The paper is well designed, nicely written, and presents a novel way to downscale precipitation data by combining semivariogram modeling into spatial random forest with well-known precipitation predictors.  Although I consider the paper can be published "as is", I have suggestions for the authors:

- The proposed method could be given a name.

-  In  5.5. Further Research,  the authors could also address some of the following:

       The upscaling from 1 km to the 10 km IMERG grid was done by pixel averaging. If some other aggregation stats were used (eg. median, max, mode), would the performance of some of the environmental predictors improve (eg. aspect)?

      Some other predictors could also be included, such as position based on metric distances instead of latitude-longitude, EVI (with is better related to water content in plants or soil than NDVI), wind or atmospheric pressure features.

    The proposed method could be transferred to use with CHIRPS or TRMM data in other parts of the world, particularly in large tracts of South America, which have complex topography and sparse gauging stations.

Easy-to-use spatial Random Forest-based downscaling-calibration method for producing high resolution and accurate precipitation data

This study used a random forest machine learning algorithm to downscale the GPM satellite precipitation measurements and calibrated them with gauge observations and the aforementioned high-resolution variables. The study is interesting and the authors presented some interesting results, while the context is hard to follow and the writing needs significant improvement. The tense of the article is very confused. The present tense and past tense are confused throughout the article, and there are also many grammatical mistakes. There are too many mistakes to point out here. It is suggested that the paper shoukd be revised by someone who is good at English.

Major comments:

• You have the daily measured precipitation data, and you can also get daily GPM precipitation data. Why don't you analyze the downscaling model on the daily scale? In fact, there have been many daily downscaling models based on Random Forest. In general, the novelty of the article is not enough.
• The author did not give a clear explanation why the environment variables could be used in RF model to derive the correlations. The relationships between these variables and precipitation should be explained in detail, not just by listing the reference papers that use these variables. The correlations between each environmental variables (NDVI, LST, DEM etc.) with precipitation and their contribution to the prediction of the precipitation should be fully discussed.
• The IMERG data were fused into precipitation products from the satellite observations and gauge data. As the gauge data had been applied in the IMERG data generation (Level3), how could you reconcile the errors in SRF with the gauge data at ground used in your study?
• Evaluation of the model was performed by considering rain gauge data as observations at ground. However, it assumed that the rain gauge measurements were representative values at their respective grid-cells. Although this was widely used in other literatures, the authors should discuss this issue to support their decision.
• This paper analyzed the accuracy of various precipitation data sets on a monthly, seasonal, and annual. There is no need to spend a lot of space in the paper to analyze the differences between seasonal and annual results, which makes the paper look more like a technical report without scientific insights. You should compare with previous literature that states if the results are in agreement or not with other studies in the field in your Discussion part. However, the present part of the discussion is rather empty, and it mainly analyzes the research results of others, and lacks the systematic analysis and discussion of the proposed method in this paper

Minor comments:

• Line 158-160: “Overall, the high spatial and temporal variability of precipitation with the complex topography makes the study site ideally suitable for the evaluation of satellite-based precipitation estimates.” Can you give the reasons or some references as evidences?
• Section 5.1. Please indicate the optimized hyper parameters (i.e., number of trees, depth of the tree, and number of features) for the Random Forest model.
• Line 269-271: “the spherical model was used since it shows better results than the others in the experiments.” You should give the analysis results (data, charts, etc.) to prove the reason for choosing the spherical model.
• Line 315-316：“Our monthly-based estimation method was compared with the annual-based SRF fraction disaggregation method (termed as SRFdis)” Please explain SRFdis in detail.
• As can be seen from Figure 5 and Table 2, the proposed method and Bi SRF have very similar performance in most accuracy index, such as CC. However, the bilinear interpolation downscaling method is obviously easier to operate than SRF downscaling method. Is it necessary to use a more complex downscaling method to improve the CC value of 0.003?
• Line 404：Change “our method” to “the proposed method”
• Line 432：Why do you say that “This is because this year has the largest precipitation (Fig. 2).” Can you explain why this is the reason for the worst results in 2018?
• Line 488: “Table 1”? Table 1 is the detailed information of the datasets used in the study.
• Line 499-500：What are the reasons that the day-night land surface temperature difference was used in this study?

The proposed article is focused on an interesting question: the improvement of satellite-based precipitation products for the estimation of month, seasonal or annual precipitation amounts. It presents an original method aiming at improving the IMERG monthly precipitation product at local scale. The original aspect of the proposal, if compared to previously published methods consists in including spatial input variables in a random forest model: longitude, latitude and above all spatially interpolated raingauge measurements based on ordinary kriging. The authors call therefore their method “spatial random forest”.

The article is interesting and overall well written and structured, but could be improved in several ways. Moreover, it suffers from an evaluation flaw that has to be corrected to provide accurate estimates of the real performances of the tested methods and fair conclusions: i.e. the performance of the proposed method should not be evaluated on a validation set, but on a test set, totally independent from the model calibration and selection step. The confusion between validation and testing is a common error in the implementation of IA methods when cross-validation procedures are implemented for the calibration and selection of the models. This pitfall has been pointed out by numerous authors and generally leads to substantially overrate the performances of the IA models (See ref 1. and 2, hereafter). The authors should not split theirs samples into two, but three subsample: a calibration set (a) and a validation set (b) (used for the cross-validation model adjustment procedure) but also an independent test set (c) used in the final step of model assessment. This has absolutely to be modified to my opinion to provide sensible results, before the manuscript can be published in HESS. I am not convinced that if really tested on an independent data set, the performances of the proposed method remain higher than the performances of the kriging method…

Some other aspects of the method and of its presentation could be improved (see also the attached annotated manuscript):

1. Some implementation information is missing and could be added in the manuscript such as the nuggets and ranges of the variograms used for the spatial interpolation.
2. The authors should provide the names of the software and possible libraries they have used for the implementation of RF.
3. Figure 13 gives an interesting insight into the calibrated model and the driving input variables. It would be interesting, to provide an even clearer insight, to test the real added value of the input variables in the SFR model. I have the impression that the dominant variable is the kriging result and not the spatial coordinates. Could the performances of the model based on the spatial coordinate only or the kriging result only be provided for a more complete discussion. I have the impression that removing the spatial coordinates from the input variables, as well as all other terrain characteristics will have little consequences on the model result. The proposed model finally mostly consists in an intelligent merging between spatially interpolated raingauge measurements and satellite downscaled precipitation. By the way, was the downscaling step really useful (see comments in the manuscript)?
4. Likewise, ordinary kriging is a relatively basic interpolation approach. I wonder if co-kriging or kriging of residuals approaches, popular for spatial rainfall interpolation, could also have been tested. But this is probably not feasible for the revised version of this manuscript but a suggestion for future developments

Cited references:

1. Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach, 2009 (3rd edition). page 709. “Peeking is a consequence of using test-set performance to both choose a hypothesis and evaluate it. The way to avoid this is to really hold the test set out—lock it away until you are completely done with learning and simply wish to obtain an independent evaluation of the final hypothesis. (And then, if you don’t like the results … you have to obtain, and lock away, a completely new test set if you want to go back and find a better hypothesis.)
2. Xu, Y., Goodacre, R. On Splitting Training and Validation Set: A Comparative Study of Cross-Validation, Bootstrap and Systematic Sampling for Estimating the Generalization Performance of Supervised Learning.  Anal. Test.2, 249–262 (2018). https://doi.org/10.1007/s41664-018-0068-2. ”There is a significant gap between the performance estimated from the validation set and the one from the test set for the all the data splitting methods employed on small datasets”

The paper applies a machine learning technique for downscaling and calibration of precipitation based on remotely sensed inputs that also aims to incorporate the spatial structure of rainfall using spatial autocorrelation. The idea of paper is interesting and it also has a organized structure which is generally well-written. However, based on the methods applied and discussion of the results, the paper has several shortcomings that need to be addressed and further explained prior to publication.

Major comments:

1. Several aspects of the OK based interpolated maps at 1k and 10k resolutions are not fully convincing. First, the accuracy of the OK-derived maps should be reported in order to determine reliability of the maps. Errors in the interpolated maps are going to be propagated to the errors in the spatial RF model because it is one of the covariates used, so they are important. It would be interesting to see if the large RMSE’s in the middle part of the study area in fig.7 also show up with large errors or variance in the OK maps.

Related to this, the authors also need to further clarify the interpolation of a 1km image based on a 10km IMERG images using OK, which is a raster-to-raster interpolation performed (lines 273-284). A coarse to fine raster-based interpolation seems unusual, so that authors need to further describe this step.

2. The parameters tested and chosen for all the models, including the semi-variogram should be reported otherwise the study is not reproducible.

3. It appears that the sRF model (also for the other ML techniques applied) did not include a separate testing phase . This is a standard approach applied when assessing the accuracy of a ML methods. I would suggest to also validate the models using an independent test that is not used in the training phase. Or re-configure the ML methods to split the total data into a training and a test set.

4. Discussion of the results focuses more on the positive aspects of using sRF but the authors do not give a balanced view by providing a critical analysis of the results of sRF. For instance, the accuracy metrics presented highlight that sRF performs well compared to the other models. However, visual comparison of the boxplots of these metrics alone in figs. 8-9 shows comparable accuracies all the models based on their range and median. Significant differences between the accuracies obtained, particularly in relation to sRF, should be reported to provide gravitas on the authors claim that sRF outperforms the other models.

5. There is an underestimation of precipitation values regardless of the model used based on fig 5 . This should be further elaborated in addition to the three accuracy metrics provided, so the bias of the estimates should also be reported. Furthermore, for very high precipitation values (e.g. >400mm), the scatter of the points in fig.5 becomes larger, indicating that all the models tested perform poorly at v. high rainfall amounts. It could be insightful to assess separately how the models compare for v. high rainfall conditions, since prediction of these extreme cases need to be generally improved.

6. It is unclear how the importance measures are calculated from fig. 13, so this should also be included in the methodology of the paper. Furthermore, discussion of the rankings could be made more in depth by determining whether they agree or deviate (and why they do) from known controls on rainfall distribution.

7. The authors already indicate that there is a delayed response of vegetation to rainfall. It is perhaps expected that the NDVI is one of the least important factors in the sRF model. But actually, this also provides an opportunity to also explore the lagged values of the predictors (and not only NDVI) with known delayed responses to rainfall.

Minor comments:

1. The captions of the figures need to be improved. Some of the features in multi-plot figures are hard to understand because of the captions are highly simplified.

2. The final version of the manuscript will benefit for another round a English check as some sentences a phrased a bit vaguely (e.g. line 150-151)