General comments

The paper presents a framework for estimating SM at 1 km resolution combining reanalysis SM (ERA5-Land, ~10km) and other variables using the XGBoost algorithm. I find it an interesting approach and certainly there is an increasing interest from the hydrological community to have similar products from various applications. 

However, to make it even more robust, I would suggest the authors to further elaborate on the following points:

- The comparison with the SMAP/S1 product is very interesting and, in principle, of high value. However, I believe that a more thorough assessment would be beneficial, as currently there is no quantitative comparison and the analysis only takes into account one single day. The authors could, for instance, calculate spatial correlation among the 2 datasets for each 12-days overlapping samples and plot the corresponding time series. 

- Similar to the previous point, I think it would be valuable to present a time series of spatial correlation between the GLASS SM and the ESA CCI products – in addition to the analysis presented by the authors. 

- As ERA5-Land SM is the dataset being downscaled, it should be included a comparison of the GLASS and ERA5-Land SM products (as advocated in Gruber et al., 2021; Crow et al., 2022). Do we see a degradation in terms of temporal metrics (correlation, ubRMSD, ..) when looking at GLASS SM compared to ERA5-Land? Or does it improve even the temporal dynamics? 

specific comments

- line 83-85: there is also an operational SM product from Sentinel-1 generated for Europe (https://land.copernicus.eu/global/products/ssm, Bauer-Marschallinger et al., 2019)

- Table 1 is not exhaustive; for instance, CGLS SSM and the EUMETSAT H-SAF SM (https://hsaf.meteoam.it/Products/ProductsList?type=soil_moisture) products are missing

- line 176: there is a more recent paper about the ISMN (Dorigo et al., 2021)

- Section 2.6: here the revisit time of the SMAP/S1 product is not specified. Please add. 

- Figure 2: I would use a different color for the training stations – purple might be not so effective depending on the background color. Also, it would be interesting to know the distribution with respect to climate classes; maybe the authors could add a table with % of land cover and climate classes

- Table 3 indicates that accounting for spatial variability is more important than accounting for the temporal component, as found in previous studies (e.g. Zappa et al., 2019).

- Figure 4: for better readability and comparison of feature importance, I would combine the 3 plots into a single one with different colors for the 3 test experiments.

- Table 4: rename the input datasets, eg coarse SM (ERA5-Land), vegetation and LST (GLASS), terrain, soil texture.

- Figure 5: create different subplots for the different metrics, as the spread of values on the y axis is rather wide.

- Figure 9: invert or change the colorbar, as it might be a bit misleading at first sight to see that high correlation (good) is shown in red, but also high RMSD (poor result) has the same color.

References

Bauer-Marschallinger, B., Freeman, V., Cao, S., Paulik, C., Schaufler, S., Stachl, T., Modanesi, S., Massari, C., Ciabatta, L., Brocca, L., Wagner, W., 2019. Toward Global Soil Moisture Monitoring With Sentinel-1: Harnessing Assets and Overcoming Obstacles. IEEE Trans. Geosci. Remote Sens. 57, 520–539. https://doi.org/10.1109/TGRS.2018.2858004

Crow, W.T., Chen, F., Colliander, A., 2022. Benchmarking downscaled satellite-based soil moisture products using sparse, point-scale ground observations. Remote Sens. Environ. 283, 113300. https://doi.org/10.1016/j.rse.2022.113300

Dorigo, W., Himmelbauer, I., Aberer, D., Schremmer, L., Petrakovic, I., Zappa, L., Preimesberger, W., Xaver, A., Annor, F., Ardö, J., Baldocchi, D., Bitelli, M., Blöschl, G., Bogena, H., Brocca, L., Calvet, J.-C., Camarero, J.J., Capello, G., Choi, M., Cosh, M.C., van de Giesen, N., Hajdu, I., Ikonen, J., Jensen, K.H., Kanniah, K.D., de Kat, I., Kirchengast, G., Kumar Rai, P., Kyrouac, J., Larson, K., Liu, S., Loew, A., Moghaddam, M., Martínez Fernández, J., Mattar Bader, C., Morbidelli, R., Musial, J.P., Osenga, E., Pal-ecki, M.A., Pellarin, T., Petropoulos, G.P., Pfeil, I., Powers, J., Robock, A., Rüdiger, C., Rummel, U., Strobel, M., Su, Z., Sullivan, R., Tagesson, T., Varlagin, A., Vreugdenhil, M., Walker, J., Wen, J., Wenger, F., Wigneron, J.P., Woods, M., Yang, K., Zeng, Y., Zhang, X., Zreda, M., Dietrich, S., Gruber, A., van Oevelen, P., Wagner, W., Scipal, K., Drusch, M., Sabia, R., 2021. The International Soil Moisture Net-work: serving Earth system science for over a decade. Hydrol. Earth Syst. Sci. 25, 5749–5804. https://doi.org/10.5194/hess-25-5749-2021

Gruber, A., De Lannoy, G., Albergel, C., Al-Yaari, A., Brocca, L., Calvet, J.-C., Colliander, A., Cosh, M., Crow, W., Dorigo, W., Draper, C., Hirschi, M., Kerr, Y., Konings, A., Lahoz, W., McColl, K., Montzka, C., Muñoz-Sabater, J., Peng, J., Reichle, R., Richaume, P., Rüdiger, C., Scanlon, T., van der Schalie, R., Wigneron, J.-P., Wagner, W., 2020. Validation practices for satellite soil moisture retrievals: What are (the) errors? Remote Sens. Environ. 244, 111806. https://doi.org/10.1016/j.rse.2020.111806

Zappa, L., Forkel, M., Xaver, A., Dorigo, W., 2019. Deriving Field Scale Soil Moisture from Satellite Ob-servations and Ground Measurements in a Hilly Agricultural Region. Remote Sens. 11, 2596. https://doi.org/10.3390/rs11222596

This paper presents a gap free 1 km global soil moisture data set from 2000 - 2020 created using a XGBoost model. Such a data set is certainly relevant for the community and the manuscript perfectly fits the scope of ESSD. The setup of using XGBoost on coarse-resolution model data together with different high-resolution variables is certainly innovative and a sound attempt to improve spatial information content over other existing products. I also very much appreciate the effort the authors have put into clear langauge and a well structured manuscript.

I only have a few comments regarding clarity and the validation approach, which I hope the authors will consider to revise the manuscript:

General comments:

- Correlations of soil moisture and in situ data above 0.8-0.9 are completely unrealistic and strongly suggest over-fitting. The authors certainly tried to avoid overfitting using different strategies (year- and site-independent sampling), but I think the manuscript can benefit from a more explicit discussion on the issue of over-fitting, XGBoost's proneness to it, and when you do and do not suspect it to occur in your particular study.

- I general, references are appropriate, but often come a bit late. For example, triple collocation and XGBoost are already discussed in the introduction, but references to these methods only two sections later. I recommend referencing always upon first occurence. Along the same lines, I would provide the map of used ISMN locations already in the data sets section where I expected it. 

- Soil moisture behaves very differently at point scale and at a 1 km scale, that is, they are in fact expected to exhibit different means and standard deviations just because of the scale difference. Training on point data somewhat imposes these point-scale properties on the data, so evaluating the model results in terms of bias, RMSD, and ubRMSD makes little to no sense, because comparing these metrics e.g., with those attained by ERA5 doesn't really say anything about which of these is actually more accurate, at the 1 km scale. This should be at least discussed.

-  When considering the case of least potential overfitting (i.e., the site-independent validation), evaluation results, at least in terms of temporal correlation (which are actually the relevent ones; see above) are VERY close to those attained by ERA5, which begs the question of why bothering with the sophisticated ML in the first place, and not use ERA5 directly? Perhaps better discuss the actual benefits of your product.

- To me, the distinction between Sec. 4 and 5 isn't clear. It feels like Sec 5 is just a continuation with more results. Also, why hasn't the use of two different model setups not been discussed in the methods section? That comes a bit out of the blue.

Specific comments:

- L65--: It's a strong claim that all global products have poor accuracy over densely vegetated areas, which I wouldn't consider to be true (considering that "densely vegetated" draws too vague a line). I recommend to delete this sentence

- L71 I suspect "combining" should be "using" or something similar? "combining", in that sentence, doesn't make sense. 

 -L70-85: Somewhat a personal preference, but I believe this paragraph is largely a repetition of what is already discussed at length in Peng et al. (2020). For brevity, I'd remove it and just refer to this paper.

L145: Blank missing before "study" and the opening bracket.

-L149: Typo: EAR5 (same in Fig. 6)

-L177: Old link. Should be "https://ismn.earth/"

Sec 2.5: Aren't these data sets also in the ISMN? Also, it'd be nice if Fig 2 would have a different color coding for training stations and these independent networks.

L211: Insufficient referencing for the CCI product; see https://esa-soilmoisture-cci.org/node/236

Sec. 3.2: Perhaps add references to previous work that has used TC as an indicator for spatial representativeness of in situ stations.

Sec. 3.3: Perhaps list the used hyperparameters

L337: I didn't undertand the difference between "representative stations selected using TC" and "stations excluded using TC". 

- L425-430: I am not convinced that your validation approach proves that your product actually contains more spatial detail. As mentioned in the general comments, you mainly show that biases relative to point scale measurements are smaller, but a 1 km product SHOULD be biased with respect to that scale; showing only that biases are smaller than those with ERA5 doesn't exclude the possibility that the ML isn't over-doing the bias correction. 

Fig. 7; Why not showing any metrics for the SMAP / GLASS SM comparison?

Fig. 9: Why not comparing ERA5 also to the CCI product, or SMAP? 

- L626: Arguably, the "relatively lower site-independent validation accuracy" is MUCH more realistic. I would rather attrobute the higher accuracy in the other cases to over-fitting.

- L665: It's a circular argument that "future studies should focus on improving the accuracy of soil moisture products at higher resolution" to then train your model to create more accurate soil moisture products at high resolution. 

- L687: This R value is clearly a result of overfitting.

- L699: As discussed above, reducing biases is not necessarly a meaningful goal when comparing 1 km products to point scale products.

- L705: remove "the"

General comments:

This paper presents a study of high-resolution seamless soil moisture estimation based on the machine learning model and multi-source datasets, which can fill the gap of the lack of long-term global soil moisture products with both high spatial and temporal resolutions and benefit a range of applications. The subject is of interest to the scientific community and the description is clear.

Three complementary model validation strategies (random, site-independent, and year-independent) were adopted by the authors, which avoids over-fitting to a large extent and makes the model more robust. I found this manuscript is well structured and suggest for acceptance after addressing the following concerns, partially related to the models and validation of the soil moisture data obtained.

Specific comments:

1. Why did you choose the XGBoost model? Have you ever tried other machine learning models (Random Forest) or deep learning models (e.g. LSTM) besides XGBoost?
2. Although the ISMN data are used as input, in most areas where there is no ground observation, the EAR5-Land soil moisture reanalysis data is the main input factor of this algorithm. The importance of the ERA5-Land data is extremely high. Does this mean that the accuracy of the ERA5-Land data itself is one of the most important factors affecting the accuracy of the developed data set? If so, the author should explain it in the text. At the same time, in view of the extremely high weight of ERA5-Land, does it mean that it is the basis or background value of the proposed 1-km data, while other parameters are only corrected in a local range (0.1°*0.1°)? If so, how is this method fundamentally different from other machine learning-based soil moisture downscaling methods?
3. While comparing the coarse scale global soil moisture product with the 1-km GLASS soil moisture product (Section 4.4), in addition to evaluating the spatiotemporal consistency between them, the authors are suggested to add a zoomed-in plot to demonstrate the superiority of the GLASS product in terms of spatial resolution.
4. To make Table 1 more comprehensive, the other soil moisture products can also be added to this table, for example a recent publicly released global product which was generated using multi-channel collaborative algorithm:

https://doi.org/10.11888/Terre.tpdc.272907

https://doi.org/10.11888/Terre.tpdc.272088

There are also others such as the SMOS 1-km soil moisture data and the Copernicus soil moisture data from Sentinel-1 (https://land.copernicus.eu/global/products/ssm) cover Europe and a similar study at:

https://doi.org/10.11888/RemoteSen.tpdc.272760

1. The location of independent in situ validation datasets (section 2.5) is also suggested to add into Figure 2.
2. What I don't quite understand is that the importance of the factors obtained through the random test, year-independent test, and site-independent are not consistent (Fig. 4), indicating that the models trained by the three methods are different, which set of models is finally used for the final soil moisture data set production? Which should be explained in the manuscript.
3. In the verification part, the author used the data from four ground observation networks to show that compared with the ERA5 data (Fig. 6), the overall bias and RMSE of the obtained 1-km soil moisture are smaller, which is a gratifying conclusion. However, the type of land cover represented by these networks are extremely limited, so the reviewers believe that this verification might not sufficient enough.
4. How does the RMSD used in Figure 9 differ from the RMSE metric should be clarified in Section 3.4.
5. In Table 5, please explain why “urban” is not excluded from the soil moisture data. Moreover, the difficulty of estimating soil moisture in forests and barren is greater than that of grasslands and croplands. However, in Table 5, there is not much difference in the accuracy of different types of results. Why? Perhaps the author can add a column of "Number of Sites" in Table 5, so that readers can better understand the accuracy of the model in various land cover types.