area (Fox et al., 2022) does not allow you to make this statement with certainty.

We will rephrase the statement in the revision to account for the lack of LAI adjustment in this study.

Furthermore the parameter estimation performed here does not seem to directly update ‘true’ vegetation hydraulic parameters (e.g. vegetation hydraulic parameters; Kennedy et al., 2019) to further test this theory. You adjust soil hydraulic parameters which could influence soil moisture dynamics, but seems redundant in that soil moisture is already adjusted directly.

It also would have been useful to include the prior and posterior values for your parameter estimation. Its unclear if the lack of impact the parameter estimation had on ET, was because of lack of sensitivity to controlling the ET, or because the parameters were not updated significantly from their prior values.

Yes, we did not take vegetation hydraulic parameters into account in this study as we only update soil hydraulic parameters which is not necessarily redundant but usually provides increased improvement in SWC estimation. However, we agree that it is not enough for ET estimation and we will take vegetation parameters and LAI assimilation into account in our future studies.

Figures 2-4: You could combine these plots to present a better comparison between OL-DAS and OL-DASP.

We separated the Figures 2-4 for visual clarity and to avoid visual noise. We will add direct comparison Figures in the revision instead.

Figure 11: This figure seems very critical, given the potential mismatch between observed and simulated LAI, and the potential impact that has on the soil water content vs. ET relationship. I feel like you also need to present the average season cycle of LAI for all these sites, because the the annual average is not enough. CLM has trouble simulating the timing of seasonal phenology of LAI especially for evergreen forested sites (not just magnitude). Would help to simulate a CLM5-BGC and CLM5-SP simulation here, where LAI is prescribed, or assimilate LAI observations directly for your experiment.

We will add a figure with the seasonal LAI cycle for each site and include a comparison with simulations where LAI is prescribed as mentioned earlier. In addition, we will work with LAI assimilation in future studies.

Figure 12: These plots could be useful to diagnose the behavior, but I am afraid information is ‘washed out’ when averaging over all years and all sites. This brings up some really important questions, such as, did the assimilation adjust the SWC all the way through the root zone, or was the adjustment primarily superficial, and limited to the upper 25-50 cm. only? Is the majority of the root zone below this, and contribute to the lack of impact on ET?

We show results averaged over all years and sites to give an overview on general trends in the results. However, we agree that we can see more with individual figures and will provide some additional figures in the revision. The data assimilation adjusts the SWC for all layers, however as shown in this figure the change is usually larger in the upper layers than in deeper layers.