The manuscript “Evaluating a Global Soil Moisture dataset from a Multitask Model (GSM3

v1.0) for current and emerging threats to crops” presents a LSTM-based machine learning model for global soil moisture estimation. The authors used a multi-loss framework to train the LSTM model, and employed multiple reference datasets for evaluation. Specifically, the work investigated spatial generalization ability by cross-validations, both randomly and continent-based sampling. The overall quality of the work is excellent and fits in the scope of GMD. With that, I do have the following comments. 

The title is somewhat confusing to me. It indicates the authors also care about crops in addition to soil moisture, as soil moisture will affect the crops. So I was expecting to read something about agricultural applications (e.g., use the DL-based soil moisture to drive some crop models). However in the manuscript, it is only mentioned before the conclusion section that this model will be put into production, and the major focus of the paper is model benchmarks on soil moisture. I suggest the authors have more discussion on crop applications, or revise the title to make it clearer. In addition, if the main focus is to evaluate the dataset instead of the LSTM model, I suggest making the dataset publicly available and adding a link to access this dataset.  

To clarify, what’s the difference between the SoMo.ml model and the authors’ LSTM model? Do they share the same dynamic and static input variables with the only difference as the loss function? In Figure 1 and 2, the authors compared the metrics derived from the training period (Multitask_train and SoMo.ml_train). The R value or RMSE from the whole training period is not important, as one can always overfit the model. I understand that a barplot for Multitask_train shows the model is not overfitting, but a comparison between Multitask_temporal and SoMo.ml_temporal would make more sense to me. 

What is ubRMSE in Table1? Does Corr represent Pearsonr correlation coefficient? In Figure 3 and 4, the meanings of colormaps are not the same. In the correlation map, greener represents better performance, but it is worse performance in the RMSE plot. I would suggest a uniform colorbar with light (dark) colors for high performance and dark (light) colors for low performance. 

The LSTM model is optimized towards both the in-situ estimation and the satellite products. In Figure1 when comparing the model performance, the authors selected the SMAP and GLDAS products, which are gridded datasets. When evaluated against the ISMN dataset, it is not a fair comparison because of the coarse resolution of SMAP. How do the authors correlate in-situ measurements with gridded datasets? Will the result change when switching to 25-km resolution (i.e., with coarser resolution, the dataset will lose more representations)? A further question is that, what is the meaning of the LSTM outputs? Is it the best estimation over the grid points (e.g., an average over the grid spacing), or the best estimation for the in-situ observations? 

The authors split the global dataset into 7 continents. Would it be a more straightforward comparison to have a similar 7-fold cross validation to match the number of continents in Section 3.2? I believe 5-fold is a common choice but 7-fold may be a more intuitive and fair comparison. 

When analyzing the factor controls, the authors selected a random forest model. What is the performance/skill of this random forest model? Is Figure 5 using spatial R as target or temporal R, or is it a multi-output model? At line 329 it shows the target is temporal R, but line 330 shows R is from either temporal or spatial tests. 

Regarding the random forest model, is the conclusion independent of the model choice (or what’s the reason to choose a random forest model)? Will we get different results if we switch to extreme boosted trees or other tree-based models? Also please check the name of the python package. It is used as “sklearn” in python but the paper referenced (Pedregosa et al., 2011) shows “scikit-learn” in the title.  

The author mentioned the driest sites are hard to predict, and suggested the reason as scarce but sudden rainfall events. Do authors believe it may be related to the loss function used in the LSTM model (i.e., some loss functions would not emphasize the extreme high/low values)? I also don’t follow the logic behind Line 345. From the error type analysis, the authors mentioned the comparison between temporal and spatial tests, especially for error type B (nonstationary). But the boxplots of R from the temporal test and spatial test over driest sites are similar to me, with a median R of approximately 0.7. 

Line 146 mentioned the use of different resolutions for the static terrain attributes. What is the aspect resolution used in the random forest model?  

The code files in the Zenodo repository are not enough to replicate the experiments. I’d suggest the authors update their repository, either during or after the peer-review process. 

This paper presents a long short-term memory (LSTM) recurrent neural network model for use in global predictions of soil moisture. The model learns from both satellite-based remote sensing and in-situ soil moisture data sets, adopting a multi-task learning approach that optimizes a linear combination individual loss functions that are computed using satellite products (SMAP L3 at 9 km nominal resolution) and in-situ data (from the International Soil Moisture Network). The authors present systematic comparisons of the model with the SMAP product, process-based land surface models, and another LSTM model (SoMo.ml) that is trained exclusively with in-situ observation network data to demonstrate the utility and limitations of the model.

The subject of this paper is a good fit for GMD, and the paper’s quality is excellent, with very good reproducibility, model evaluation, discussion, and writing. I have little to suggest that could improve the paper without expanding its scope (which would be unfair to the authors and is not necessary). I have only two minor criticisms / comments to mention:

1. I believe that the title is misleading because of the phrase “for current and emerging threats to crops”. I understand that one of the main motivations for developing this model is for it to serve as part of what might be termed an early-warning system for threats to crops in Africa. However, this system does not yet appear to be fully in production, or, if it is, the authors have not given a full description or examples of its application. An accurate soil moisture model has clear benefits (and possibly is an unavoidable necessity) for such a system. But since there is no substantial discussion of how the model is being used in this context, I suggest either dropping the phrase “for current and emerging threats to crops” or modifying the title to change the “for” to “with potential application to”. (I see that Anonymous Referee #1 has raised the same concern.)
2. I was confused by “Normalized Difference Vegetation Index” (NDVI) being listed as one of the static landscape attributes in the paragraph beginning at line 147. I have worked extensively with NVDI time series (in the context of emerging threat detection, incidentally) and it is something that I think of as the antithesis of a static landscape attribute (outside of a place like a barren desert environment). I had to look at the author’s code in their Zenodo repository to see that this is actually NVDI averaged over some long time period (though I didn’t dig enough to figure exactly what this time period is or how the average is computed). I would appreciate it if the authors would amend the paper to add something about what sort of NVDI average is being used here.

I must apologize to the authors for not having been able to review their paper sooner. Before I got to it, I see that Anonymous Referee #1 did an excellent job of providing feedback on the paper, and he or she has already mentioned some of the other concerns or questions that I had. I do not need to repeat these here, as I see that the authors have already responded appropriately to the referees comments.

Minor copy editing comment:

• The second-to-last sentence of the abstract, “Factorial analysis shows that the LSTM model’s accuracy with representing impacts of terrain aspect, resulting in lower performance for dry and south-facing slopes or wet and north-facing slopes”, does not make sense to me. I believe it needs to say something like “Factorial analysis shows that the LSTM model accuracy can be strongly impacted by terrain aspect, resulting in…”