the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 15 Jan 2026
Attribution of Changes in small and large Floods across Brazil
Abstract. In tropical regions, flood changes are driven by a combination of event rainfall characteristics and antecedent wetness changes. However, how the interactions between storage capacity, event rainfall, and antecedent wetness influence flood changes across event magnitudes is elusive. Here, we explore the causes of changes in small and large floods by combining flood elasticities with trends in event rainfall peak and pre-event antecedent wetness of 765 catchments in Brazil. Our results suggest that large floods are increasing more than small ones, corresponding to 80 % of substantial flood increases. While those changes in large events are usually rainfall-driven, changes in small floods are mostly aligned with changes in antecedent wetness. We find that in regions with high water storage capacity, antecedent wetness drives changes in both small and large floods. Conversely, in regions with low water storage capacity, changes in small floods are driven by antecedent wetness, whereas large floods are mainly rainfall-driven, as rainfall outweighs antecedent wetness in those fast-saturating catchments. Our findings highlight that reliable predictions of flood responses to climate change should account for both event magnitude and catchment storage capacities, as climatic drivers alone are insufficient to fully explain flood changes.
- Preprint
(2473 KB) - Metadata XML
-
Supplement
(1531 KB) - BibTeX
- EndNote
Status: open (until 11 Mar 2026)
- RC1: 'Comment on egusphere-2025-6456', Anonymous Referee #1, 18 Feb 2026 reply
-
RC2: 'Comment on egusphere-2025-6456', Anonymous Referee #2, 03 Mar 2026
reply
The manuscript “ATTRIBUTION OF CHANGES IN SMALL AND LARGE FLOODS
ACROSS BRAZIL” by Anzolin et al. analyses trends in floods of different magnitudes in Brazil and the relative contribution of two drivers (peak rainfall and antecedent wetness state) to these changes. The authors find that in low water storage capacity regions, different drivers explain trends in floods of different magnitude, while, in high water storage capacity regions, antecedent wetness explains both. The analyses and findings are clearly presented and well supported by figures. Please find my comments below.
L11: “corresponding to 80% of substantial flood increases “. Unclear, please rephrase.
Section 2.2: I have some doubts about the event selection criteria adopted. Authors sample flood events based on a peak-over-threshold approach based on rainfall magnitudes and further refine the selection by excluding small flood peaks. Their rationale is not to bias the selection towards events with wet antecedent conditions. However, peak rainfall is also a driver used later in the attribution analysis and I believe that the event selection should be independent of the drivers. With the current criteria, events with very wet initial conditions and very small event rainfall are excluded from the analysis. Isn’t the selection then biased towards high rainfall events? Also the fact that further refinement (too small flood peaks are later excluded) is, to me, an indication that this is more an “event selection” rather than a “flood event selection”. I suggest repeating the analysis with different event selection criteria (e.g. pot or annual maxima based on floods only) that are independent on the drivers at least to prove that the selection is not biasing the results.
Section 2.3:
- The attribution model is based on local regressions between flood quantiles and the drivers. I am wondering how robust the regressions and results over large floods are with limited flood series. I suggest that the authors could add something about the uncertainty of the results. Perhaps a regional approach (e.g. over the hotspots) could increase the robustness of the attribution results.
- Small and large floods are defined by 0.5 and 0.95 flood quantiles. It would be useful to add a statement about equivalent return periods of such flood events in order to make findings comparable with previous studies and also more easily interpretable by practitioners.
- What about catchment area effect? Your catchments vary across 5 or 6 order of magnitude. One would expect that different processes may be relevant.
L173-175: unclear sentence. Do you mean heavier distribution tails?
Section 3.2: please check units of elasticity. Is it % or %/% (as it appeared to be defined in section 2.3)?
L194: is the median calculated over the hotspots?
Figure 7: it would be useful to see results about the absolute contributions as well. Those are useful because show whether drivers contribute to positive or negative flood changes. Also, relative contributions alone may return a biased picture in cases where absolute changes are small.
L259-270: I find that the comparison of the findings with the literature should be a bit clarified. No direct comparison is possible because of the different small/large flood definitions (here I think large floods correspond to ~10year return period, which is not really comparable to 100yr or event with significant initiation. Also the (flood) event selection used here and, in some cases, the local vs regional nature of the findings, does not allow a direct comparison.
L303: unclear sentence.
Citation: https://doi.org/10.5194/egusphere-2025-6456-RC2 -
RC3: 'Comment on egusphere-2025-6456', Anonymous Referee #2, 03 Mar 2026
reply
The manuscript “ATTRIBUTION OF CHANGES IN SMALL AND LARGE FLOODS
ACROSS BRAZIL” by Anzolin et al. analyses trends in floods of different magnitudes in Brazil and the relative contribution of two drivers (peak rainfall and antecedent wetness state) to these changes. The authors find that in low water storage capacity regions, different drivers explain trends in floods of different magnitude, while, in high water storage capacity regions, antecedent wetness explains both. The analyses and findings are clearly presented and well supported by figures. Please find my comments below.
L11: “corresponding to 80% of substantial flood increases “. Unclear, please rephrase.
Section 2.2: I have some doubts about the event selection criteria adopted. Authors sample flood events based on a peak-over-threshold approach based on rainfall magnitudes and further refine the selection by excluding small flood peaks. Their rationale is not to bias the selection towards events with wet antecedent conditions. However, peak rainfall is also a driver used later in the attribution analysis and I believe that the event selection should be independent of the drivers. With the current criteria, events with very wet initial conditions and very small event rainfall are excluded from the analysis. Isn’t the selection then biased towards high rainfall events? Also the fact that further refinement (too small flood peaks are later excluded) is, to me, an indication that this is more an “event selection” rather than a “flood event selection”. I suggest repeating the analysis with different event selection criteria (e.g. pot or annual maxima based on floods only) that are independent on the drivers at least to prove that the selection is not biasing the results.
Section 2.3:
- The attribution model is based on local regressions between flood quantiles and the drivers. I am wondering how robust the regressions and results over large floods are with limited flood series. I suggest that the authors could add something about the uncertainty of the results. Perhaps a regional approach (e.g. over the hotspots) could increase the robustness of the attribution results.
- Small and large floods are defined by 0.5 and 0.95 flood quantiles. It would be useful to add a statement about equivalent return periods of such flood events in order to make findings comparable with previous studies and also more easily interpretable by practitioners.
- What about catchment area effect? Your catchments vary across 5 or 6 order of magnitude. One would expect that different processes may be relevant.
L173-175: unclear sentence. Do you mean heavier distribution tails?
Section 3.2: please check units of elasticity. Is it % or %/% (as it appeared to be defined in section 2.3)?
L194: is the median calculated over the hotspots?
Figure 7: it would be useful to see results about the absolute contributions as well. Those are useful because show whether drivers contribute to positive or negative flood changes. Also, relative contributions alone may return a biased picture in cases where absolute changes are small.
L259-270: I find that the comparison of the findings with the literature should be a bit clarified. No direct comparison is possible because of the different small/large flood definitions (here I think large floods correspond to ~10year return period, which is not really comparable to 100yr or event with significant initiation. Also the (flood) event selection used here and, in some cases, the local vs regional nature of the findings, does not allow a direct comparison.
L303: unclear sentence.
Citation: https://doi.org/10.5194/egusphere-2025-6456-RC3 -
RC4: 'Comment on egusphere-2025-6456', Anonymous Referee #2, 03 Mar 2026
reply
The manuscript “ATTRIBUTION OF CHANGES IN SMALL AND LARGE FLOODS
ACROSS BRAZIL” by Anzolin et al. analyses trends in floods of different magnitudes in Brazil and the relative contribution of two drivers (peak rainfall and antecedent wetness state) to these changes. The authors find that in low water storage capacity regions, different drivers explain trends in floods of different magnitude, while, in high water storage capacity regions, antecedent wetness explains both. The analyses and findings are clearly presented and well supported by figures. Please find my comments below.
L11: “corresponding to 80% of substantial flood increases “. Unclear, please rephrase.
Section 2.2: I have some doubts about the event selection criteria adopted. Authors sample flood events based on a peak-over-threshold approach based on rainfall magnitudes and further refine the selection by excluding small flood peaks. Their rationale is not to bias the selection towards events with wet antecedent conditions. However, peak rainfall is also a driver used later in the attribution analysis and I believe that the event selection should be independent of the drivers. With the current criteria, events with very wet initial conditions and very small event rainfall are excluded from the analysis. Isn’t the selection then biased towards high rainfall events? Also the fact that further refinement (too small flood peaks are later excluded) is, to me, an indication that this is more an “event selection” rather than a “flood event selection”. I suggest repeating the analysis with different event selection criteria (e.g. pot or annual maxima based on floods only) that are independent on the drivers at least to prove that the selection is not biasing the results.
Section 2.3:
- The attribution model is based on local regressions between flood quantiles and the drivers. I am wondering how robust the regressions and results over large floods are with limited flood series. I suggest that the authors could add something about the uncertainty of the results. Perhaps a regional approach (e.g. over the hotspots) could increase the robustness of the attribution results.
- Small and large floods are defined by 0.5 and 0.95 flood quantiles. It would be useful to add a statement about equivalent return periods of such flood events in order to make findings comparable with previous studies and also more easily interpretable by practitioners.
- What about catchment area effect? Your catchments vary across 5 or 6 order of magnitude. One would expect that different processes may be relevant.
L173-175: unclear sentence. Do you mean heavier distribution tails?
Section 3.2: please check units of elasticity. Is it % or %/% (as it appeared to be defined in section 2.3)?
L194: is the median calculated over the hotspots?
Figure 7: it would be useful to see results about the absolute contributions as well. Those are useful because show whether drivers contribute to positive or negative flood changes. Also, relative contributions alone may return a biased picture in cases where absolute changes are small.
L259-270: I find that the comparison of the findings with the literature should be a bit clarified. No direct comparison is possible because of the different small/large flood definitions (here I think large floods correspond to ~10year return period, which is not really comparable to 100yr or event with significant initiation. Also the (flood) event selection used here and, in some cases, the local vs regional nature of the findings, does not allow a direct comparison.
L303: unclear sentence.
Citation: https://doi.org/10.5194/egusphere-2025-6456-RC4
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 195 | 82 | 13 | 290 | 34 | 11 | 9 |
- HTML: 195
- PDF: 82
- XML: 13
- Total: 290
- Supplement: 34
- BibTeX: 11
- EndNote: 9
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This manuscript examines changes in small and large floods in Brazil as well as explores potential causes of these changes. The manuscript finds that large floods are increasing more than small floods and that climate drivers are not able to fully explain the reasons for these changes - this is an important and interesting finding with implications for management of hazards and water supply. For these reasons, the manuscript is suitable for HESS.
I have a number of comments to improve support for the findings and one methodological question regarding the use of multiple linear regression to quantify the elasticity of different explanatory variables to changes in small and large floods. Overall, these comments tend towards Minor Revision but a few could be considered Major comments, in that new analysis may be needed rather than edits to the existing text.
I will note that the text of the flood event detection in Section 2.2 is thorough and thoughtful in the application of trend detection methods. In particular, it is worth noting the excellent attention to details regarding the assurance that methods to isolate large and small floods are robust to catchments of varying size. This is a careful and important consideration that is often overlooked.
The one major comment that I have is with respect to the use of coefficients resulting from a linear regression model as elasticities. This approach is used to assess the sensitivity of changes in large and small floods to changes in rainfall and antecedent wetness, forming the foundation of these conclusions. Linear regression finds the optimal model coefficients that minimize the sum of square errors between a dependent variable and its predictor variables. The estimated coefficients are, therefore, subject to factors like omitted variable bias and multicollinearity thus making the coefficients a less desirable approach to understand the sensitivity of a dependent variable to changes in a predictor variable. The coefficients are appropriate for prediction but not very useful for physical interpretation. There are other regression approaches, namely panel regression, that can formally treat regression coefficients as elasticities. There is also a large literature on formal causal attribution approaches.
For this manuscript, I wondered why these approaches were not applied and how the results and interpretation might differ if these more formal methods were applied to assess causes for changes in large and small floods.
In the current version of the manuscript, I would request that the introduction be expanded to discuss these approaches and provide justification, if possible, for why the application of linear regression to assess elasticity is appropriate. The introduction at present does not provide any justification and elasticity is only briefly mentioned in L40. The introduction should be expanded to include (1) a discussion of the use of causal attribution for the attribution of flood changes, (2) a discussion of panel regression and why it was not applied here, and (3) justification for the use of elasticities via linear regression to attribute reasons for flood changes.
Later in the text, section 2.3 explains that “the log-log regression allows us to interpret the model coefficients as elasticities” but I do not think this is correct given the discussion above. If the authors continue to use the linear regression approach, a much more detailed justification is needed as to how omitted variable bias and other limitations are addressed. The authors do note that they consider multicollinearity but not omitted variable bias which could also affect the coefficient estimates.
A potential implication of the use of linear regression comes in the form of some counter-intuitive results. For example, L216 states that “there is a noticeable shift in the sign of rainfall changes (from negative to positive) between large and small events.” This seems counter intuitive that the contribution of rainfall would not always be positive in terms of its effects on flood changes. A negative sign indicates that more rainfall would lead to less flooding, which does not make sense (assuming I am reading this statement correctly). Could this result be an artefact of the use of linear regression-estimated coefficients for the elasticities?
Another potential indication that this may not be the appropriate approach is that you needed to apply a constraint on the parameters because you are getting negative elasticity values (L152).
Minor comments:
- The period of analysis ends in 2018, which is now almost 8 years ago. Is it possible to include the additional 7-8 years of data beyond the measurements included in the CAMELS-BR dataset? For example, can you use the streamgauges in the CAMELS-BR gauge list but obtain the additional years of data?
- L78: What is the justification for using the multiplier of 1.11? Was a sensitivity analysis conducted here to assess this choice or was that part of your visual inspection and other checks on the approach? It might be helpful to clarify this.
- L75-85: Was a software package used to perform this analysis or did you use your own original code to make these calculations? If you did use an existing software package, please cite.
- Figure 2: It is very difficult to distinguish between the grey and black lines. Perhaps consider another way to present this information.
- On L142-143, the comment is made: “Assuming that the contribution of the drivers is additive…” I wondered if this is correctly stated because taking the logs of the data and using the log values in the linear regression would imply that the contribution of the drivers is multiplicative.
- L145: The text says “we check for these correlations [between the drivers]” however, I think this more precise to say that we “check the variance inflation factors for correlations between the drivers.” This is an important point because variance inflation can lead to problematic coefficients and therefore unreliable elasticity estimates.
- L171: Please include the percentage of small floods in parenthesis exactly as you showed for the proportion of large floods. You cannot assess the accuracy of this statement without the other percentage shown.
- L173-174 are confusing. The sentence reads that large events are increasing more and decreasing less; however in L172, it appears that large events are decreasing more and increasing less (the range shown is -76.6 to +59.2). Can you clarify this?
- L176-180 refers to “hotspots”; however, it is unclear what geographic region corresponds to these hotspots because there is no reference to a figure to support these statements. Please add supporting evidence.
- L200: The general comment is made that “The elasticity of floods to antecedent wetness is noticeably lower than that to rainfall peaks…” Clarification is needed as to whether you intend to say all floods (small and large floods). Later in the sentence you differentiate between large and small floods. I would change to read: “The elasticity of both large and small floods to antecedent wetness is noticeably lower than that to rainfall peaks…”
- Please provide evidence to support your statements in L228-230.
- Figure 7 is not standalone and the colors are not ideal. The hotspot abbreviations should be spelled out in the caption or on the figures. It is also confusing to use blue and green in the boxplots to differentiate between wetness and rainfall peaks and then re-use the same two colors to designate hotspots in panel c, which have no relation to the interpretation of those colors used in panels a and b.
- In Section 4.2, consider also citing Collins et al. (2022), who also show large floods in the US have not changed over the past 50 years.
Collins, M. J., Hodgkins, G. A., Archfield, S. A., & Hirsch, R. M. (2022). The occurrence of large floods in the United States in the modern hydroclimate regime: Seasonality, trends, and large-scale climate associations. Water Resources Research, 58, e2021WR030480. https://doi.org/10.1029/2021WR030480
Editorial comments:
- L76: Consider rewording as: “To accomplish this, local minima are identified…”
- L255: Consider changing “timidly explored” to “explored only in a limited way in Brazil”
- L259-270: Nice discussion section overall; however L265 contains an incomplete sentence. The text reads: “In Australia (Wasko & Nathan, 2019), and globally (Wasko et al., 2021).”
- L284: Change to read: “...and driving mechanisms for these changes in Brazil.”
- L285 and L284: Change the word “sign” to “direction”. Consider this change across all mentions of “sign” in the text to improve clarity.