Response to Anonymous Referee #1

We thank the reviewer for taking time to read our manuscript and we appreciate their comments. It appears that reviewer # 1 has reviewed our manuscript from a hydrology lens, as opposed to a land surface modelling point of view, and this is reflected in their feedback around calibration of the model and the use of the bias corrected data. 

In the following we address all comments from anonymous reviewer #1’s. The reviewer’s comments are shown in bold and italic and our response is shown in the non-italic font.

The paper discusses the two methods used to conduct these sorts of studies and selects one approach without substantiating the claim that it will give robust results despite the biases shown in the paper regarding the main climate variables (temperature and precipitation) as well as runoff.

We would like to clarify here that we did not intend to advocate the use of one approach over the other one by saying the selected approach is superior. The two approaches: 1) the use of raw climate model output, and 2) the use of bias corrected model output to drive a given model offline, both have their strengths and limitations for assessing climate change impacts on hydrology. The introductory section (lines 51-92) of the manuscript highlights the strengths and limitations of the two approaches that are used to study climate change impact on hydrology of river basins. Here, we have followed the first approach, clearly outlining its strengths (that the simulated change is consistent with the prescribed change in radiative forcing, and across Canada) and its limitations (primarily that model climate is not perfect). The second approach has its own limitations as the introductory section in the manuscript thoroughly explains. Bias-correction is not entirely a bias-free methodology. Furthermore, setting up and calibration of a hydrologic model for six large river basins across Canada by correctly representing regulation, and projecting future climate impacts by downscaling and bias correcting ensemble of GCMs is very a challenging task, and that is why it has only been done for a selected few river basins in Canada and that too considering one river basin at a time. We believe that the two  approaches 1) and 2) are complimentary to each other. The question investigated in our study is “can important insights be gained in the context of hydrological climate changes impacts in large basins, in spite of the biases in climate?”. We believe our results demonstrate the value that can be derived from analyzing raw climate model outputs. If given the opportunity to revise our manuscript, we will clarify these points further. 

The approach or routing climate model runoff is not new as it has been around for a few decades and the authors do not develop it further to include anthropogenic effects of regulation, nor natural effects of lakes and river ice processes except in a very simplistic way.

Routing is an age-old exercise. The objective of this manuscript is provide a consistent overview of climate change for the entire Canadian land mass on simulated components of the water balance including streamflow and not to refine routing by including regulation, lakes and ice processes. Please note that we do highlight limitations in our routing model and that’s why we use naturalized streamflow for the Columbia River. We were unable to obtain naturalized streamflow for the Nelson River from the Manitoba hydroelectricity company. Regardless, even if we had included anthropogenic flow regulation for the Nelson River and included a full lake model for the St. Lawrence River, the streamflow at the mouth of the Nelson and St. Lawrence rivers shows very little seasonality. For both these rivers climate change information is provided by changes in mean annual quantities of runoff. 

There is no attempts either to calibrate it to improve the match between observed and simulation streamflow.

This argument follows directly from use of models that are driven offline (i.e. not within the framework of climate models) and typically driven with bias-corrected climate data. We would like to note that land surface models (LSMs) which are used in climate models are different from hydrological models. The primary model output quantities from hydrological models are runoff and streamflow, and calibration until a good fit with observed streamflow is obtained is a routine exercise for hydrologic modelling studies. This is not the case for LSMs which cannot be calibrated for a given or multiple river basin(s). LSMs are process-based and the only thing that spatially varies in the models are the prescribed geophysical fields such as land cover, soil texture, and soil permeable depth. The model (and its parameterizations) when driven with spatially varying climate data and spatially varying input geophysical fields lead to spatial distribution of runoff and other model simulated quantities. LSMs simulate energy, water, momentum, and CO2 fluxes and it is not possible to calibrate LSMs for a given river basin just on the basis of runoff or streamflow. Again, the important question here is “can useful information be derived in the context of future hydrologic changes from land surface models based raw climate model output which unlike hydrologic models cannot be calibrated against a single model output?”. We will provide further clarification on this question if given the opportunity to revise our manuscript.

The authors also claim that their results would be useful for adaption but adaptation is usually done at much smaller scales than presented and the changes to the flow regimes at the mouth of a large river basin are not indicative of changes occurring at the sub-basin level of those rivers which are known to be heterogenous (e.g. the Mackenzie western headwaters are very different from the eastern Canadian shield tributaries, the Nelson headwaters are mountainous but then it flows through the prairies where usual routing fails) .

It seems that this the following sentence …

“Despite this, however, the response to climate change is relatively robust and there is useful information in the simulated change that can be used to inform adaptation measures.”

… in our manuscript is the source of this comment. This was a generalized statement. The scope of our study  covers the whole of Canada and considers only the major river basins. We do not imply that the results at the large basin outlets are valid for sub-basins levels. We agree with the reviewer that flow responses can be very different at sub-basin scales.  The routing model output will be made available publicly and discharge can be extracted and evaluated at major sub-basin levels. However, within the scope of this manuscript it is not to delve into smaller basins. If given the opportunity to revise our manuscript, we will revise this sentence and provide further clarifications on the above mentioned points.

The only positive about the paper is that it uses a consistent set of forcing data from CanRCM4 to drive the routing model. However, using a monthly time step for presenting the results does create artefacts for hydrograph time shifts and peaks as the precision is months, not days or weeks. Overall. I do not see any novelty in the study to qualify it for publishing as a research paper. It can be a climate change impact study report.

Thank you for noting that using a consistent set of results from CanRCM4 is a positive aspect of our manuscript. 

In regard to daily vs. monthly streamflow, we do have daily streamflow available from our model output. If given the opportunity to revise our manuscript and if instructed by the handling editor, we will be happy to modify monthly streamflow plots to daily streamflow plots. Please note that, however, flow duration curves are based on daily streamflow. Furthermore, if given the opportunity, we will elaborate on the novelty of our study including the use of physically consistent climate variables (as the raw output from a climate model is), that our projections are consistent with existing studies based on hydrological models (especially at the annual time scales), and that our projections are also continuous over Canada without any gaps.

The paper overlooks the availability of a bias-corrected and downscaled dataset for the CanRCM4 model which can be used to drive the model, at least to assess whether the validity of the bold claim of robustness made at the start. This dataset covers most of the North American domain and can be found at:  https://doi.org/10.20383/103.0622 and is described and partly assessed for the Mackenzie River Basin in Asong et al. (2020) (https://doi.org/10.5194/essd-12-629-2020).

As mentioned above, the use of bias-corrected climate data is one of the two approaches. The suggestions of calibrating our model to match streamflow and the use of bias corrected data, and the overall negative evaluation of our manuscript, indicates reviewer’s strong viewpoint that the use of hydrological models driven by bias-corrected climate data is the only viable approach to study climate change impacts on streamflow. However,  both approaches have been used in the existing literature  including in the IPCC’s assessments. The IPCC WG1 report is based primarily on raw climate model output while the WG2 report is based primarily on results from impact models (typically driven offline with bias corrected data). Both approaches have their pros and cons, and our study has demonstrated the value of using approach 1). In the end, it’s the consensus obtained by both approaches that leads to adaptation measures. 

Eq 2 lacks units, and S (which the channel slope) is not defined and contradicts the use of "S" on Figure 2. No mention was made on how S was obtained and how the equation is solved to get both the depth and velocity of flow.

using the mean +/- 1 std dev gives a confidence interval of 0.68 not 0.90 as suggested on line 652 - assuming the ensemble results is normally distributed - another assumption that was not substantiated.

Thank you for catching these typos. Yes, S is used twice. We will address this if given the opportunity to revise our manuscript. We started out by plotting the +/- 1 std dev range but later switched the plots to 90% confidence range. You are absolutely right +/- 1 std dev range corresponds to 68% confidence range. 

Response to Anonymous Referee #2

We thank the reviewer for taking time to read our manuscript and we appreciate their comments. 

In the following we address all comments from anonymous reviewer #2’s. The reviewer’s comments are shown in italic and bold font and our response in the non-italic font.

This paper suggested an approach that the outputs of climate models can be directly incorporated into climate change impact assessment at a basin scale. Although the authors addressed pros and cons on two approaches which has been usually employed for climate change impact assessment, there are main concerns for publication.

Mismatch of spatial resolution between CanRCM4 and river network: CanRCM4 provides outputs at 0.22° and 0.44° while river networks are from the TRIP dataset at 0.5°, which indicating the spatial resolution of runoff outputs were downgraded to the river networks. The main advantage of the approach employed in this study is to directly incorporate the outputs of climate models into impact studies at a watershed scale. However, this study downgraded the spatial resolution by interpolating outputs. As the river networks can be defined by topographical characteristics usually obtained from a digital elevation system (DEM), I recommend to newly define river flow directions at each grid cell of CanRCM4 to preserve the spatial resolution at 0.22° and 0.44° as well as to avoid additional biases from spatial interpolation.

Thank you for raising this point. We do not agree the difference between the resolutions of runoff (0.22 and 0.44 degree resolutions) and routing (0.5 degree) is a major source of bias in any way. We would like to draw reviewer’s attention to the following paper which addresses this exact same concern when it was first raised in the context of performing routing at a much bigger spatial scale in the CCCma climate model. 

Arora et al. (2001) Scaling aspects of river flow routing

https://onlinelibrary.wiley.com/doi/epdf/10.1002/hyp.161

In this paper the same amount of runoff, regridded to two different resolutions, is routed through two routing schemes. The first routing scheme is the same as used in this manuscript and routing is performed at a very coarse 3.75 degree resolution (~350 km) with only 20 grid cells representing the Mackenzie River basin. The second routing scheme called WATROUTE uses a resolution of around 25 km (~2700 grid cells). This is a very large difference in resolution. The results of this study showed that routing at these two vastly different spatial scales yields very similar streamflow characteristics at the mouth of the Mackenzie River. The primary purpose of routing, if done properly, regardless of the scale is to lag and attenuate the runoff over distances. The difference in resolutions is our manuscript (0.22/0.44 vs. 0.5 degree) is much smaller that the resolution difference in the Arora et al. (2001) study. As a result, we do not expect somewhat different spatial resolutions for runoff and routing to alter the results obtained in this study in any significant way. 

Please also note that upscaling river networks is not a trivial task. This is especially tricky given the rotated latitude longitude projection of the CanRCM. This is illustrated in Arora and Harrison (2007) (https://doi.org/10.1029/2007GL031865) who show that upscaling river networks is a cumbersome process and manual corrections are difficult to avoid. Given the minimal effect of scale on routed streamflow at the basin outlet as found in Arora et al. (2001) study mentioned above, it is does not seem that the effort involved in upscaling river networks to 0.22 and 0.44 degree resolutions is necessary in our study.

Regridding runoff: I do not agree that regridding runoff is straight forward as the responses (i.e., runoff) to climate forcing are depending on hydrologic characteristics (e.g., slope, soil, land cover, etc) averaged within a grid cell. In other words, different resolutions will induce different responses (runoff).

This comment is related to the previous comment. If regridding of runoff was not be performed then we would have had to perform routing at 0.22 and 0.44 degree resolutions by finding river flow directions at CanRCM’s rotated latitude longitude projection. Please note that regridding the rotated lat-long grid to regular lat-long grid is necessary, even if we were to use routing models at 0.22 or 0.44 degrees resolution, because as discussed above, developing routing network at the native rotated lat-long projection of CanRCM4 is a cumbersome task that requires manual corrections.  Also, as discussed above, routing is not expected to result in substantially different discharge responses, and is not a source of bias here.

Note that runoff is a response of the model to climate forcing (precipitation, temperature, radiation, etc.) and to spatial distribution of geophysical fields (soils, vegetation characteristics, land cover, etc.). This is why the model response (runoff in our case) is slightly different at 0.22 and 0.44 degrees, even when routing at the common 0.5 degrees resolution. The difference in model results at these two resolutions, amongst other reasons, is the result of non-linear response of runoff to climate forcing and geophysical fields. We are regridding the model response i.e. runoff for routing purposes not the actual forcings. Given these points, we believe it is reasonable to to regrid it conservatively to 0.5 degrees resolution, as discussed in the response to your previous comment in the context of the Arora et al. (2001) study.

Accuracy in primary variables and streamflow: From Figure 4, a considerable systematic bias in precipitation and streamflow was found, especially winter and spring flows at Columbia and Fraser, which may be from a combination of biases in climate forcing and a poorly-calibrated parameter set of CLASS. Thus, the bias may bring a question to me how reliable the climate change impact assessment is. I was wondering if a bias-correction method can be applied to the routed streamflow as a post process to improve the reliability of impact assessment.

It appears just like reviewer #1, reviewer #2 is also seeing our results from a hydrology lens. As we mentioned in our response to reviewer #1, land surface models (LSMs) cannot be calibrated to a single quantity (runoff/streamflow in this case). Unlike hydrological models, where streamflow is the primary output, LSMs simulate all energy, water, and CO2 fluxes and have 100s of parameters. Calibrating a model to produce streamflow that compares well with observations at one gauging station, would inevitably make other quantities compare worse with observations.  

If bias correction were to be applied to routed streamflow then it requires to make the questionable assumption that the nature of those biases is time invariant and applies to future climate as well. Bias correction of routed streamflow is especially problematic for snowmelt dominated streamflow regimes in Canada. A bias correction based on historical streamflow will not be likely suitable for future when shifts in peak streamflow and overall seasonality occur.  

The approach suggested in this study can be applied for only an existing river network (TRIP at 0.5°). It would be great if the authors suggest a general approach that can be applied to any resolutions (e.g., sub-watershed scales) by directly incorporating the outputs of CanRCM4.

We are unclear what this comment implies. When analyzing streamflow, the spatial resolution of model that generates runoff and that performs the routing may or may not be the same. The reason for choosing a 0.5 degree river network in this study was to be able to use runoff from both CanRCM resolutions given the cumbersome task of upscaling river directions to a rotated latitude longitude projection and the results from the Arora et al. (2001) study which indicates that routing is broadly insensitive to spatial resolution. Developing a generalized approach for routing CanRCM4 outputs at sub-watershed scales will need more investigation and evaluation, for example the minimum sub-basin size that can be suitably discretized at the CanRCM4 grid size. Such evaluation is beyond the scope of this study.

Novelty and contribution: I appreciate it if the authors prominently address the novelty and contributions of this study.

The novelty of our study is that it provides a consistent view of hydrological changes across Canada in a single study. While climate change impact studies are available for individual river basins, we are not aware of any Canada-wide study. If given the opportunity to revise our manuscript, we will highlight this novelty even more clearly. 

Specific comments

Line 68-70: Climate model driven climate signals need to be preserved. However, it is questionable that the information can be used to inform adaptation measures.



As noted in response to reviewer # 1 this was a generalized statement. The scope of our study covers the whole of Canada and considers only the major river basins. We will revise this statement if given the opportunity.

Line 84-88: There are bias-correction methodologies that preserve GCM-driven climate signals.

The intent here was to indicate that just like climate models have their distinct biases, bias-correction methodologies have their limitations as well. 

Line 295-296: The Nelson River is affected by heavily regulated flow. Thus, I do not think that this basin is a good testbed where the approach can be applied unless a set of naturalized flows can be employed.

Yes, we agree. As noted in response to reviewer #1 we were unable to obtain naturalized streamflow for the Nelson River from the Manitoba Hydroelectricity company. We were able to get naturalized streamflow for the Columbia River as noted in our manuscript. There’s a bigger question here as well. While as hydrologists and land surface modellers we may analyze climate change impact on naturalized flows, the fact of the matter is that rivers like Nelson will stay heavily regulated for the foreseeable future. In that sense, future change in mean annual quantities is of interest to the hydroelectricity companies, which this manuscript provides.



Line 669-672: If we do not have to look at the magnitude of changes, without routing processes, simply aggregating the runoff within a basin seems to be enough to evaluate climate change impacts.

This comment also appears to be related to earlier comments raising doubts on whether runoff can be routed after regridding. The regridding of runoff, as noted above in the context of the Arora et al. (2001) study, before routing is not the source of bias in any significant way. In addition, simply aggregating runoff within a basin cannot account for the lag and attenuation of runoff. Please note that while observations-based estimates for streamflow are available, there are no such direct estimates for runoff.

------------------------------------------------------

Overall comments for the handling editor

Regardless of critical reviews, this exercise has been enlightening for us as a land surface modelling group. Hydrological models have been traditionally calibrated against streamflow and the community accepts this norm as the basis of climate change studies as well. However, this is not possible for land surface models used in climate models. It appears that the hydrologic community may not be aware how land surface models are used within the framework of climate models where model parameters do not vary in space as a function of a given river basin. Rather model parameters are only dependent on geographically varying geophysical fields (soils, lands cover, vegetation types, etc.)

If given the opportunity to revise our manuscript, we will certainly address the broader issue of calibration, biases in raw climate model output, and the specific points raised by both reviewers.