This study has investigated the impact of simultaneously assimilating satellite soil moisture data from L-band SMAP, C-band ASCAT and X-band MWRI into the Common Land Model (CoLM) for improving soil moisture estimates. The impact was evaluated by comparing both spatial and temporal soil moisture patterns with in situ data from the ISMN and the CMA networks. While the results of the study are interesting, further details are needed regarding the methodologies employed and the results obtained. Additionally, some interpretations presented appear speculative and require stronger justification.

MAJOR COMMENTS

1.    The assertion that this work represents a 'paradigm shift' (Lines 414ff) is not supported by the existing literature. Numerous studies, including those conducted within the framework of the ESA CCI programme, have previously utilized complementary information from different satellite soil moisture sensors and developed dynamic weighting schemes.

2.    The authors do not properly describe active and passive microwave sensors, and their specifications and characteristics. Most importantly, differences in the performance of the products cannot merely be explained by differences in frequency given that SMAP and MWRI are passive sensors and ASCAT is an active sensor. This leads to wrong statements and interpretations such as:

a.    On line 72 they write: “Active microwave sensors offer finer resolution but lower temporal sampling due to long revisit periods.” This is only true for SAR, but not scatterometers. 

b.    Line 75: “L-band missions (SMOS, SMAP) penetrate vegetation well but exhibit larger errors over complex terrain.” Complex terrain is challenging for all microwave sensors, but certainly even more problematic for active than passive sensors.

c.    Line 76: “C-band sensors (ASCAT, Sentinel-1) have high temporal resolution but variable accuracy across vegetation types and seasons.” Only ASCAT has a high temporal resolution, not Sentinel-1. One can note that – while it is true that the accuracy varies by vegetation type and season – this is nothing unique for active sensors. The same applies of course for passive sensors.

d.    Line 90ff: “L-band sensors penetrate vegetation well and perform best in moderately vegetated areas; C-band sensors have moderate penetration and are sensitive to vegetation changes; X-band sensors have weak penetration and mainly capture near-surface signals.” This description is oversimplified and partly not true, e.g. ASCAT can be better than SMAP over moderately vegetated areas.

3.    There are some choices made in the assimilation (Section 3.3) which need much better justification than merely citing past research:

a.    There is no logic in the selected pertubations for the different layers. Explain.

b.    The authors state that the satellite soil moisture data are mapped into the second layer (7-28 cm). This counter intuitive given that the satellite data match much closer to the first layer (0-7 cm).

4.    The evaluation methods are not adequately described. For example, it is not fully clear what how the spatial correlation coefficients (R) and root-mean-square error (RMSE) relative ERA5-Land, shown in Figure 2, are calculated.

5.    The interpretation of why L-band assimilation decays more rapidly is extremely speculative (lines 298ff: “While L-band performs best in regions with dense vegetation and high precipitation (Mousa and Shu, 2020), soil moisture in these areas is frequently influenced by strong meteorological forcings such as rainfall. As a result, the assimilated information is more likely to be masked by subsequent hydrometeorological variability, leading to faster loss of forecast skill.”) Please provide evidences that can support this interpretation.

6.    The statement (Lines 301ff) that “In contrast, C- and X-band retrievals perform better in regions with low-to-moderate vegetation cover and more arid conditions.” is not fully correct. For ASCAT, it is known that subsurface scattering effects degrade the quality of the data in arid region.

7.    The statement that “L-band performs well in densely vegetated areas, …” (Line 315) is highly problematic as also L-band cannot sense soil moisture below dense forests.

8.    Considering the very small differences shown in Figure 11, it is hard to agree with the statement “the NEW experiment exhibits notably better performance” (Line 372).

MINOR COMMENTS

Section 2.1 must be much improved, providing more details about each data set used.

Section 2.1. Distinguish between spatial sampling and spatial resolution! E.g., the 12.5 km for ASCAT refer to the spatial sampling, not the spatial resolution (which is 25 km).

Section 2.2: Show a map with the location of the ISMN and CMA in situ stations.

Section 2.2: Specify how many ISMN and CMA stations were used in the study.

Figure 5: Not all names of vegetation types are self-evident. Describe.

Figure 9: Not clear what is shown. E.g. which sensor is R-single referring to?

The study presented in this paper investigates the individual and synergistic assimilation of satellite microwave soil moisture products from three sensors at different frequencies (SMAP at L-band, ASCAT at C-band, and MWRI at X-band) into the Common Land Model (CoLM). The skill of the various data assimilation experiments as well as the skill of forecasts initialized from the respective analyses is evaluated against soil moisture observations from the ERA5-Land reanalysis, as well as in situ soil moisture observations from the International Soil Moisture Network (ISMN) and the Chinese Meteorological Administration (CMA). The authors find that generally the combination of all three sensors yields the best performance, except over densely vegetated areas where the inclusion of X-band retrievals leads to a skill degradation. To address this, they present a vegetation-adaptive data assimilation framework that shows slightly better overall performance.

Overall, this is an interesting paper, and the inclusion of X‑band observations offers a novel and valuable perspective. However, several major issues need to be addressed before the manuscript can be considered for publication.

Major comments:

1. This paper lacks essential details regarding the methodology that are needed to properly assess and interpret the results presented. In addition, several statements within the methodology section appear contradictory, further complicating the interpretation of the findings. Specifically:

• More detail is needed regarding the preprocessing of the satellite observations. For example, what form of quality control is applied for each of the satellite datasets? Is this based on quality flags provided with each product? Are the observations bias-corrected to the model climatology? The authors mention that “the retrieved soil moisture from satellite products is mapped to the model’s second soil layer”. Does that mean the retrievals are CDF-matched to the second layer soil moisture or is a different approach chosen? Why was the second soil layer selected and what is the depth of this layer in CoLM?
• More details are needed regarding the actual data assimilation. How is the observation error for each of the three satellite products defined? The authors mention that the observation error statistics are derived from the RMSE between the open loop (CTL) and ERA5, but it is unclear how this would provide any information about the uncertainty of the three satellite retrieval products. Also please specify what quality control is applied within the assimilation system. For example, do you assimilate observations over frozen soils, when there are open water bodies present in the grid cell, or if the departures between the model background and the observations are very large?
• More details regarding the model are needed, specifically the distribution of the seven soil layers and the temporal resolution of the model outputs.
• More details are needed about the processing of the in situ station data. Was any quality control applied based on the quality flags, the observing conditions (e.g., frozen soil), or the completeness of the observation time series? How many stations were used in total? It is also unclear which in situ networks are used when. The methodology section mentions both ISMN and CMA, but the map plots in the results section (Figures 12 and 13) only show stations in China and the text for Figure 6 also mentions that only stations in China are used. Which of the figures in the results section include the ISMN data? Please label all plots clearly to show which networks were used.
• It is unclear which SMAP product is used. Line 107 states that the L2_SM_P_E product is used and the reference for this product is given, but line 108 states that the 36-km product is used. Please clarify.

2. One of the main findings from this paper is that the optimal combination of satellite sensors used in the assimilation is dependent on the observing conditions, specifically the vegetation density, which impact the uncertainty of the assimilated products. The authors address this by introducing an adaptive framework whereby the combination of satellite sensors changes depending on the vegetation density. However, in a traditional Kalman filter DA system, this sort of adaptive weighting of the impact of different observations is controlled through the observation error and its impact on the Kalman gain. So, I wonder whether a cleaner implementation would be to use spatially and temporally varying observation errors for the three satellite products, which presumably would also reduce the impact of the X-band observations over densely vegetated areas. However, in the current manuscript it is unclear how the observation errors were characterised for the three satellite products or why the vegetation adaptive scheme was chosen instead. Addressing these two questions would greatly strengthen this paper.

3. The suitability of the chosen spin-up‑ approach is unclear, as the methodology section does not provide sufficient detail to fully assess it. The authors describe a 342-‑year spin-up prior to the ‑two-month experiment period, but it is not clear whether this spin-up was conducted separately for each experiment configuration. For example, in the L+C configuration, does the spin-up already include the assimilation of L+C observations? Or was the ‑spin-up performed as an ‑open loop run with no data assimilation? If the latter is the case, then an additional ‑spin-up‑ period after activating the DA would be necessary to allow the system to adjust to the new configuration. The behaviour shown in Figure 2, specifically the continuously increasing correlation values and decreasing RMSE throughout the assimilation period, suggests that the system may indeed have been initialised from a no DA ‑spin-up‑and was still adjusting during the experiment period.

4. Generally, the experiment period of 2 months seems much too short. It does not capture any seasonal or interannual effects and raises questions about the transferability of the conclusions drawn here to other parts of the year. I invite the authors to provide a rationale for choosing such a short experiment period (is it limited by data availability?) and consider extending the period if possible.

5. For the global evaluation, ERA5-Land is used as a reference. However, ERA5-Land is forced with meteorological fields from ERA5, which does assimilate ASCAT soil moisture observations. So, there is an indirect influence from ASCAT soil moisture retrievals on the ERA5-Land soil moistures. Also, the CoLM model is forced with ERA5 meteorological fields that are impacted by the ASCAT SM assimilation. It is possible that the very good performance of the C-band data assimilation in Figure 2 is a side effect of this? I would suggest to either use a more independent reference for the global evaluation or to at least discuss the possible impact of ASCAT.

Minor Comments:

1. The naming convention in the plots is all over the place. The assimilation experiments are sometimes referred to as ‘L’, ‘C’, and ‘X’, sometimes as ‘SMAP’, ‘ASCAT’, and ‘MWRI’, sometimes as ‘DA_L’, ‘DA_C’, and ‘DA_X’. The X-band experiments are referred to as ‘X’, ‘DA_X’, ‘MWRI, or ‘FY3D’ depending on the plot. Please make sure you pick one naming convention and use it consistently throughout. I would also suggest using different names/label when you are evaluating data assimilation experiments and when you are evaluating retrievals.
2. The differences between the various experiments are often very small. For the bar graphs, I would suggest including confidence intervals (or some other measure of significance) as well as some indication of the number of stations used in the evaluation against the in situ data.                                   
3. Please include more detail regarding the satellite data that are being used. I would suggest including the data period of each sensor, the spatial resolution, the revisit time, the temporal resolution of the satellite product, and the retrieval approach used.
4. Each of the retrieval algorithms used the generate the products used here makes assumptions about the observing conditions, such as the vegetation state or the soil temperature. That means that to some extent differences in the performance observed here could be due to assumptions about for example the vegetation state made by one retrieval algorithm matching better with the vegetation state in the reference data. These effects can be limited by assimilating radiances instead of retrieval products. Could you please comment in the text why you chose to assimilate retrieval products rather than satellite radiances?
5. Please include units and colorbar labels on the plots where they are missing.

Detailed Comments:

l.18 “significantly improves” Including confidence intervals as suggested above would help support this claim.

l.25 “rises by about 0.25”. Is this compared to the open loop? Please specify.

ll.57-59: SMAP only provided active and passive data in its original configuration for a few months (until the failure of the radar). If you are referring to the SMAP/Sentinel product here, please include the appropriate reference.

l.63 Draper and Reichle 2015 uses GEOS/Catchment, not SiB2. Please adjust the text or include the appropriate reference for SiB2.

ll.66-69: For the discussion of the SMAP data assimilation I suggest also including the discussion of the SMAP Level-4 SM (data assimilation) product that is part of the official product suite (Reichle et al., 2019), as well as studies that have investigated the impact of SMAP DA in NWP systems (e.g., Carrera et al, 2019)

Reichle, R.H., Liu, Q., Koster, R.D., Crow, W.T., De Lannoy, G.J., Kimball, J.S., Ardizzone, J.V., Bosch, D., Colliander, A., Cosh, M. and Kolassa, J., 2019. Version 4 of the SMAP level‐4 soil moisture algorithm and data product. Journal of Advances in Modeling Earth Systems, 11(10), pp.3106-3130.

Carrera, M.L., Bilodeau, B., Bélair, S., Abrahamowicz, M., Russell, A. and Wang, X., 2019. Assimilation of passive L-band microwave brightness temperatures in the Canadian land data assimilation system: Impacts on short-range warm season numerical weather prediction. Journal of Hydrometeorology, 20(6), pp.1053-1079.

l.72 “long revisit periods” This is really only true for SAR or other high-resolution sensors.

l.75 “L-band missions (SMOS, SMAP) penetrate vegetation well” They penetrate vegetation better, but L-band sensors also have difficulties “seeing” the surface under dense vegetation (VWC > 5 kg/m^2)

ll.83-86: The statements made in this section are not quite accurate. There are plenty of studies that have investigated the synergistic assimilation of soil moisture from different sensors, for example:

Draper, C.S., Reichle, R.H., De Lannoy, G.J.M. and Liu, Q., 2012. Assimilation of passive and active microwave soil moisture retrievals. Geophysical Research Letters, 39(4).

Lievens, H., Reichle, R.H., Liu, Q., De Lannoy, G.J., Dunbar, R.S., Kim, S.B., Das, N.N., Cosh, M., Walker, J.P. and Wagner, W., 2017. Joint Sentinel‐1 and SMAP data assimilation to improve soil moisture estimates. Geophysical research letters, 44(12), pp.6145-6153.

Girotto, M., Reichle, R.H., Rodell, M., Liu, Q., Mahanama, S. and De Lannoy, G.J., 2019. Multi-sensor assimilation of SMOS brightness temperature and GRACE terrestrial water storage observations for soil moisture and shallow groundwater estimation. Remote Sensing of Environment, 227, pp.12-27.

Additionally, there are many studies investigating the complementary synergy of different satellite retrieval products, for example in the context of the ESA-CCI project. Finally, the studies that are cited here do indeed investigate the complementarity of different sensors. I suggest rewriting this section to acknowledge the existing body of work, whilst also highlighting the novel contributions from this study (e.g., the inclusion of X-band data).

l.93: The Kerr et al., 2010 reference is not about X-band sensors, so maybe not an appropriate reference here.

l.151 Please include the citation for ERA5:

Hersbach H, Bell B, Berrisford P, et al. The ERA5 global reanalysis. Q J R Meteorol Soc. 2020;146:1999–2049. https://doi.org/10.1002/qj.3803

ll.155-159: It is unclear to me whether this text refers to ERA5 or the CoLM. Please clarify.

l.167 Please also include de Rosnay et al., 2013 as a reference.

De Rosnay, P., Drusch, M., Vasiljevic, D., Balsamo, G., Albergel, C. and Isaksen, L., 2013. A simplified extended Kalman filter for the global operational soil moisture analysis at ECMWF. Quarterly Journal of the Royal Meteorological Society, 139(674), pp.1199-1213.

l.196: Could you please include some motivation for conducting the assimilation once daily?

Figure 3. It is a bit difficult to see the differences between the different products. You could consider changing this figure to show the ERA5-Land soil moisture and then three maps showing the differences of the three DA experiments with respect to ERA5-Land. Please also include a label and units on the colorbar and change the title of the bottom right map to “ERA5-Land”.

l.256: How was the vegetation density computed?

Figure 5: If these are organized according to increasing vegetation density, then why is “bare soil” not in the first slot? And since you are showing differences, please label the plots as ΔRMSE etc. (same for Figure 6).

ll.269-270: Why did you choose not to include ISMN in this evaluation?

ll.286-288: I do not see the described behaviour in the plot. Could you please clarify what you mean?

ll.288-289: Qualitatively, ASCAT and SMAP seem to show a much more similar performance than ASCAT and MWRI.

ll.297-304 I am not sure how you can draw these conclusions about the impact on forecast skill when you have not shown the forecast skill across different land cover or climate regions.

Figure 9: The y-axis label here is confusing. If I understand the text correctly, you are comparing the skill of a 3-sensor combination against the skill of a 2-sensor combination, not a single sensor. Please clarify this.

Figure 10: A number of comments on this figure:

• The whole discussion here seems to contradict the findings from Figure 9, where the conclusion was that adding X-band only had minimal impact on the skill or even degraded the performance.
• Could you please clarify how the three stations that are shown here are selected? Could you include information on where they are located? And also, please comment on or show how the impact of adding X-band at these stations compares to other stations of the same land cover.
• Please define the ‘AWS’ acronym for the in situ data in the caption or just label the green line ‘in situ’.
• The blue and black dots for the C- and X-band retrievals are difficult to tell apart. I suggest using different marker symbols to make it easier to distinguish them.

ll.367-369: Could you please include information on how the threshold for including X-band observations was defined?

Figure 12: I do not quite follow why a different reference was used for the central-west and south-east regions in the bottom plots. Could you please clarify?

Figure 12 and 13 In the discussion of both figures the focus seems to be on the skill improvements of NEW compared to the CTL. I think it would also be relevant to discuss here how NEW compares to the L+C+X assimilation experiment that does not use the vegetation-adaptive approach.

General Comment:

'A preliminary study on a synergistic assimilation scheme for multi-band satellite soil moisture data' presents the assimilation, on a global scale, of three soil moisture retrievals from the SMAP, ASCAT and MWRI satellite instruments into the Common Land Model. The three observations differ significantly in terms of the spectral bands used, measurement swathes, resolution and other factors. The study discusses the pros and cons of co-assimilating these products with regard to modelled soil moisture. In the field of geosciences, there is a common debate about the benefits of assimilating complementary or redundant observation datasets.

The results of assimilation experiments involving one, two or three of these observation datasets are carefully compared. The skill of each experiment is evaluated in relation to ERA-5 land reanalysis and in situ global soil moisture observations taken at a depth of 10 cm. From these comparisons, the authors highlight significant regional differences in the experimental results and link them to the distribution of vegetation types, justifying this by the sensitivity of L-, C- and X-bands to dense vegetation, and of C- and X-bands to low-to-moderate vegetation cover and more arid conditions. This evaluation per vegetation type was then used to evaluate the synergies induced by the assimilation of different observations. The authors concluded that the co-assimilation of SMAP+ASCAT offers the best overall improvement and that the addition of MWRI-X-band should only be applied to areas with sparse vegetation to avoid any anti-synergistic effects. 

While these conclusions are understandable, there is a major issue with the experimental setup. All experiments were conducted over a consecutive two-month period only. This means that any seasonal effects on vegetation, soil moisture, observation availability and quality could not be considered. Without information on how the analysis performs at other times of the year, the links between regional differences, vegetation types and recommendations for co-assimilating different observation datasets remain unclear.

For this study to be considered for publication, either the simulation period should be extended, or a more thorough investigation into the relationship between localisation, vegetation type and observation availability should be conducted.



Major remarks:

• l 106 - You only selected June to August 2022 as your study period. Why did you choose such a short period? Moreover, 2022 was an unusual year in Europe, with severe heatwaves, droughts and wildfires in the summer.
• l 121 - These satellites are in a sun-synchronous orbit. How do you schedule assimilation with the SEKF all over the globe? You say that you assimilate all observations at 00:00 UTC daily. Do you only consider observations available at that specific hour, or do you move all observations from throughout the day to 00:00? If so, can you justify that this has no significant impact on the soil moisture diurnal cycle?
• l 131  - Please justify the depth taken for this study.
• l 150 - Could you clarify whether your LAI is prognostic (modelled) or forced? If it is forced, which climatology database or observation product are you using?
• l 185 - Please describe the values sets in B and R error covariance matrices.
• Figure 3 - Could you show the bias maps averaged over your assimilation period using ERA5-Land instead? That is what you are describing in the text. This would make it easier to read.

Minor remarks:

•  l 43 - ... to improving model performance -> to improve forecast skills of the model. The initial conditions do not change the model.
•  l 93-97 - You discuss the complementarity of the observations used according to their vegetation sensitivities, stating that you will capitalise on this to engender synergistic improvements. However, as will be seen later in the article, complementarity between observations does not always lead to synergistic improvements. 'Complementary' does not mean 'in agreement with each other'. What you are discussing here is the necessity of a screening step to select the observations to assimilate to avoid anti-synergistic behaviour. The terms 'to engender' synergistic improvements feels awkward, you should only expect them to occur.
• l 105 - Precise the frequency of the product (daily, 10-days,...).
• l 119 - Please detail the retrieval model you used for MWRI or cite it.
• Figures - units please