Thank you for providing me the opportunity to review the entitled manuscript " A two-step merging strategy for incorporating multi-source precipitation products and gauge observations using machine learning classification and regression over China". It’s an interesting study and fits the scope of the Journal. However, there are several major and minor flaws that authors should take care of and revise before final consideration in high-quality peer-review Journals like Hydrology and Earth System Science.  Authors should improve the quality of the manuscript including research outcomes, discussion, and unique conclusions.

My major comments are as follows;

1- In the abstract, there are too many simple conclusive statements in the abstract, which are well-known to people who are involved in this area. You should first layout a background information description. Second, point out what is the most important in the current field of research and what has not yet been solved. Third, explain your novel method in detail how you solve this problem. Last, a brief description of your own findings should be presented. If your methods and findings are novel and interesting to people who are involved in this area, they will continue to read your main text.

2- It is recommended that the author rewrite the INTRODUCTION, increase the citation of the literature, and extract questions and useful information from the literature. Through literature review, point out the shortcomings of existing research, thus leading to the article's hydrological and environmental significance and purpose. In this section, the literature review needs to be more critical.

3- The authors should detail the methodological novelties with the vast amount of existing literature in this area in the Introduction.

4- The authors should address the clear objectives of the study in the introduction section.

5- In the discussions, comparisons of the results obtained in this manuscript with the extensive existing literature on Satellite-based precipitation and the methodologies used need to be expanded. I recommend authors should compare results with previous approaches. In discussions, it must add what the results mean with respect to what is already known and highlight how your results support or refute the current hypotheses in the field if any. More references should be added to that section. Underline how your results make a significant move in the working field forward.

6- It’s is recommended to improve the quality of grammar and take care of grammar mistakes.

7- Line# 115-120 Expand the hydro-metrological features of the study region with more explanations.

8- Line# 230-240 I am wondering if the technique applied by authors is correctly classified for wet and dry events. Authors only attempted to correct precipitation events fall in wet events. I recommend techniques should test with combined dry and wet days because measurement techniques for all precipitation datasets are quite different and definitely there would be a lag between wet and dry events for all different datasets which could create outliers in applied techniques.  

9- Authors have merged ground observations data with different precipitation estimates. While different precipitation products have different spatial resolutions which could cause outliers when merging with ground observations.  Apart from these, the authors used a simplified approach to merge coarse resolution precipitation products with ground observations data. Rainfall within a single satellite pixel could vary considerably by 38% between two gauges located within 4 km × 4 km. The deviation between gridded precipitation and single gauge observation is due to the discrepancy of scale and could be reduced by increasing the validation stations or downscaling the TRMM precipitation to a finer resolution. Therefore, authors should firstly downscale all grided precipitation datasets at a finer scale and then merge with ground observational data to make the approach more novel towards environmental significance. The author can learn lessons from the following papers.

Arshad, A., Zhang, W., Zhang, Z., Wang, S., Zhang, B., Cheema, M.J.M. and Shalamzari, M.J., 2021. Reconstructing high-resolution gridded precipitation data using an improved downscaling approach over the high altitude mountain regions of Upper Indus Basin (UIB). Science of The Total Environment, 784, p.147140.

Gebremichael, M. and Krajewski, W.F., 2004. Assessment of the statistical characterization of small-scale rainfall variability from radar: Analysis of TRMM ground validation datasets. Journal of Applied Meteorology, 43(8), pp.1180-1199.

Harmsen, E.W., Mesa, S.G., Cabassa, E.D.V.I.E.R., Ramírez-Beltran, N.D., Pol, S.C., Kuligowski, R.J. and Vasquez, R.A.M.Ó.N., 2008. Satellite sub-pixel rainfall variability. International Journal of Systems Applications, Engineering & Development, 2(3), pp.91-100.

10- The study region covers complex hydro-topographical features and some of the stations are located in snow and glacier coverage regions (e.g., Tibetan Plateau) and hence observed precipitation in these regions is unreliable and unavailable. Therefore, the orographic correction of precipitation based on the vertical gradients along with glacier mass balance is required to retrieve an accurate precipitation dataset in high-altitude mountain regions such as Tibetan Plateau and some others. Authors can take glacier mass variations from GRACE data and try to correct precipitation for high-altitude regions.

11- Add a limitation section that explains any limitations that your hypothesis or experimental approach might have and the reasoning behind it and some of them I have clearly mentioned. This will help the field in generating hypotheses and new approaches without facing the same challenges. The discussion becomes well-rounded when you emphasize not only the impact of the study but also where possibly it falls short. Consider posing a few questions or directions, preferably in the form of a hypothesis, to provide a launchpad for future research.

12- Conclusions need to revise and well-round major and significant findings of current study

The article titled "A two-step merging strategy for incorporating multi-source precipitation products and gauge observations using machine learning classification and regression over China" assesses the effectiveness of three machine learning-based algorithms to merge precipitation products over China. The article is very well written, concise, relevant, and I enjoyed reading it. I believe that it fulfills the requirements to be published in HESS. In the following points, I describe some points and suggestions that I think can be considered to increase the quality of the manuscript.

ERA5 has already superseded ERA-Interim; therefore, I recommend using ERA5 instead. Similarly, the authors used TMPA 3B42 (with IMERG), which is no longer in production. For reproducibility purposes, it is essential to use products that can still be acquired.

I suggest including a figure showing the number and location of stations used in the GPCC product in the appendix or supplement material. This information will be beneficial for the readers.

I liked that the authors obtained the meteorological data from GLDAS and not from ERA-Interim or ERA5. This can help avoid introducing biases towards the ERA product (although they used the cloud cover from the ERA product).

It would be helpful to describe whether the ground-based data were quality controlled.

The readers would benefit substantially from an improved description of the methodology. I liked Section 3.1 as it is very clear and informative, as well as the description of the machine learning algorithms and their respective figures. However, I believe that the description of the two-step merging strategy (Section 3.2) can be further improved. The authors mention that first, the gauge observations are classified into wet and dry days to be used as the benchmark for classification models. Can the authors explain what the reason behind this is? It is not clear how the dry days were separated from wet days. How is a day classified as dry? Is the classification performed by grid-cell and by day? What is the final result of this classification? How can data scarcity affect this classification? Solving these questions is crucial for the understanding of the manuscript. Additionally, the authors mention that a regression process was applied to the wet days for the cold and warm periods. How were the models trained independently for the warm and cold periods? Perhaps it would be helpful to do a diagram for the first and second steps where the process is clearly explained. This will increase the manuscript's impact and is essential as the article presents these novel merging procedures.

I liked the idea of using the semivariogram as a spatial autocorrelation variable :). Can the authors discuss the influence that the selection of a particular semivariogram model can have in applying the method?

For reproducibility purposes, please mention the parameters of the RF, GBDT, and XGBoost that were used while training the models during the warm and cold periods.

The authors evaluated the performance of two categorical indices (CSI and HSS) over different precipitation intensities. This is a very good idea as the detection of no-precipitation events mainly masks the categorical performance. In this sense, as the objective of this article is to assess and compare the effectiveness of merging precipitation products using different ML techniques, I suggest evaluating all categorical metrics over these precipitation intensities. This separation into rain intensities will provide additional insights regarding the performance of these merged products.

I want to congratulate the authors for their figures. They are very nice and informative :)

The first two sections of the Results are a bit puzzling for me. In Section 4.1, "Performance assessment for classification results", the results obtained during the first step of the merging procedure are shown. However, in this section, the authors evaluate precipitation intensities (see Fig 6), which I believe, according to the methodology, are the result of applying the regression models over the wet days. Later on, in Section 4.2, "Performance assessment for regression results", the authors mention that regression models predict the final results presented in this section. By improving the explanation in the methodology, these two sections will be much clearer, and this issue can be solved.

I liked Section 5.2. I think that it will be helpful to perform a simple statistical test to assess whether the performance of the different methods is statistically significant or not.

Can the authors discuss the influence of the gauge density in the performance of the merging methods? This will be very helpful for the readers as in many regions, there are fewer rain gauges than those used in this study. This can give an insight on how to apply these techniques over diverse areas with different gauge densities.

I also liked the experiment performed in Section 5.1, where different approaches are evaluated and inter-compared.

The Discussion section could be improved by comparing the results of the present research with the wider literature. This is important because it will help to highlight the novelty of the approach. Similarly, the authors can discuss the implication of using products with different spatial resolutions and the representativeness of the rain-gauge stations with their respective grid-cells.

L355: significantly is a statistical loaded term, which must be accompanied by its respective p-value. If the p-value is not provided, I suggest using the word "substantially" instead. Please apply this throughout the manuscript.

L395: The authors mention the following "...although Kriging exhibits better performance than original MSPs, its accuracy is strongly dependent on gauge density. This only gauge-based interpolation method would have limited in complex mountainous areas with few gauges." I would be cautious with this statement. Although I agree with it, the ML algorithms cannot predict values outside of their training range, which could be translated into plausible underestimating precipitation over high elevations. Additionally, these techniques are also affected by the size of the training sample. Therefore, the ML techniques, in a sense, have limited performance in complex mountainous areas with few gauges as well.

L400: CC or r is also a component of the KGE. Also, see L412. I suggest including it inside of the KGE.

Figure7: Nice Figure! I think that it should be KGE instead of KEG.