A Hybrid Artificial Neural Network to Estimate Soil Moisture Using SWAT+ and SMAP Data
, Scott C. James 1,3,*,†, Joseph D. White 4, Peter M. Allen 1 and Jeffery G. Arnold 5Round 1
Reviewer 1 Report
This paper has demonstrated that the neural network approach was better able to encapsulate the complexity and variability of weather time series and hydrologic data dynamic systems than even a computationally expensive physically-based model.
The topic is interesting and matches well for MDPI Mach. Learn. Knowl. Extr. journal.
The paper contains a review of related works and shows good simulation studies.
However, the paper has some unclear points and the following minor concerns.
- In fig. 5,6,7 it is difficult to discern the nature of the error. The authors should consider adding inserts with enlarged fragments of graphs.
- You must also provide the ability to separately download code and data. At the moment, the file size for downloading is more than 3 GB
- L485 “https://storage.googleapis.com/breen_make]_supplementary/Breen_supplementary.7z” → “https://storage.googleapis.com/breen_make_supplementary/Breen_supplementary.7z”
Author Response
Please see attached PDF for a point-by-point response to each reviewer's comments.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The authors present two approaches to using ANN to determine soil moisture. I know a little less about ANN than the authors, so I will not comment except to say that it looks good within the bounds of my knowledge. However, there are some concerns about the methodology, and serious concerns about knowledge of the relevant aspects of hydrology.
Major concerns:
- The last statement in the abstract is unsupportable. A long-term trend is usually much more subtle than seasonal changes.
- Line 31: This is true, yet there are also models that do these things. This statement is misleading.
- Line 28: The statement that weather characteristics and soil characteristics are independent was highly controversial last time I checked, and is certainly not true in all cases. If this is a critical assumption, please add the appropriate caveats and propagate any changes through the paper.
- Lines 86 and 87: The results were evaluated in part against data that was included in the training set. This is extremely bad form, making any conclusions about accuracy invalid. This must be fixed (or somehow justified) prior to publication.
- Accuracy cannot be meaningfully assessed based on figures 6 and 7. Please make a quantitative assessment including rms differences, biases, best fit slopes or similar diagnostics. A scatterplot (or density plot) would be most welcome to visually show the effectiveness of the technique. The work is unpublishable without a reasonable assessment of accuracy.
Minor concerns
- Introduction, line 27: Soil moisture is NOT a flux, and hence not a key component in a hydrological flux. It is a key variable. While this appears to be an English error, it suggests to the reader (and reviewer) that the authors have little knowledge of hydrology (or fluxes for any field of study).
- Define acronyms, like ML, before they are used.
Author Response
Please see the attached PDF for a point-by-point response to each reviewer's comments.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
The authors are to be complimented for responding well to the vast majority of the concerns. The improvements are quite substantial, and greatly enhance the value of this contribution. I still find the writing style (not grammar) to be awkward and to reduce clarity. I encourage the authors to have non-experts read the paper, and point out parts that they find confusing, and to improve the clarity in those sections.
I also suggest that the authors state the interpretation R^2 values, as they strongly support the value of this work. I also urge that they note that results with substantially different slopes (e.g., PGLF in the 3rd row of Fig. 7) can still have large values of R^2 (the departures from the mean are proportional), which is why both the R^2 and RMSE are used as diagnostics.
The conclusion could be improved to be much more specific about these results. That would likely have a large impact on the perceived value of this work. Some excellent points are made earlier, and it would be good to see them more clearly presented in the conclusions.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
