Improving the Accuracy of Hydrodynamic Simulations in Data Scarce Environments Using Bayesian Model Averaging: A Case Study of the Inner Niger Delta, Mali, West Africa

Round 1

Reviewer 1 Report

This work describes the calibration of several model domains some with different forcing conditions, though the reasoning for the different forcing conditions is lacking.  And then describes the application of a Bayesian Model averaging to better calibrate the model domains used.  The purpose of this calibration, model comparison and Bayesian model averaging is to build a 2D hydrodynamic model to simulate variables such as water level, discharge, inundation extent.  While a decent introduction describing why understanding the hydrology of the inner Niger Delta is given, and as stated on L61-62 that accurate simulations of inundation could help our understanding, no link was provided regarding how the model domains built will be applied outside this model calibration exercise.  Essentially this work amounts to a model calibration exercise, and more disappointing is the lack of insight, beyond which models calibrated better, was proved.  Furthermore, it is plainly obvious that a Bayesian approach that weights model performance will provide a better calibration.  But what understanding does this provide?  How will it be applied to the problems of the Niger Delta?  Beyond fitting the models to an observed curve, what can be said about this work? When done right calibration and model comparison can provide new information about what processes are important for a given site, and why some models perform differently than others.  Unfortunately, this manuscript provides no insight into the problem and only states that the model domains perform differently, and that Bayesian approaches, which is essentially a form of cure fitting, will calibrate better.  The usefulness of this information is not provided.  Given that I do not find scientific value of the exercise, I cannot recommend this manuscript for publication.      

It should also be noted that no reasoning was given to why some parameters were calibrated, and others that govern water balance, i.e., evapotranspiration and rainfall were not, or were neglected altogether within the domain of interest.  Given this lack of process representation, and specifically the lack of reasoning as to why these processes were neglected, I have serious doubts about model application outside of the calibration exercise. 

Author affiliation for Sara Karam was not given

Figure 1.  Map should be shown with in larger the West African region to better familiarize readers of the location.

L75-59:  it would be good in include citations for this sentence.

L70: Replace ‘got’ in ‘…but model results got affected due to …’ with were.

L71:  The kinematic and diffusion wave approximation equations are very different from each other that respond to 0 slope in very different ways, i.e. the diffusion wave equation still should be valid. 

L72: what is ‘rive’?

L75: Replace ‘in’ in ‘…finite-volume method in unstructured grid…’ with ‘on’

L79:  provide citation or link for TELEMAC 2D model.

L222-228:  Lacks sufficient description of the SWAT model.  Is it just a surface energy and water balance model?  How is discharge from SWAT calculated?

L294-298:  The description of boundary conditions (i.e., the inlet and outlet forcing conditons), does not fit with ‘Mesh generation’.  This should be in it’s own subsection and needs to be described in better detail. 

L297: Why was rainfall not considered?  This could be a major pitfall of the modeling effort.

L298:  Justification for considering a constant evaporation flux needs to be provided.  Again is a possible pitfall of the modeling done here.

L306: Omit ‘keep’ in ‘…floodplain was keep constant during …’

L315:  cloud should be plural ‘clouds’. Same and L318.

Section 2.7.  No justification was given as to why mannings coefficient was calibrated, but not other uncertain parameters, chiefly water balance forcing’s of evapotranspiration and rainfall.   

L410: ‘Results show that downstream boundary condition[s] affects model performance.’  Is there a table or figure that show’s this? Please direct the reader there.  If there Is not a table for figure, then there should be.

Conclusions section:  These conclusions provide little to no insight into the why the models behave the way they do, but rather only offer a description of model performance.  I’m not at all certain how this is useful.

Author Response

Reply submitted as an attached file

Author Response File: Author Response.docx

Reviewer 2 Report

An interesting paper.  As I understand it, the problem is to predict the annual flooding of a remote location, which lacks accurate map measurements.  This paper solves the problem by taking all 6 combinations of the 3 data sets for elevation and 2 data sets for downstream boundary conditions (and some reasonable assumptions, given the lack of reliable data), and then combining their various predictions using Bayesian model averaging.  The climax is Table 7 which validates the approach, in that the combination is better than any single model for half the columns, and only slightly worse in the other half.  The approach has more general significance as a way to reduce model risk where good maps are unavailable, and also a specific significance for predicting the future in this particular part of the world.

My comments are mostly about presentation and minor glitches.

Page 3, line 72, should "rive" be "river"?  A spell-checker might not pick this up, because "rive" is an valid word (even though it's a bit old fashioned).

Page 3, line 79, should "suit" be "suite", as in the sense of a collection of bits of software?  That sentence is not quite clear.

Page 10, line 304, should "bellow" be "below"?  Again, a spell-checker might not pick this up, because "bellow" is also a valid word.

Table 4 has 6 rows, but in my copy the last row is at the top of the following page, i.e., an orphan.  Is it possible to keep the whole table on the same page, perhaps by moving it to the top of the next page? 

On page 11, line 341 has "manning's value" with a lower-case "m".  Should that be upper-case, as in "Manning's value" or perhaps "Manning's coefficient" to be consistent with earlier in the paper?

It might be that I don't have the correct fonts installed, but on my computer the four graphs in Figure 8 and the similar four graphs in Figure 11 have a Chinese character in the labels for the x-axis, "月" meaning "month", which is correct for what the graph is saying.  If it isn't my incorrect font, would it be possible to change the labels to lose that character, for those readers who don't read Chinese so well?

Section 3.4 does not have any statistical or numerical commentary, so it is just an aside to comment on inundation extent?  Or is it a critique of model 5.  The section seems like a bit of an add-on. 

Page 6, line 197, at the bottom of the page, there is an unusally large gap between the end of the second-last-sentence, and the last sentence.  This might be an artifact of Figure 5 appearing at the top of the following page.

Author Response

Reply submitted as an attached file

Author Response File: Author Response.docx

Reviewer 3 Report

The manuscript deals with the simulation of the inundated areas in a data scares region; the simulation is performed by using hydrologic/hydraulic modeling that make use of global datasets or proxy information. Specifically, hydraulic modeling is based on a global DEM with two different resolutions or a DEM derived from inundated areas satellite images; furthermore, two different boundary conditions are used for hydraulic simulation, namely a stage-discharge relationship and water level time series. By combining DEM source and boundary conditions, six different hydraulic models are derived. The model are calibrated and tested based on the available observations; they perform slightly different in terms of NS and r^2 metrics, yet none of them outperforms the others. Bayesian Model Averaging (BMA) is proposed and tested against single simulations.

The manuscript is well organized and almost well written. Based on my opinion, the Authors should spend some more efforts to better explain the rationale behind BMA, especially to let the readers understand the difference between the NS and r^2 metrics and the performance score used by BMA, that is the Bayesian Model Evidence (BME). Indeed, the Authors compare results in terms of BMA weights (that are nothing else that the normalized model evidence) with those obtained by using NS and r^2 without explaining the meaning of the former quantity; I also suggest to improve the mathematical notation of BMA for the sake of clarity. Furthermore, the Author should also better explain the motivation of their work, by discussing why first witch are the limits of single model simulation and then providing the conditions under which BMA might be a solution. As for the results, I do not see a large improvement on simulation accuracy by using BMA instead of single models, especially if we compare results at the two stations both in the calibration and validation periods; I would expect the Authors to better discuss on this.    

Author Response

reply submitted as an attached note

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors have done a great job improving the manuscript and added merit to this publication.  I appreciate the effort and recognize the improvement in the manuscript.  While, I do not view this manuscript having high scientific impact in that it remains a calibration exercise with a Bayesian Model Averaging approach, I never the less now see the value of identifying appropriate DEM data to use in modeling flood propagation.  Furthermore, the authors have now done a robust job of describing why they are interested in this type of data, and why they choose specific calibration parameters.

Still my main comment in the first review is that it is plainly obvious that a weighted average of modeling attributes as done in the BMA.  Here, I am not as convinced the authors have fully addressed my gripe.  If the purpose is to simply provide ‘more reliable water lever and discharge simulations’ as stated in the last sentence L802-803, then more description of how the BMA model will be applied outside of the calibration/verification exercise is needed.  But, if however,  the purpose of the BMA exercise was to identify more and less reliable model attributes from the original six models, or to provide information regarding how the ‘best model’ might look,  then I would suggest highlighting that more. 

Author Response

Please find attached the answer.

Author Response File: Author Response.docx

Reviewer 3 Report

The Authors have almost addressed all the reviewer comments. They could have spend much efforts to see the original mathematical notation used by Hoeting et al. instead of considering only that used in more recent papers. Anyway, the manuscript could be of interest for the scientific community and I believe it deserves to be published. 

Author Response

Please find attached the answer.

Author Response File: Author Response.docx