Urban Flood Prediction Using Deep Neural Network with Data Augmentation

Round 1

Reviewer 1 Report

The paper combined the deep neural network (DNN) and SWMM model to predict the total accumulative overflow of urban areas, especially in data insufficient consequence. Although the authors provide some exciting results, there remains several points need to be further clarified, which can be summarized as follows:

1. As a paper in machine learning field, what we care about most is the data preparation. However, the data preparation of DNN scattered in different sections (from 3.3 to 3.5). I strongly suggest the author modified the sec 3.4, and provide enough information for readers.

• Line 244-245: “70% of input data was used for training, 20% was used for validation, and 10% was used for testing.’ How many rainfall events for each datasets? How to select them? I suggest the authors add a table to describe them.
• The authors use a DNN with complex structure. How to determine the structure? Please describe.
• How to apply data augmentation? Why are 10 rounds taken in the research? Please describe.
• A comparison results table in training, testing and validation datasets is needed.

2. The SWMM simulation results need further proof. The difference of Fig.5(a) and Fig.5(b) can't be ignored. Considering the SWMM simulation results is set as DNN's target, the model errors will highly affect the results. I suggest:

• add more storms to validate the SWMM model.
• Estimate the total cumulative overflow of each storm and compare it with simulation. NOT just the flood mark.

3. The data augmentation method should be explained to the readers in detail. It is the emphasis of the paper, but described too brief. Too many efforts are engaged in describing how to apply noise, but the strongpoint of the method lack analysis. I suggest

• Add the description of the reason why add noise can improve the model efficiency.
• Improve the flowchart in Fig.2. Especially, how to execute outlier check, the criteria?

Author Response

Thank you for reviewing our paper, we attached the file for response.

Author Response File: Author Response.docx

Reviewer 2 Report

Dear authors, 

Please, read the attached file concerning  some questions and suggestions for improvement.

Your reviewer

Comments for author File: Comments.docx

Author Response

Thank you for reviewing our paper. We attached the file for response.

Author Response File: Author Response.docx

Reviewer 3 Report

General comments:

1. The paper is of research interest and sheds light on the way of using DNN and data augmentation for supporting and improving flood forecasting in urban environments. However, some aspects of the manuscript should be improved, as specified in the following comments.
2. In the Introduction, a deeper review of the state of the art on application of Neural networks and urban floods is requested. For this purpose, I suggest to see Talei et al., 2010 and Abou Rjeily et al., 2017. Instead of only citing the scientific literature and then presenting their work, the authors should underline the research gap they are proposing to fill with their research as respect to the current scientific literature.
3. As limitation of the work, in the Discussion, the authors should mention the fact that they are considering the SWMM results as “true”, but this kind of 1D model simplifies complex phenomena and can be affected by strong limitations (Mark et al., 2006).
4. It can be interesting to see if there is a relation between the performance of the DNN and the total rainfall (or rainfall intensity), showing the results for specific rainfall events.
5. In the Discussion, the authors should underline the fact that for a proper flood forecasting, decision makers are interested not only on the total amount of overflow, but also on WHERE the overflow occur in order to identify the flood risk areas. Since DNN approaches are not physically based, the author should mention, as potential future research, the need of combining the total amount of overflow with the delineation and quantification of flood risk areas in a spatially distributed way.

Specific comments are reported below:

1. Page 1, Abstract, Line 5. Remove the comma after “(deep learning)”
2. Page 2, Table 1, Second heading: Replace “Observed Date” with “Days of observation”
3. Page 4, Figure 1 b) Place an x as name of the horizontal axis.
4. Page 4, Figure 3. A frame or inset with a zoom out (at Country/city scale) of the study area should be added.
5. Page 7, Section 3.2. Since FLO-2D is able to couple 1D SWMM storm drain and the 2D surface runoff, did the authors took advantage of this module or they run separately the SWMM and the FLO-2D model? This aspect should be clarified in the mentioned section 3.2. Moreover, the title of the Section seems not appropriate. The authors did not perform a real calibration. The parameters of the model were not changed to test the best ones as respect to the Flood Trace Mark. In this case, the term validation is maybe more appropriate.
6. Page 10, Lines 241-244. The authors should better explain of they set the number of hidden lyers and their related nodes. What was the criteria of the trial and error? How the authors set the epoch of maximum learning? Did they find a sort of stationarity of the MAE/ MAPE?
7. Page 24, Table 6. To which simulation is referred the Average of Absolute R-square? Please enrich the caption of the table.
8. Page 15, Line 297. Replace “could be performed” with “was performed”.
9. Page 15, Line 325. Replace ”2–3 s” with “2–3 seconds”

References

Abou Rjeily, Y., Abbas, O., Sadek, M., Shahrour, I., & Hage Chehade, F. (2017). Flood forecasting within urban drainage systems using NARX neural network. Water Science and Technology, 76(9), 2401-2412.

Mark, O., Weesakul, S., Apirumanekul, C., Aroonnet, S. B., & Djordjević, S. (2004). Potential and limitations of 1D modelling of urban flooding. Journal of Hydrology, 299(3-4), 284-299.

Talei, A., Chua, L. H. C., & Quek, C. (2010). A novel application of a neuro-fuzzy computational technique in event-based rainfall–runoff modeling. Expert Systems with Applications, 37(12), 7456-7468.

Author Response

Thank you reviewing our paper. We attached the file for response. We are impressed with your academic review.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

I have readed the revised paper carefully, and found that the issues that I once concerned have been modified accordingly. To my point of view, the paper has reached the requirments of publication. Thanks for the efforts of the authors.

Author Response

Thanks you for review. Please see the attachment. 

Author Response File: Author Response.docx

Reviewer 2 Report

I agree with almost all explanations and improvements provided by you. However, is difficult for me to understand why “However, this study applied 0.014 as roughness of urban area by referring to the road design manual (KICT, 2001).”

Please, clarify the following:

• The water is flowing on the whole floodplain? In this case, the roughness coefficient should be much larger than 0.025 (to take into account the hydraulic resistance of the buildings). May be 0.080 or even 0.100 are the adequate values.
• The buildings’ areas were blocked for water flow? In this case, the water is flowing among the buildings on the streets and green areas, and I can accept a synthetic roughness coefficient of 0.025 (a weighted average for different land uses). According to your text, a value of 0.014 was applied as roughness coefficient of urban area, which I totally disagree.
• There is also another contradiction between the roughness coefficient of the roads: 0.014 (according to the road design manual) and 0.047 (row 227: while ?1, ?2, and ?3 are 0.06 (farmland), 0.047 (road), and 0.05 (others), respectively).

Author Response

Thanks you for review. Please see the attachment. 

Author Response File: Author Response.docx