|This paper has been significantly revised following my earlier comments with substantive new work undertaken to address the points I raised. My thanks to the authors for engaging so constructively and I hope they feel the process was helpful.|
This new works address most of my key concerns. It better explains, and provides evidence, that CEHGear-1Hr can be considered as a benchmark, the rainfall data themselves are now analysed, background information has been added, and a more substantive analysis has been undertaken including further validation of the hydrodynamic model. All this work more clearly identifies the differences between reanalysis precipitation products and a national data set derived from rain gauge data.
There were, however, still a few things in the new explanations that I was either not fully convinced by or did not understand. In these cases, I think a modest amount of further explanation or justification is necessary because the current arguments do not, in my view, stack up. These remaining points are:
1. Lack of river channels in the hydrodynamic model. The revised paper argues that representation of river channels is not necessary for large events and cites two papers that it is claimed back this up (new text on lines 88-93). I’m a co-author on one (Neal et al, 2021), and I have to say this is absolutely not what our paper shows. In Neal et al (2021) there is no comparison of models with and without bathymetry, but there is in his Jeff’s 2012 paper:
Neal, J., Schumann, G.J.-P. and Bates, P.D. (2012). A simple model for simulating river hydraulics and floodplain inundation over large and data sparse areas. Water Resources Research, 48, Paper no. W11506, (10.1029/2012WR012514
And what this shows is crystal clear: representation of channels is critical for correct simulation of flood propagation. In fact, the reason for developing the different ways of representing channels in Jeff’s 2021 paper was precisely because the 2012 findings showed that this was so necessary. The paper by Dey et al (2019) has also been misunderstood. Like Neal et al (2021), Dey et al compare different ways of implementing the bathymetry in large scale hydrodynamic models and also do not ever consider the ‘no bathymetry’ case. The statement quoted in the text that “the choice of bathymetric model becomes irrelevant at high flows for predicting hydraulic outputs” does NOT mean that bathymetry irrelevant at high flow as claimed. Rather, it says that the choice of channel shape and how much longitudinal spatial variability needs to be included is unimportant, but the paper is clear that a channel of some form is still needed. In fact, two sentences on from the quoted test, Dey et al make the statement “1D hydraulic models for high flows do not require incorporation of geometric variability in channels and even using a simple triangular or rectangular shape is sufficient for flood modeling purposes”. Dey et al’s opening sentence is “Accurate representation of river bed topography, commonly referred to as bathymetry, plays a critical role in a variety of hydrologic and hydraulic applications including but not limited to flood modeling”. The papers by Dey et al and Neal et al are remarkably consistent, but do not show what you claim. Moreover, the importance of bathymetry has been recognised all the way back to the very earliest papers on large scale hydraulic modelling e.g. Kate Bradbrook’s work for JBA in the early 2000s.
The use of citations in the revised paper to support not including bathymetry is therefore both wrong and selective: you must surely have read the rest of the paragraph in Dey et al to see that they flatly contradicted the point you were seeking to make only a couple of sentences further on?
Moreover, there are good physical reasons why channels are central to flood routing and inundation prediction even during extreme flows. First, even at high flows a good proportion of the discharge still goes through the channel because of the high velocities there. From flume experiments and field observations we know that floodplain flow velocities are often an order of magnitude lower than those in the channel, so channel conveyance is still a major component of the total flux. Let’s just think about the Carlisle 2005 event. In Carlisle, the Eden is ~70m wide and ~5m deep. With a conservative average channel flow velocity during the event of 1m/s that’s 350 m3s-1 out of a total flow of around 1600 m3s-1. And in terms of channel flow:floodplain flow ratio, Carlisle 2005 is likely to be an extreme (low) end member. Channel velocities during the 2005 flood were probably even higher, and the event itself was 150 year return period so there was a lot of floodplain inundation. If you have no channel, then the unaccounted for channel flux has to be distributed over the floodplain and that must lead to a significant overestimation of flows there. Second, flood waves do not propagate in a physically realistic way without a fast-moving channel flow filament set within a slower moving floodplain background field. The Neal et al (2012) paper shows this very clearly, even when momentum exchange between channel and floodplain is not included.
I fully accept that your model is getting reasonable results, but the justification for not including bathymetry does not stand up to scrutiny, and actually runs counter to everything we know about channel-floodplain hydraulics. The errors so generated will be significant (my back of an envelope calculation using the real geometry suggests ~20% of the discharge for Carlisle 2005) and leads to the suspicion that your model may only be getting the right results for the wrong reasons.
I’m sorry, but you are going to have to provide improved text to justify not including channels. Iif plausible arguments cannot be found then you will have to be clear about the errors this can generates and explain why your model still appears to perform well in spite of these (i.e., explore what might be going on to compensate).
I think you also have to better explain how the model can get the stage-discharge relationship correct, as Figure 7 seems to show, when you do not have any channel bathymetry in the model. My understanding of the model simulations is that you run a ‘warm up’ period to get ‘normal flow’ in the river channels (line 116) which I assume means below bankfull discharge for the real geometry. The event simulation then takes the flow from what would, in real life, be in- to out-of-bank flow. Without a channel geometry (however approximated) how can stage in the model respond correctly to changing discharge? Figure 7 shows that the change of stage with discharge in the model is similar to the observations, but the latter implicitly includes channel bathymetry whereas the model does not. I am at a loss to explain this and just don’t understand how this can work.
2. Calibration. One thing that might be a compensating effect which you don’t discuss at all in the paper is calibration. How was this done, if at all, and what model parameters values did you end up using? I should probably have identified this in my previous review, but could you also add a short section outlining how the model was calibrated and commenting on the extent to which the optimum parameters are in the right physical range?
3. Lack of infiltration. I was more convinced by your justification for not including infiltration during extreme events as that seems at least physically plausible, but the paper you cite to justify this (Hossain Anni, 2020) is a small-scale study in an urban catchment, so again does not show quite you claim. Please could you replace with references showing the same effect at catchment scales.
4. Lack of flood defences. I was surprised how you got such good results in Carlisle without including flood defences. Even in 2005 Carlisle was well defended and I don’t think these were always overtopped, so this needs a couple of sentences of explanation in the paper.
5. Line 116. Could you define ‘normal’ flow? I assume it is somewhere below bankfull discharge (see above) but I think this needs to be clear.
6. Digital Elevation Model. I still can’t quite understand what is happening with the DEM data. Thanks for confirming that OS Terrain 50 is photogrammetry and not LiDAR, but that does mean you are using a DEM with 4m RMSE of vertical error. If that is genuinely the case, how did you get inundation and water depth predicted as well as you do? Either the actual vertical error over the areas you simulate is substantially lower than 4m or there is something odd going on in the model to compensate.
First, I think you need to be clear about the DEM error in the current paper. You don’t mention this anywhere, but it is huge (DEM RMSE >50% of the observed flood wave amplitude for all catchments).
Second, you need to explain how the model is still able to produce the results it does despite these terrain data errors.
Lastly, your statement regarding OS Terrain 50 that it “has been shown to perform best for flood risk modelling in a comparison with other DEMs (McClean et al., 2020)” derives from a paper that did not include LiDAR data in the comparison. Instead, your previous work only compares OS Terrain 50 to a basket of global DEMs, which unsurprisingly have higher error. Numerous previous studies, starting with Sanders (2007):
Sanders BF. 2007. Evaluation of on-line DEMs for flood inundation modeling. Adv.Water Resour. 30(8):1831–43.
Have shown airborne LiDAR data sets to have considerable advantages for flood inundation modelling. In the above text from the paper you have used the word ‘best’ when you mean ‘better’ and have not qualified the subset of DEMs that your study refers to. Because of this your statement is misleading. The Yunus study you quote is also not the greatest evidence here because: (a) it is a bathtub model which ignores hydraulic connectivity and (b) the events simulated are extreme and valley filling such that inundation extent becomes a very insensitive metric for comparing DEMs. There are now many papers on DEMs in flood inundation modelling and you need to consult a range of these to revise this text instead of using a self-citation in an inappropriate way.
As an aside, I still don’t understand why you are not using LiDAR data where it exists? Most river floodplains in the region are now mapped and you could use these data and patch with OS Terrain 50 in hillslope areas away from the floodplain. LiDAR vertical errors are ~10cm RMSE and it is the gold standard.
7. Figures. Some of these are difficult to read because the panels are small compared to the detail you are trying to show or because you have used a low-resolution raster file format instead of vector. The latter issue means that text, axes and other figure elements are quite pixelated. You need to redo these images at higher resolution or (better) as vector graphics.
I don’t think too much additional work is needed to address these points assuming there are plausible arguments that can be made in response, but they are important to the credibility of the work. A further point I would make is that in a number of responses you mis-represent previous work (Neal et al, Dey et al) or quote it out of context (Hossain Anni, McClean et al) in order to justify some aspect of your approach. Everyone makes this mistake occasionally, but in my experience it is unusual for this to happen as often as it does here.
I hope these further comments are useful
University of Bristol