This paper develops SGD-SM 2.0 framework for reconstruction of seamless global daily soil moisture dataset from 2002 to 2022, based on the development of the LSTM-CNN method. The method fuses the global daily precipitation products and is able to consider sudden extreme weather condition. Generally, the topic is interesting, the method makes sense and the results are supportive. Some minor comments before positive publication are as follows.

1. Please provide the parameters and descriptions of AMSR-E, AMSR2, and WindSat to demonstrate the rationality of using three heterogeneous sensors.
2. The original SM data acquired from AMSR-E, AMSR2, and WindSat sensors were used to generate the seamless soil moisture dataset. As we know, although the frequency band of both AMSR-E/2 and WindSat have the same frequency band to retrieve the soil moisture, the GHz of WindSat sensor is different from that of AMSR-E/2. Is there a big difference in accuracy between 2011.10.5 to 2012.07.02 (i.e., using WindSat) and other periods (i.e., using AMSR-E/2)?
3. Section 4.3 only exhibits the dynamic change of the SGD-SM 2.0 dataset. Perhaps, the advantages of version 2.0 can be demonstrated by introducing version 1.0 as the reference in this section. Moreover, this revised description is different from Section 5.2 (the time series in the precipitation area).
4. Please provide the website for collection of in-situ data, if the data are public.
5. Why the partial CNN and mask updating were used to reconstruct the missing regions?
6. What is the relation between the two sub figures in Fig. 1?
7. Please consider including more up-to-date references on gap filling, such as [Remote sensing image gap filling based on spatial-spectral random forests. Science of Remote Sensing, 2022, 5: 100048].
8. Check the caption of Fig. 6.

This manuscript presents a novel study on use of three sensors AMSR-E, AMSR2 and WindSat to reconstruct SGD-SM 2.0 products. One of the novel aspects of this study is that global daily precipitation products are wisely assimilated into the proposed LSTM-CNN, to fill gaps in daily soil moisture products. This methodology represents a substantial advancement in generating global soil moisture products that synergistically incorporate soil moisture and its closely associated hydrological variable, precipitation from the last precipitation satellite, i.e., Global Precipitation Measurement. The improved SGD-SM 2.0 product has been shown to outperform the previous SGD-SM 1.0 product in terms of accuracy and time-series consistency. I recommend accepting this wonderful work after minor revision.

-Page 2 Line 24: AMSR2 and WindSat products in caption (a) and (b) are incorrect.

-Page 3 Line 64: Word ‘description’ is repetitive in this sentence.

-Page 5 Line 101: IMERG precipitation products should be given the full name.

-Page 9 Line 185: Why did the authors use the global land mask M_L in the loss function?

-Page 12 Line 215: ‘drawing into global daily precipitation products’ should be revised as ‘assimilating global daily precipitation products.’

-Page 16 Line 281: The authors claimed that the reconstructed SGD-SM 2.0 points behave more consecutive around their adjacent original soil moistures points than SGD-SM 1.0. More explanations need to be given for this attribute.

-Page 17 Line 297: Data availability. Current descriptions about SGD-SM 2.0 in the website are not Data availability. The authors may want to supplement specific information for possible users.

This paper presents an improved seamless global daily soil moisture dataset from 2002 to 2022 (SGD-SM 2.0) based on the three satellite soil moisture sensors AMSR-E, AMSR2 and WindSat and Global daily precipitation products. A new convolutional neural network approach is used to fill the gaps and missing regions and ISMN data is used for the validation of SGD-SM 2.0.

The topic is of general interest due to the increasing drought and overexploitation of water resources in many regions of the world due to global climate change and fits well within the scope of the ESSD. The applied methods mostly appropriate and the manuscript is mostly well written but contains some incorrect wording and phrasing (see specific comments). My main concern is that the authors used only six stations for the validation of the global SGD-SM 2.0 data set, which is not inappropriate. The authors should make an effort to test whether SGD-SM 2.0 data accuracy is independent on the environmental conditions. The SGD-SM 2.0 data product would be well received by the science community working on Global Change issues and can be recommended for publication after all issues detailed below have been appropriately addressed.

General comments:

The authors only show the averaged evaluations indicators from all selected ISMN stations and only six in-situ soil moisture stations were actually used for the validation of global SGD-SM 2.0 data set. In my view, this is not enough to appropriately demonstrate the accuracy of a global SM data set. There much more data is available at ISMN. In addition, other in-situ soil moisture data products are freely available, e.g. Bogena et al. (2022). In this way, potential users could also see if the SGD-SM 2.0 data accuracy is independent on the environmental conditions, e.g. soil properties, vegetation coverages, climate zone.

Some soil moisture data shown in Fig. 8 show extremely SM high values of more than 80 Vol.%. Such high values are very unlikely, as soil porosity in most soil is typically between 40-50 Vol.%, indicating measurement errors in the in-situ data or soils with extremely high organic matter or clay content. Indicating a reference site description will help to understand this better. On the other hand, the SGD-SM 2.0 data the same high values, which is astonishing. In my view, these data outliers could be the result of SM overestimation by the CNN procedure due to the precipitation consideration. In addition, single outliers can be found in Figs. 9d and 10a. Again, this indicates the influence of precipitation. Maybe the data should be cleaned with an outlier detection method? Please add at least a discussion on these issues.

The in-situ soil moisture data from ISMN are treated anonymously in this work. However, the site owners that work hard to maintain the soil moisture stations should be better cited. This will help the site owners to ensure funds for the costly operation of the stations and data management. Therefore, the authors should add at table with basic information on the soil moisture data using, including the name of the site owners and/or monitoring networks instead of just presenting the station coordinates. See Bogena et al. (2022) for a great example. The necessary information is available in the metadata descriptions at ISMN.

Throughout the manuscript, you use the term “assimilation” in the context of including precipitation data in your CNN based data interpolation method. However, I think this is not appropriate as the term “data assimilation” is generally used optimally combine numerical models with observations.

Specific comments:

L17: Please cite the more recent ISMN publication of Dorigo et al. (2021)

L21-22: Incorrect phrasing.

L23: Change to “approximately 20% to 80%”

L30: “words”

L31: Change “destroys” to “degrades” or similar

L35-36: Citation is missing.

L43-44: Reads awful, please rewrite.

L56-57: Incorrect phrasing.

L70: Please mention the source of the in-situ data.

L78: The GES DISC website should be referenced

L85: Reads awful, please rewrite.

L97: Please cite the more recent ISMN publication of Dorigo et al. (2021)

L103: Change here and elsewhere to “long and short-term”

L130: Change to “soil moisture and precipitation products”

L132-133: Can you estimate the average time scales of the long and short-term memories and their variabilities? It would be interesting to know how different the time scales are.

L190: The term “epoch number” should be explained.

L290: Change to “the soil moisture time-series of”

Figure 11: Please show the precipitation in reverse order and as bar chart, which the standard way of presenting precipitation and much better to understand.

Literature

Bogena, H.R., M. Schrön, J. Jakobi, P. Ney, S. Zacharias, M. Andreasen, R. Baatz, … and H. Vereecken (2022): COSMOS-Europe: A European network of Cosmic-Ray Neutron Soil Moisture Sensors. Earth Syst. Sci. Data 14: 1125–1151. DOI: 10.5194/essd-14-1125-2022

Dorigo, W., I. Himmelbauer, D. Aberer, L. Schremmer, I. Petrakovic, L. Zappa, W. Preimesberger, A. Xaver, F. Annor, J. Ardö, D. Baldocchi, M. Bitelli, G. Blöschl, H. Bogena, … and R. Sabia (2021): The International Soil Moisture Network: serving Earth system science for over a decade. Hydrol. Earth Syst. Sci. 25: 5749–5804. DOI:10.5194/hess-25-5749-2021

This paper addresses the emergent need to increase soil moisture information access, quality, and quantity for multiple users and applications. The authors present an interesting study about reporting a new data version of a seamless global soil moisture product that increases both the quality and accuracy of the previous version of this product. The methods are sound and novel, particularly the development of a deep learning algorithm to fill daily gaps in soil moisture estimates. The authors compare the old and new versions of the datasets, and they provide a robust quantitative accuracy benchmark between versions. 

The paper is generally well written. However, from the narrative, I feel that there are missing technical details. For example, the use and role of the three passive microwave sensors in modeled soil moisture values in the presence of the precipitation dataset is unclear. Also, can the authors elaborate on prediction variance or model-based uncertainties? I feel that uncertainty of estimates is commonly not presented in soil moisture gap-filling efforts despite being helpful for assessing the reliability of soil moisture predictions. 

It is also my opinion that the accuracy limitations or advantages of the new product version are relative to the reader. For example, the authors poorly discuss their accuracy findings against previous research or gap filling efforts of satellite soil moisture estimates across scales.

The first version of the product has a relatively good number of citations, meaning that the community uses the product and that the methodological approach is being compared with similar research. The authors provide a thorough comparison between product versions, but they do not present the discussion of findings against previous research. I would appreciate more discussion about the potential implications of using the product's old or new version in multiple applications in terms of other available soil moisture estimates. 

The paper leaves the value of this product relative to the reader as the comparison is made only between versions one and two, and it does not consider the large availability of other soil moisture estimates for multiple uses and applications. Many (hundreds if not thousands) studies currently report alternatives to downscale or fill gaps in satellite soil moisture data. I invite the authors to provide a more extensive literature review and discussion of previous research to support the value of their product. 

I invite the authors to discuss the main implications of accuracy metrics to assess modeled soil moisture values. Can the authors describe the accuracy of the soil moisture sensors used? I invite the authors to use community accepted standards to report errors on soil moisture products, e.g., ubRMSE https://www.sciencedirect.com/science/article/pii/S0034425720301760, and discuss the accuracy of the soil moisture modeled values against other products or gap-filling efforts. A simple demonstration of the new knowledge that users can obtain from the new product would increase substantially the value of this excellent modeling framework applied to soil moisture satellite estimates. 

Finally, but more importantly in my opinion (considering that this is a dataset journal), please consider publishing your code in order to fulfill the FAIR principles and contribute to the open-science culture transparently e.g., https://bg.copernicus.org/preprints/bg-2021-323/bg-2021-323.pdf. .

Specific comments

L30 I recommend to avoid the word ‘destroy’ as aggregated soil moisture values are useful for multiple applications (e.g., to constrain long-term Earth system models). The cited references do not deal with gap-filling daily soil moisture values, please revise. 

L35 sentence relative to the reader, can the authors be more specific and quantitative and include supporting references, e.g., which quantitative indexes?

L4065 Consider combining each weakness or limitation in v1 with their corresponding advantages in v2,  instead of two separated lists.

L83 how they are employed?

L98 What is the criteria to select those sites?

L102 It seems to me that the authors solve a regression problem (where soil moisture is a response of precipitation and time) using deep learning, but they use the word assimilation, which is relatively fine for me given how the algorithm they use works. However I recommend to elaborate on the concept of data assimilation applied here for a better and broader understanding of narrative flow.

L108 what soil moisture product? the authors use three products  

L170 40x40 what?

L119 What was the criteria to select those 124 sites? Can the authors provide a map of points showing in colors the correlation between in-situ and their product for all the stations? I like the presented information but this is a global product and I think it will be useful to interpret the reliability of the product elsewhere.  Also for bias indicators (MAE, RMSE), it would be nice to see a map of errors to identify areas with high or low quality of predictions. Please consider also the ubRMSE as it has been a widely discussed metric validating local to global soil moisture predictions. Discuss please the values of accuracy metrics in this and other products. 

L225 can the authors highlight these points in figure 3b? 

L285 the temporal resolution depends on the application.

Figures 9 and 10, consider using lines instead of points.