A Review of Hydrodynamic and Machine Learning Approaches for Flood Inundation Modeling
, Lachlan Tychsen-Smith 2 and Lars Petersson 2,3
Round 1
Reviewer 1 Report
This paper mainly gives a brief summary of ML and DL models for flood inundation modelling. It is indeed helpful for readers in flood modelling to have such an overall review. However, after reading through this paper, there are quite a few issues to be answered in this paper; particularly, critical comments are needed, and the prospect of modelling for the future is lacking. In addition, I also have the following comments:
1). The title is a bit over vague: suggest to replace conventional and data-driven approaches by machine and deep learning appoaches
2) L225-245, The topic claimed is 'Other ML related models used for flood inundation, but the cited references are not those for the implementation of the relevant models on flood innudation.
3) Before Section 3, need to state clearly what are the issues for ML, DL or conventional approaches for flood inundation modeling?
4) In Table 2, the author list would be better to list in a certain order, e.g. in year or model/method; also need a concluding comment or remark on each type of modelling
5) The same issue on the list in Table 3 rises as that in Table 2 (see the comment above).
6) It seems there are some overlaps on some models in Table 2/3
7) Section 3.3 (Datasets), despite data described in the text and summarized in Table 4 for the models in Table 2/3, it is suggested to have a concluding comment or remark at the end of section
More specifically (note: the authors need carefully check typos or grammatical errors, here only give a few lists)
L6, Do you mean physical based methods, rather than physical methods?
L7, models...suffer, not model...suffer
L70, Modelling, not modelling
L78, contraint, not constrains
L97, delete central, which is not needed as necessary here
L105, Why do 14 rows (not 12 rows or another number) represent 12 hour periods?
L117, Why W is a 6x5 element matrix? Is it 16, not 6 here?
L151, take a value 1, 2 or 3, not values 1, 2 or 3
L166, what is the D?
L183, What is the short PCA?
In the title of Figure 2. Schematic diagram, not schematic; the same applies to Line 220
L222-223, please rewrite the sentence (grammatical error)
L252, Figure 3, not Figure.3
2 lines above eq (1), delete a
L287, CNN is not defined before
L314, What is LSTMs?
L323, is presented ..., not are presented ...
L263, One of ..., not one of ...
L450, Hosseiny [64] without et al.
Author Response
1). The title is a bit over vague: suggest to replace conventional and data-driven approaches by machine and deep learning approaches
Response: As per suggestion, we have revised the title as ‘A review over hydrodynamic and machine learning approaches for flood inundation modeling’.
2) L225-245, The topic claimed is 'Other ML related models used for flood inundation, but the cited references are not those for the implementation of the relevant models on flood inundation.
Response: To avoid confusion we have re-phrased the section as ‘Other ML related models’, ‘used for flood inundation’ is deleted. In this section we have summarised common ML models. Cited models are applicable to flood data analysis.
3) Before Section 3, need to state clearly what are the issues for ML, DL or conventional approaches for flood inundation modeling?
Response: We have provided detail of ML and DL methods in the Section 2.2. Issues for ML and DL approaches for flood inundation modelling are described in Section 4.2.
4) In Table 2, the author list would be better to list in a certain order, e.g. in year or model/method; also need a concluding comment or remark on each type of modelling
Response: Re-arranged based on application area & year of publication.
5) The same issue on the list in Table 3 rises as that in Table 2 (see the comment above).
Response: Re-arranged based on application area & year of publication.
6) It seems there are some overlaps on some models in Table 2/3
Response: Yes, some methods (e.g. RF, MLP) are used in both traditional ML and DL methods.
7) Section 3.3 (Datasets), despite data described in the text and summarized in Table 4 for the models in Table 2/3, it is suggested to have a concluding comment or remark at the end of section
Response: In this section we have described the data sets that were used in different studies. The importance of open-source data and availability of different data sets are provided in Section 4.2.2.
Comment: L6, Do you mean physical based methods, rather than physical methods?
Response: Right terminology is ‘physically based models’. Corrected
Comment: L7, models...suffer, not model...suffer
Response: Corrected
Comment: L70, Modelling, not modelling
Response: Corrected
Comment: L78, constraint, not constrains
Response: Corrected
Comment: L97, delete central, which is not needed as necessary here
Response: deleted
Comment: L105, Why do 14 rows (not 12 rows or another number) represent 12 hour periods?
Response: This is an example for 1 week period. At 12-hour interval there are 2 data in a day and 14 data in a week. For 6 sensors there are 14×6 data.
Comment: L117, Why W is a 6x5 element matrix? Is it 16, not 6 here?
Response:
Equation is an example of mapping between the input data X (16x5) and output data Y (14x6) by the assumption that “each row in Y can be recovered by applying a fixed but unknown linear function to the corresponding row in X and adding some unobserved noise.” And thus the 6x5 is correct as it refers to the number of columns not rows of X, Y. The number 6 is also correct as it refers to the number of columns of Y.
Comment: L151, take a value 1, 2 or 3, not values 1, 2 or 3
Response: Corrected
Comment: L166, what is the D?
Response: D is the dimensionality of the new space, which is explained in line 170: “The dimensionality D of the space to which f maps may be very large, or even infinite.
Comment: L183, What is the short PCA?
Response: PCA stands for ‘principal component analysis’
Comment: In the title of Figure 2. Schematic diagram, not schematic; the same applies to Line 220
Response: Corrected
Comment: L222-223, please rewrite the sentence (grammatical error)
Response: Revised as ‘Boosted regression tree (BRT) combines a boosting algorithm with regression tree to increase the performance and reduce the model variance’.
Comment: L252, Figure 3, not Figure.3
Response: Corrected
Comment: 2 lines above eq (1), delete a
Response: deleted
Comment: L287, CNN is not defined before
Response: CNN stands for ‘Convolutional neural network’. Defined in the previous paragraph.
Comment: L314, What is LSTMs?
Response: LSTM stands for ‘Long short-term memory’. Defined in Line 314 (Line 360 in revised manuscript)
Comment: L323, is presented ..., not are presented ...
Response: Corrected
Comment: L363, One of ..., not one of ...
Response: Corrected
Comment: L450, Hosseiny [64] without et al.
Response: I think ‘Hosseiny et al. [64]’ is the right citation style. However, I will cross-check with the MDPI’s publication team.
Reviewer 2 Report
This paper reviewers the machine learning and deep learning methods to predicate flood inundations. Not only weakness and strength of these data driven methods are introduced, but also challenges and future research avenues have been discussed. It is a well review article for flood inundation modeling.
Author Response
This paper reviewers the machine learning and deep learning methods to predicate flood inundations. Not only weakness and strength of these data driven methods are introduced, but also challenges and future research avenues have been discussed. It is a well review article for flood inundation modeling.
Response: Thanks for your valuable comments and positive feedbacks. We have revised the manuscript further to improve readability.
Reviewer 3 Report
In my opinion the paper could be useful for people who want to have a first idea of the topic about data-driven approaches for flood inundation modeling. In general the paper is well structured, but session "5. Conclusion" is more similar to a very short summary. I suggest two different possibilities: a) to change session 4 in Discussion and Conclusion and to remove session 5. Conclusion or b) to identify the bullet points of session 4. and to rewrite the session 5 conclusion as a short synthesis of final considerations about the topic. Probably the last one would be better.
Author Response
In my opinion the paper could be useful for people who want to have a first idea of the topic about data-driven approaches for flood inundation modeling. In general the paper is well structured, but session "5. Conclusion" is more similar to a very short summary.
I suggest two different possibilities: a) to change session 4 in Discussion and Conclusion and to remove session 5. Conclusion or b) to identify the bullet points of session 4. and to rewrite the session 5 conclusion as a short synthesis of final considerations about the topic. Probably the last one would be better.
Response: As suggested we have revised the section 5 with additional concluding remarks
Reviewer 4 Report
The authors present a review and comparison of machine learning (ML) and deep learning (DL) methods/approaches to flood inundation modeling by reviewing recent review papers and then studies that have used these approaches. The strength of this paper is the number of studies reviewed and the depth of detail provided. Additionally, Section 4 is a very useful section in outlining some of the challenges in this field. I think this could be a really useful, highly cited paper, and Water is exactly the place to publish this, but before this is published I think revisions are needed to strengthen the paper and its contributions to the field.
The strengths of this paper could be boosted in several ways:
- Clarifying and standardize terminology – something that is clearly problematic with this field (e.g., ANN and MLP being used interchangeable across studies, and even ML vs DL). By providing readers with a central and standardized (or at least more consistent) list of terminology and definitions, this paper will become the ‘go-to’ in this field for referring to these various tools
o While I understand that a comprehensive introduction of machine learning is not the aim of the paper (Line 99) – clarification and consistent use of various terms is needed in order to effectively communicate the authors’ ideas.
o Perhaps a very brief intro (1-2 lines) of how ML and DL are different – at least in the context presented here (something beyond this was how the terms were used in primary papers reviewed (e.g, Line 141-143)). This distinction seems to be important to establishing the knowledge gap that this review paper aims to address, but by not defining these clearly, this distinction is lost.
o In the opening line of the introduction, it says DL is a type of ML, but then goes on to treat these as two separate methods – maybe it is as subtle as referring to ML as TML (traditional machine learning) and DL as DML (deep machine learning)? Addressing this would then allow the authors to replace ‘data-driven’ with Machine Learning throughout the paper (and address a further concern that the authors are implying that physical models are not data-driven).
§ Authors suggest that physical models are not data driven (e.g., Line 1, Line 6-7 in the abstract and throughout the paper). I don’t think this is on purpose, perhaps this is a word choice issue – and I suggest careful consideration here (and throughout the paper) is needed so as to avoid this insinuation.
- The authors state that one of the aims for this paper is to make the case for increasing user-expertise in ‘validation’ steps to consider the model outputs – I agree this is important, however I wasn’t convinced by the argument presented. Clarifying/strengthening this aspect of the paper will improve the paper’s utility to a wider audience of readers. One way this might be accomplished is by presenting a synthesis of the examples in Section 3, rather than a summary of each study. By providing the readers with the key pieces synthesized, I believe readers could use this paper as a way-finder for particular methods or tools in their own work and help them make decisions about what approach to take. In its current presentation, Section 3 is more along the lines of an annotated bibliography.
- The limited number of references in section 2.2 (and its subsections) is concerning. Including additional references in each paragraph, for example similar to how this is done in Line 101, Line 181-184, Line 302 or Line 307, would be most helpful to the reader looking for more information. And again I think this would set up this review as the ‘go-to’ for users.
- Section 4 should be last section – Section 5 doesn’t add anything to the paper and Section 4 is well-organized, and reads as a conclusion anyway.
- Perhaps including a list of abbreviations would improve the readability.
Specific Comments
Russell Tsuchida’s email is included in author/affiliation info, but is not listed as an author.
Throughout the paper there are typos and grammatical inconsistencies, these should be addressed
MI vs ML (line 6 – might be a typo?)
Author Response
The strengths of this paper could be boosted in several ways:
- Clarifying and standardize terminology – something that is clearly problematic with this field (e.g., ANN and MLP being used interchangeable across studies, and even ML vs DL). By providing readers with a central and standardized (or at least more consistent) list of terminology and definitions, this paper will become the ‘go-to’ in this field for referring to these various tools
o While I understand that a comprehensive introduction of machine learning is not the aim of the paper (Line 99) – clarification and consistent use of various terms is needed in order to effectively communicate the authors’ ideas.
Response: We have revised the ‘Introduction’ section and made our best efforts to increase the readability and flow of information.
Comment: Perhaps a very brief intro (1-2 lines) of how ML and DL are different – at least in the context presented here (something beyond this was how the terms were used in primary papers reviewed (e.g, Line 141-143)). This distinction seems to be important to establishing the knowledge gap that this review paper aims to address, but by not defining these clearly, this distinction is lost.
Response: Thanks for this valuable comments. Yes, I agree that ML and DL are used interchangeably, but it is important to highlight the difference between ML and DL. We have added following information. ‘It is important to note that both ML and DL are part of artificial intelligence (AI). The ML algorithms learn from structured data to identify patterns in that data and DL algorithms are based on highly complex neural networks that detect patterns in large unstructured data sets’.
Comment: In the opening line of the introduction, it says DL is a type of ML, but then goes on to treat these as two separate methods – maybe it is as subtle as referring to ML as TML (traditional machine learning) and DL as DML (deep machine learning)? Addressing this would then allow the authors to replace ‘data-driven’ with Machine Learning throughout the paper (and address a further concern that the authors are implying that physical models are not data-driven).
Response: Thanks for the suggestion. Yes, DL is also a machine learning that uses advanced technique such as Convolution Neural Network (CNN). To avoid confusion, we have now replaced Machine Learning with ‘Traditional machine learning’ and kept ‘Deep learning’ as is.
Comment: Authors suggest that physical models are not data driven (e.g., Line 1, Line 6-7 in the abstract and throughout the paper). I don’t think this is on purpose, perhaps this is a word choice issue – and I suggest careful consideration here (and throughout the paper) is needed so as to avoid this insinuation.
Response: Physical models also need data to configure and calibrate the model parameters. To avoid confusion data-based models (e.g. ML/DL), we replaced the ‘data-driven’ with ML/DL as appropriate.
Comment: The authors state that one of the aims for this paper is to make the case for increasing user-expertise in ‘validation’ steps to consider the model outputs – I agree this is important, however I wasn’t convinced by the argument presented. Clarifying/strengthening this aspect of the paper will improve the paper’s utility to a wider audience of readers. One way this might be accomplished is by presenting a synthesis of the examples in Section 3, rather than a summary of each study. By providing the readers with the key pieces synthesized, I believe readers could use this paper as a way-finder for particular methods or tools in their own work and help them make decisions about what approach to take. In its current presentation, Section 3 is more along the lines of an annotated bibliography.
Response:
Comment: The limited number of references in section 2.2 (and its subsections) is concerning. Including additional references in each paragraph, for example similar to how this is done in Line 101, Line 181-184, Line 302 or Line 307, would be most helpful to the reader looking for more information. And again I think this would set up this review as the ‘go-to’ for users.
Response: We focused mostly on recent publications on ML application for inundation modelling.
Comment: Section 4 should be last section – Section 5 doesn’t add anything to the paper and Section 4 is well-organized and reads as a conclusion anyway.
Response: I fully agree with the reviewer. However, we prefer to keep Section 5 as a short summary of key findings (as the Section 4 is very long).
Comment: Perhaps including a list of abbreviations would improve the readability.
Response: Yes, agree. A list of abbreviations is added at the end of the paper.
Specific Comments
Comment: Russell Tsuchida’s email is included in author/affiliation info, but is not listed as an author.
Response: Deleted Russell Tsuchida’s email
Comment: Throughout the paper there are typos and grammatical inconsistencies, these should be addressed
Response: Carefully read the paper and corrected typos and grammatical inconsistencies.
Comment: MI vs ML (line 6 – might be a typo?)
Response: Yes, corrected to ML.
Round 2
Reviewer 1 Report
The authors have made some changes accordingly after the reviewers' comments. Despite no detailed response letter to the reviewers' comments, I can see the track of change which responds to most of my comments. But I still could not see any reply to this comment as follows:
7) Section 3.3 (Datasets), despite data described in the text and summarized in Table 4 for the models in Table 2/3, it is suggested to have a concluding comment or remark at the end of section
Author Response
Comments: Section 3.3 (Datasets), despite data described in the text and summarized in Table 4 for the models in Table 2/3, it is suggested to have a concluding comment or remark at the end of section
Response: we have added the following paragraph at the end of Section of 3.3: “A summary of all datasets is presented in Table 4. It demonstrates a wide variability in the source of data and software that were used to generate input data. All studies emphasized the importance of large data sets for training and validating ML models. Majority of these studies used hydrodynamic model-generated flood data for training ML/DL algorithms because measured and/or remotely sensed flood inundation data were insufficient for training/validating such models. Lack of publicly available benchmark data limits the rigorous testing of ML and DL technologies in the field of flood inundation. These challenges are also mentioned in other studies (e.g. Sit et al. [2]).”me concluding remarks are now added at the end of section 3.3:
