Integrating Satellite Imagery and Ground-Based Measurements with a Machine Learning Model for Monitoring Lake Dynamics over a Semi-Arid Region

Round 1

Reviewer 1 Report

Lines 73-75: Revise the sentence.

The literature lacks the literature covering the machine learning applications for similar purposes. I could not find the novelty of this research. Please highlight the literature gaps and the contributions made by this study to the body of knowledge.

What I see from Figure 1 is that the lakes’ surface areas significantly different. Please discuss how this affect the attained results? How the authors can explain the impacts of hydro-meteorological variables in this regard?

I suggest adding a table that summarizes the properties of utilized remote sensing products (Landsat, CHIRPS, MODIS). Compared to giving within the text, this can provide better overview regarding the used tools.

I am not sure that Section 2.3.1 Supervised classification is necessary. Please incorporate the corresponding into the Section 2.3.2. Also, the titles of 2.3.1 and 2.3.2 are same. Please adjust it.

Provide more information regarding the Random Forest method including their weaknesses (being prone to overfitting) and how you dealt with these challenges. Why you used random forest over either more simplified (e.g., decision trees) or complex (e.g., XGBoost) tree-based ensemble algorithms? Enhance this section with supportive materials. Also, in my opinion, Figure 2 is not necessary as the RF method has already long been adopted and similar figures widely recognized in the literature.

What is “non-values”? Do you mean missing values?

I could not find any information regarding the hyperparameters of the RF algorithm. What are their values? How did you find them? Did you apply any optimization technique in order to configure your RF structure?

There was little information about the inputs used in the study. Regarding the statements between L:318 and L:323, please explain how many of the total number of samples are urban, forest, grassland, etc. Also, was there an imbalance between these classes, and if so, how they affect the results? This part is not clear for me. Please explain if they do not have impact on the results.

Please provide more explanations regarding Figure 9 and provide more discussion about the huge difference in the determination coefficients.

Incorporate the limitations of this study into the conclusions and provide implications for further works.

 Overall, please go through the entire manuscript to eliminate the typos and try to improve the language of the paper. 

Author Response

We, the authors, are thankful to the reviewers for their very constructive and fruitful comments. which were focused primarily on the methodology part and improving the results section. Recommendations supplied in the review report were carefully followed; the text was revised accordingly. 
The comments were simply copied from the referee report and pasted below after which our responses are given in red. Changes made in the manuscript are given in red.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper uses Landsat imagery between 1999 to 2021 to estimate lake area extent in two geolocations, the Great Salt Lake and Lake chad. The lake area extent estimation were done using a random forest supervised classification. Inter-annual variability of lake area extent is compared against trends in precipitation, evapotranspiration, and in situ water level. Their results showed a more steady 50% overall decrease in area extent for the Great Salt Lake between 1999 and 2021, whereas for Lake chad, their results indicate ~30% overall decrease with larger fluctuations in area extent for the same time period. Overall this paper is well structured, the writing of this paper is easy to follow. There are certain aspects/details that require further clarification from the authors, and I have listed them in my comments below. 

 

More information on the parameters of the random forest supervised classification is needed. For example, what is the total number of decision trees?  What is the maximum depth of the trees? Tuning these different parameters may greatly impact the accuracy of your classification results, and it would be good to justify its impact, and also provide rationale on why these parameters were selected in constructing your random forest.

 

In selecting those Landsat imagery, even though the annual image processing were done using all images intersecting the target area, I think you need to be cautious and keep track of what time of the year (or what season) did most of the images come from. Since seasonal variability may come into play and convolve with the inter-annual variability that you see in the lake surface area time series. I think it would be good to justify the possible impact of the seasonal variability when discussing the causalities. 

 

When you discuss the influence of climate variability on the observed lake dynamics, it would be nice to also investigate the relative impact of the variables that were examined (namely annual precipitation, evapotranspiration, and water depth), perhaps implement a multi-linear regression and examine the R-squared or the explained variance of each variable.

 

Minor:

Line 45 and Line 47: remove “-“ in the words.

Line 101 to 103: I would be more careful with the use of the word “fusion”, since in this study you are mostly just using data products from different sources trying to explain the variability of the lake dynamics time series. Strictly speaking, you are not fusing them together, but just using them separately.

Line 141 to Line 154: The information listed here is a bit all over the place. I think it can be cleaned up quite a bit by simply listing which Landsat datasets were used during which time period, and then list the reasoning, other than listing the time periods of the availability of each dataset…

Line 157: 50% cloud filter was mentioned for the first time but not explained, only to find explanation on Line 164. Consider bring the 50% cloud cover filtering to Line 157.

Line 181: Incorrect sectional title, it should be MODIS Dataset or MODIS Evapotranspiration dataset.

Line 201-202: You can probably put this sentence in the data availability section at the end of the paper. 

Line 222: Would be helpful to also list the section number for the random forest subsection.

Line 223: Repeated Section title with 2.3.1, I think you meant “2.3.2 Random Forest classification”.

Line 244: the description on accuracy metrics in this entire sub-section is somewhat confusing. I think to improve the clarity of this section, you should be clear on how these metrics were calculated (e.g., which is the reference image and which is the target image), and possibly step through an example of a classification case. Simply read through the [28] reference that you cited and see how they explained each metric/term.

Line 247: “producer accuracy” instead of “product accuracy” here?

Line 263: What is “Tri“ in equation 2?

Line 272: The fontsize of this equation is smaller than the previous ones. Please try to be consistent.

Line 334 to 336: In the table captions TA and KTA are referred to as test sample accuracy and kappa test accuracy, yet in the main text Line 338 to 339 they are referred to as training accuracy and kappa’s training accuracy. Please be consistent.

Line 455 to 463: I am totally confused by what you want to convey here. Please consider rephrase this paragraph.

Line 485: “Pham-[19]” reference not listed.

Line 503 to 504: I thought you just said “that the rapid expansion of the lake's surface area was consistent with high annual precipitation…” on Line 496. 

Line 534 to 535: Please improve the grammar of this sentence.

Author Response

We, the authors, are thankful to the reviewers for their very constructive and fruitful comments. which were focused primarily on the methodology part and improving the results section. Recommendations supplied in the review report were carefully followed; the text was revised accordingly. 
The comments were simply copied from the referee report and pasted below after which our responses are given in red. Changes made in the manuscript are given in red.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Thanks for addressing my comments.

Author Response

The authors would like to thank the reviewer for his fruitful comments and suggestions.

Reviewer 2 Report

The authors have addressed all my questions and concerns. I believe the manuscript has been sufficiently improved to warrant publication in Hydrology after fixing the following minor issues.

Line 285: "won't" to "will not"

I think the reference section needs some reformatting. Please follow MDPI's recommended reference list and citation styles https://www.mdpi.com/authors/references. Please note that each reference should start with Author last name followed by first name initials. Also, for cited online resource, for example, reference number [16], [32], and [40], you should probably also provide the URL to the resource.  I also believe that reference [40] you should be citing the user guide: https://lpdaac.usgs.gov/documents/494/MOD16_User_Guide_V6.pdf.

Author Response

The authors would like to thank the reviewer for his fruitful comments and suggestions.

The reviewer's concerns were addressed as follows:

• English language and style were checked.
• Line 285-286: "won't" was changed to "will not"
• The reference style was revised according to the journal guideline
• Online resources for references [16], [32], and [40] were added.