Review of “Using hydrological and climatic catchment clusters to explore drivers of catchment behavior” by Jehn et al.
Summary
The authors attempt to address the question: “where is hydrologic behavior similar across the contiguous United States?” They use Principal Component Analysis and quadratic regression to cluster catchments in the CAMELS data set located in the Contiguous United States. The variables of interest are 6 hydrologic signatures that earlier research has shown to have high spatial predictability for this dataset. The authors use 15 catchment attributes that were shown to strongly correlate to these 6 signatures to explain the generated clusters in terms of catchment similarity. They discuss which attributes are most influential in determining each cluster and take some steps towards interpreting this from a hydrological processes point of view.
I have read this paper with great interest. I find the separate correlation plots between the east and west CONUS (Figure 6) very interesting and am curious about the differences in eco-regions and hydrologic behavior between the east and west that this might imply. My main concern is that apart from Figure 6, the manuscript mostly seems to confirm earlier work by e.g. Addor et al. (2018), Berghuijs et al. (2014), Knoben et al. (2018) and Kuentz et al. (2017). Confirming findings is not a bad thing, but I think the authors are missing out on an opportunity to go beyond these studies. The authors spend some time in the main manuscript (L219-229; L256-258; Table 2 to some extent) speculating about hydrologic behavior in each of their clusters. More of these thoughts are hidden in the appendices (L460-511). I believe the manuscript would become much stronger if the authors would make this the main topic of the manuscript and spend more time on trying to understand the hydrologic behavior each cluster represents in terms of dominant processes, as this would be a novel contribution to the field. This could be structured similar to Berghuijs et al. (2014) but the CAMELS dataset gives the authors the catchment information needed to go beyond that work. Addor et al. (2018) could also help to outline potential changes to the manuscript. I also have a few methodological concerns about the way the signatures and attributes have been selected and how the number of clusters has been determined, and how these choices might limit the authors’ ability to go beyond these earlier studies (see below).
Major comments
1. The authors use the six most predictable signatures from Addor et al. (2018) for analysis here. They use 15 catchment attributes that Addor et al. (2018) indicates as having the highest ability to explain these signatures, with climatic attributes having the strongest connection to signatures. Earlier work (Berghuijs et al., 2014; Kuentz et al., 2017; Knoben et al., 2018) has also shown the strong influence climatic conditions have on hydrologic behavior and advocate for further studies that investigate the impact of less clear relations between catchment attributes, such as resulting from geology or vegetation, and hydrologic behavior. These relations can be seen in well-monitored experimental catchments and must logically exist in all other catchments, but in large-sample studies this has so far not been conclusively shown (Addor et al., 2018, provides various compelling reasons for why that might be the case). Progress towards understanding these relations would be an important contribution to the literature and interpreting cluster analysis from a dominant hydrological process perspective could be a first step. The authors take some steps in this direction, but unfortunately they are limited by their study setup to mostly confirm what has already been shown before, without having the necessary information available to go beyond these earlier studies. Restructuring of some of the study setup and analysis might be needed (see points 2 and 3 below).
2. Hydrologic interpretation is limited by the choice of signatures. I appreciate that the authors were looking for signatures that are easily predictable in space, but this limits the generality of the conclusions that can be drawn. The chosen six signatures do not describe the full hydrologic regime, focusing mostly on flow magnitude (mean annual, summer and winter flow, runoff ratio, Q95) and somewhat on seasonality (half-flow date), with no signatures dedicated to low flows, intermittency of flows, or response time of the catchment. Therefore statements such as the following are too general for the supporting analysis and should be rephrased to account for the specific conditions these 6 signatures describe (please note that this list could be incomplete):
- L95. “These two principal … overall hydrological behavior.”
- L188. “ … lead to similar (equifinal) discharge behavior”
- L464. “So over one third of the catchments in CAMELS show a relatively similar behavior.”
- L470. “… catchments with very different attributes can produce very similar discharge characteristics, …”
- L480. “an example of different catchment attributes being able to create similar discharge characteristics concerning their signatures, while having different catchment attributes”
Related to this, both Addor et al. (2018) and Knoben et al. (2018) show that these particular signatures correlate strongly with climatic conditions in the catchment. I doubt whether there is much to be learned about the influence of non-climatic attributes on hydrologic behavior by looking only at signatures with such strong connections to the prevailing climate. Using a wider range of signatures could allow more in-depth analysis of the relation between attributes and signatures. E.g. McMillan et al. (2017) could be of use in choosing different signatures:
McMillan, H., Westerberg, I., & Branger, F. (2017). Five guidelines for selecting hydrological signatures. Hydrological Processes, 31(26), 4757–4761. https://doi.org/10.1002/hyp.11300
3. Hydrologic interpretation is also limited by the choice of attributes, because the selected attributes are strongly correlated with one another. I had already written a few comments on this before reaching Figure 6, which shows that the authors are aware of these correlations. This knowledge should play a much larger role in the earlier parts of the paper, where the study setup is decided (i.e. which attributes to use) and where the importance of attributes for clustering is discussed (for example, the 5 most important attributes in cluster 3 are essentially 2 factors spread out over 5 attributes: snow & elevation are the first (r = 0.8), and various aspect of vegetation are the second group (r=0.7, r=1 and r =0.8). A different selection of attributes might be needed. I also believe that enforcing 3 attributes per attribute category is unnecessarily limiting and ignores some of the current understanding of drivers of hydrologic behavior, such as not using a climate seasonality metric (further details below).
Minor comments
L45. Addor et al. (2018) identified these signatures as having low spatial predictability in the US. Is it correct to assume that these conclusions also apply to the study domain of Kuentz et al. (2017), i.e. Europe?
L68. Is there a reason to assume that the most diverse total information is retained by using 3 attributes each from climate, topography, vegetation, soil and geology? To what extent are all CAMELS attributes correlated and to what extent is the subset of 15 attributes correlated with one another?
L69. [Adding to the previous comment] Among others, Berghuijs et al. (2014), Addor et al. (2018) and Knoben et al. (2018) have found that climatic seasonality is an important control on hydrologic behavior. The authors have included ‘frequency of high precipitation events’ over a climate seasonality metric. I agree that there can be good reasons to include the frequency of high P events metric but because the authors limit themselves to 3 attributes per attribute category, they cannot include a seasonality attribute even though current theory indicates that seasonality can be an important control on hydrologic behavior. Given this, I think the choice of 15 attributes and how they are distributed between the different categories needs to be better justified and possibly changed.
L95-97. I don’t think the two PCA’s of the six signatures can be seen as “describers of the overall hydrologic behavior”. This sentence and the next one need to be more nuanced, because the authors state in section 2.1 that no low flow signatures are part of their selection. Other possibly relevant aspects of the flow regime, such as baseflow or flashiness, are also not covered in this selection of signatures.
L119. Kuentz et al. (2017) use 10 clusters to group >35000 catchments using 16 different signatures. I expect that choosing to use 10 clusters in this study with >600 catchments and 6 signatures might provide unnecessary granularity. Can the authors somehow quantify the difference between each pair of clusters to show that 10 is an appropriate number? If such quantification is not possible, did the authors investigate the impact of using fewer or more clusters?
[additional note] Seeing that cluster 3 only contains 7 catchments and that cluster 5 only has 9, but that cluster 1 has 230 catchments in it, I think that some more discussion of the number of clusters is warranted. Cluster 5, 6 and 7 also look very similar, possibly indicating that too many clusters have been used. Some questions that come to mind:
- What is the explanatory power of a cluster with only a handful of catchments in it?
- Is the distribution between clusters so skewed because the catchment sample is not uniformly distributed across the selected attributes?
- Would more and/or different attributes provide more balanced clustering results?
- If the catchment sample is not uniformly distributed across attribute space, does this influence the PCA results?
L126. That aridity and forest fraction score highest could possibly relate to the high correlation between these two attributes. Investigating the correlations between the 15 catchment attributes could show how much independent information is contained in each. The same could be said about fractional snowfall and elevation. – Note: upon further reading I see that these correlations are in Figure 6. This information should be part of the text here.
L137. I don’t think calling these six signatures “more hydrologically meaningful” is supported by the findings of Addor et al. (2018). “more gradually varying in space” perhaps.
L144. “This can probably be extrapolated to most catchments in the continental US without human influence, as the CAMELS dataset contains large samples of undisturbed catchments”. This sentence is speculation and should be removed. If the authors want to keep this statement it could for example be supported by calculating the climate attributes used by Knoben et al. (2018) and comparing these to the range of values for these attributes found across the CONUS. This would show how climatically representative the CAMELS catchments are for the wider CONUS.
L185. See also Berghuijs et al. (2014) who find hydrologic similarity across comparable distance in the CONUS; or Kuentz et al. (2017) who find hydrologic similarity across comparable distances in Europe; or Knoben et al. (2018) who find catchments with similar hydrologic regime on different continents, using only climate indicators to describe similarity.
L189-195. I suspect that if correlations between attributes are taken into account, many of the attributes that are of high importance in each cluster turn out to be quite directly related one another. For example, (cluster 1) high aridity and low forest fraction & green vegetation fraction maximum will be inversely correlated; (cluster 3) precipitation and snow and elevation will be correlated, as will forest fraction and LAI maximum and green vegetation fraction maximum. Therefore I expect that this part of the analysis will be more instructive if these correlations are accounted for, either in selection of the attributes or by lumping correlated attributes into groups in some fashion. Changes to Figure 5 might be needed.
L214. “While aridity … single clusters (Figure 5).” Implying that aridity is not important in most of the clusters seems a bit of a stretch. Aridity is the most important attribute in 4 out of 10 clusters, and the second-most important in another 2. It appears in the top 5 of important attributes in 8 out of 10 clusters (and in the remaining 2 clusters the correlated forest fraction appears), more often than any other attribute.
L248. “Therefore, our selection of hydrological signatures seems to allow a better identification of hydrological similarities.” Unfortunately I think this argument can be reversed as well, in the sense that this selection of signatures might not capture enough of the details of the individual regimes to give the clustering approach any trouble. Because these 6 signatures are strongly related to climate (e.g. Addor et al., 2018; Knoben et al. 2018), and the relevant climate indices are (mostly) included in the clustering approach, it is not surprising that these signatures cluster easily. The fact that the authors don’t use a climate seasonality attribute, which has been shown to be an important driver of hydrologic differences, could potentially explain why their Cluster 2 does not seem to have any distinct character. Instead of making this statement and moving on, a strong contribution would be if the authors can determine how to make hydrologic sense of all the catchments that don’t seem to follow any obvious pattern. Would different attributes solve this?
L301. I’d argue that Cluster 4 seems to be firmly placed in the non-arid & snow-dominated region of the climate space. There are more catchments in this climate region that belong to different clusters but this is (1) inherent to imposing binary boundaries (catchments are either cluster X or Y, even if they are 49% similar to X and 51% similar to Y) and (2) because the climate plots in Figure 8 only look at a limited selection of possibly influential attributes (climatic or otherwise).
L310-315. This connection between signatures and climate can also be seen in Knoben et al. (2018) and Kuentz et al. (2017). Addor et al. (2018), Knoben at al. (2018) and Kuentz et al. (2017) (among others) acknowledge that using climate alone is not sufficient to produce a catchment classification system. This should probably be mentioned as part of this section (or in the introduction of the paper, because it provides a compelling reason for investigating catchment attributes).
Figure 8. Is the aridity axis upside down in these plots? More arid catchments seem to have higher flows.
Figure A2. I like the way violin plots look, but kernel density smoothing does not respect physical boundaries very well and distorts the data being plotted. See for example cluster 3 and the mean winter discharge signature, which is, according to the violin, a negative flux for some of the catchments in this cluster. Histograms or box-and-whisker plots would more accurately reflect the data.
Figure A3. See comment above.
Figure A4. Is the aridity axis upside down in these plots? More arid catchments seem to have higher flows.
Typographical
L30. “those” > “this”?
L119. Kuentz et al. (2018) > Kuentz et al. (2017)
P11. Caption of Figure 5. “For the catchment clusters.” should not be a stand-alone sentence.
L215. “single” > “individual”?
L251. I understand what this sentence is meant to say but it doesn’t quite work. Is “This human influence might mask otherwise apparent patterns.” better?
L261. “have” > “has”
L265. “cluster” > “clusters”
L463. “… the low elevation those catchments are located, …” > “… the low elevation those catchments are located at, …”
L500. “cluster catchments”. Should the word “catchments” be here? |
- Printer-friendly version
- Supplement
