Drainage Pattern Recognition of River Network Based on Graph Convolutional Neural Network

Round 1

Reviewer 1 Report

This paper describe a drainage pattern recognition method based on deep graph convolutional network architecture. By focusing on four drainage patterns the authors show that the proposed approach gave interesting and promising recogition results. Several concern is expressed in my following comments and recommendations are suggested.

An important concern is that the authors need to explain why they focus on the four specific drainage patterns? how are they important than other patterns and for what purposes? Such motivations are necessary.

Format issues: white spaces are needed before the citation brackets.

Experiments could be improved, see my specific comments.

Specific comments:

Sect 3.1. It is not necessary to describe the mathematical principles behind GCN. It is a standard component in current deep learning applications. This section should be greatly shortened to improve readability of the paper. 

P5, L209-210: why did the authors choose to use the kernel size of 1 in their attention pooling layer? what about using kernel size larger than 1? I recommend the authors to do some comparason evaluations to reveal more insights. 

In general, the idea behind using SAGPooling layer is not explained: based on what criterion the ranks/weights of nodes in a graph are computed? or, what is it that makes a node in a graph more important than the other nodes? so that appropriate computational measures are chosen to capture that rank. This point should be briefly discussed here.

Sect 4.4 title: It is a bit confusing to see network structure here, as here are two networks: one is the river network (physical), the other is the graph convolutional network (model). Please revise the section title for clarification.

Sect 4.3.2: it is not crystally clear how the relative poistion (left or right) is encoded into the river segments. To do so, you need to determine the main stream in the first place. How is the main stream identified for a subbasin? what is the coding of the main stream segments themselves? are they left or right? it is enough to use just binary codes (0 and 1)?

P8, L330: a reference is needed for Strahler's stream coding system.

Sect 5: An important issue seems to be omitted: small scale sub-networks in river systems are isolated and used directly in the experiments. In fact, real-world river networks are not easy to be identified and segmented from each other; they usually are woven into larger and larger networks. Perfact sub-basin that contain a single pattern usually does not exist: many are much more noisy, mixing more than one prototypical drainage patterns, some may lie between the boundary of two or more sub-basins. How does the model perform for such data? The author should include a discussion of this point (could be interesting for potential readers) and list it as a limitation for future work.

Fig 11: what is the only one descriptive index? please be more specific? the one that is under each column? all indices: how many? e.g. 4 remaining features?

P15, L515: "only some river network models are considered when selecting samples" please clarify what do you mean by this? models seem to refer to deep learning models in this paper. better to use a different term and clarify the intended meaning.

Author Response

We are grateful for the helpful and insightful comments from the reviewer. Many changes were made improving the manuscript based on these remarks. Our response and changes can be found in the attached file.

Author Response File: Author Response.docx

Reviewer 2 Report

Summary:

In this paper, the authors apply a graph convolutional Neural Network to classify river characteristics of 2000 rivers in China. The model is innovative and could be useful to other scholars and government agencies who may need this information and automated process. Particularly for planning and floodplain mapping (however this is not mentioned in the paper). The methods and results sound valuable – but the writing needs to be improved. Please see my comments below.

General comments:

Focusing on categorizing river structure between sinuosity and other characteristics is interesting, valuable and innovative.

Great points are made throughout the paper – for example line 113 -115 about subjective decisions. However – subjective decisions also go into the automated models. This point was not mentioned in the paper. How will your method be more uniform?

What about spatial scale? Will that influence categorization? How would that influence the model and the output?

Specific comments:

Writing could be smoother. For example:

There are many word repetitions. For example, the phrase “based on” is at least three times in the abstract alone.

Some word choices make the reader try to guess what the authors are saying. Especially in the abstract. “…realize the river network data structure.” Is this paper going to be about data structures?

Segues between ideas and different sections are needed.

Some places ideas are repeated several times, too many places to mention. Sometimes they are repeated ideas with many words. Other places it is literally repeated words – such as in line 516.  In some places it feels like repetition and the English is unclear (for example: line 91-94 on page 3- this is a very unclear set of sentences – please re-read and re-focus). Many sections the English needs to be improved significantly.

The tense of the writing changes from first person (“we try to fine” p 15 line 504 to third person “this paper”) pick one and stick with it. My personal preference is first person.

More citations to support claims are also needed. For example – you make many claims about cartography without any references. Another section needing references relate to your description of CNN and GCN – examples of how they have been used in the past are needed.

In the abstract – first state the specific problem that this research addresses right away. This is not clear to the reader.

What do you mean by “river as the basic unit” Define river as a unit. Rivers are long – how long is the unit? What connections count in this unit?

“we try to fine the most suitable feature combination…” line 503-506 – this is important and should have been stated earlier in addition to here. Same for the statement “it should be mentioned that the purpose…” line 384 – 386 State the purpose on page 1.

Figure 1 and 8 are almost the same – how do they differ? Formatting – figure figure 8

Methods- GDEM-V2 – tell the reader what kind of data this is – true color? Digital elevation model? Multispectral data?

Table 2 is very basic. Is it needed? In the title shouldn’t you state these are the number of samples in the training set specifically?

Did you segment the 2000 river networks?

Section 4.3.1 could be in the background section – state clearly how this type of information is used – why it is important to do this classification – how it could help a specific problem…

Suggest making a specific results section – I had to search for them.

Line 455 which features are critical? Which not important? Do you know? Do you have an idea? Tell the reader. This is important. This paragraph is the most important part of the paper but is still vague. I would not be able to improve this study if I were to build upon it with this vague information.

In the results and conclusions – I would love to see at least a screenshot of the output not only the confusion matrix. In the conclusion – this is very surficial and nothing new or amazing is shared. This research is novel – tell the reader why. What did you find? The last sentence hints at future research – be more specific. Don’t make the reader guess what you mean or what would be a good idea. Tell them.

Author Response

We are grateful for the helpful and insightful comments from the reviewer. Many changes were made improving the manuscript based on these remarks. Our response and changes can be found in the attached file.

Author Response File: Author Response.docx

Reviewer 3 Report

Dear all,

This work deals with the recognition of four drainage patterns through deep learning. This is a very interesting topic and the developed approach is promising for this application, however there are significant issues that need to be addressed prior to publication. Kindly find attached my detailed comments. 

Comments for author File: Comments.pdf

Author Response

We are grateful for the helpful and insightful comments from the reviewer. Many changes were made improving the manuscript based on these remarks. Our response and changes can be found in the attached file.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Authors have addressed all my concerns and I have no further questions.

Author Response

Dear Reviewer,

We would like to express our heartfelt gratitude for your meticulous review of our manuscript. Your insightful comments and thoughtful suggestions have been invaluable to us. We deeply appreciate the time and effort you have dedicated to providing such comprehensive feedback.

Thank you once again for your invaluable contribution to our work.

Warm regards.

Reviewer 2 Report

Thank you for making my recommended changes - i still think river unit should be better, more precisely defined earlier in the paper. River network is still vague. When does the binary tree stop being followed? When does it end? What is the geographic threshold? 

Author Response

Dear Reviewer,

We thank the reviewer for providing these helpful and insightful comments. Multiple improvements were made throughout this manuscript. Many were textual additions providing clarity. Our response and changes can be found in the attached file.

Thank you once again for your invaluable contribution to our work.

Warm regards.

Author Response File: Author Response.docx

Reviewer 3 Report

The authors addressed my concerns satisfactorily, thus I recommend the publication of their work in it's present form.

Author Response

Dear Reviewer,

We would like to express our heartfelt gratitude for your meticulous review of our manuscript. Your insightful comments and thoughtful suggestions have been invaluable to us. We deeply appreciate the time and effort you have dedicated to providing such comprehensive feedback.

Thank you once again for your invaluable contribution to our work.

Warm regards.