Uncertainty of Rainfall Forecasts for Impact-Based Flood Warning in the Cagayan River Basin, Philippines

Kurihara, Yuta; Miyamoto, Mamoru

doi:10.5194/egusphere-2026-3482

Preprints

https://doi.org/10.5194/egusphere-2026-3482

Preprints

26 Jun 2026

| 26 Jun 2026

Uncertainty of Rainfall Forecasts for Impact-Based Flood Warning in the Cagayan River Basin, Philippines

Yuta Kurihara and Mamoru Miyamoto

Abstract. Effective impact-based flood early warning requires not only information on when heavy rainfall may occur, but also a calibrated estimate of the uncertainty in the resulting impacts. This study develops a framework for translating multi-centre sub-seasonal to seasonal (S2S) rainfall forecasts into probabilistic, municipality-level impact-based early-warning information for staged flood preparedness in the Cagayan River Basin, Philippines. Daily Gamma-kernel Bayesian model averaging is first applied to ensemble forecasts from ECMWF, NCEP, and UKMO to generate continuous predictive distributions of basin-mean rainfall. Leave-one-year-out verification for 2015–2025 identifies the ECMWF+NCEP+UKMO combination as the most robust tested input, with useful daily rainfall information mainly retained up to approximately lead days 5–6. The daily predictive samples are then accumulated into rolling seven-day rainfall distributions, because flood impacts in the basin are more closely related to multi-day rainfall than to isolated daily totals. Threshold-based probability recalibration improves the reliability of seven-day exceedance probabilities, raising Brier Skill Scores from −0.02 to +0.08 at 100 mm per seven days and from −0.07 to +0.02 at 150 mm per seven days relative to a monthly climatological baseline. The recalibrated seven-day rainfall distributions are subsequently propagated through municipality-level rainfall–damage functions to estimate probabilistic impacts on affected population, building damage, rice damage, and maize damage. An application to Typhoon Ulysses in November 2020 demonstrates how forecast-state-dependent impact intervals evolve as the event approaches and how municipalities with potentially large impacts can be prioritised. The results show that calibrated S2S rainfall probabilities can support uncertainty-aware, impact-based flood preparedness, while also highlighting limitations related to lead-time skill, basin-mean rainfall representation, upper-tail rainfall coverage, and the validation of rainfall–damage functions.

Received: 14 Jun 2026 – Discussion started: 26 Jun 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Yuta Kurihara and Mamoru Miyamoto

Status: final response (author comments only)

RC1:
'Comment on egusphere-2026-3482', Anonymous Referee #1, 10 Aug 2026

This paper evaluated uncertainty of rainfall for the impact based flood warning. Figure 6, 7 and 8 show the uncertainty and range of the predicted rainfall, damages and affected populations. These figures are good to prepare the coming disasters. Then, I have a quesition. Is it possible to make these figures for water levels in rivers, flood inundation area and zone in the flood plaine? Before damage and population, many people want to know the river water level and flood inundation situations, uncertainty and ranges. Thanks in advance.

Citation: https://doi.org/10.5194/egusphere-2026-3482-RC1
- AC1: 'Reply on RC1', Yuta Kurihara, 03 Oct 2026
  
  We agree that water level and inundation extent are the quantities most users want to see first, and that they sit naturally between rainfall and damage in the uncertainty chain. We will add them. The rainfall–damage functions in the manuscript are derived from RRI flood simulations run for rainfall scenarios of 50–800 mm (Sect. 2.7, Appendix B). The same simulations provide, for each scenario, the river water level at reference points and the inundation area and inundated zones for each municipality. We will therefore construct seven-day rainfall–water level and rainfall–inundation area relationships in the same way as the damage functions, and propagate the recalibrated seven-day rainfall samples through them. Running RRI directly for every forecast sample is not feasible, because thousands of simulations would be needed per forecast issue; we discuss this in our response to RC3-2.
  
  Citation: https://doi.org/10.5194/egusphere-2026-3482-AC1
RC2:
'Comment on egusphere-2026-3482', Anonymous Referee #2, 15 Sep 2026
General comment
This manuscript presents a clear and well-structured framework for converting multi-centre subseasonal-to-seasonal (S2S) rainfall forecasts into probabilistic flood-impact information for municipalities in the Cagayan River Basin, Philippines. The four-step methodology (forecast-source selection, Bayesian Model Averaging (BMA), seven-day rainfall accumulation and recalibration, and conversion of rainfall forecasts into damage estimates) is logically organized and generally well explained. The use of leave-one-year-out validation, Monte Carlo simulations, and uncertainty analysis is appropriate. The Typhoon Ulysses case study is a useful example of how forecast uncertainty changes as an event approaches. The authors are also appropriately transparent about the limitations of the study, including the short forecast skill horizon, use of basin-mean rainfall, limited representation of extreme rainfall, and lack of independently validated damage functions. Overall, the manuscript is a useful contribution and is close to being suitable for publication. However, several issues should be clarified or addressed before publication.

Main Issues
1. Practical benefit of combining forecast centres
The improvement from combining ECMWF, NCEP, and UKMO forecasts is relatively small. In addition, the Brier Skill Score (BSS) remains negative at the 20 mm threshold. The authors should discuss whether the improvement is statistically meaningful and whether it provides a significant practical benefit.
2. Lack of end-to-end validation
The framework is evaluated separately for rainfall, seven-day rainfall, and damage. However, the complete system is not validated against observed flood impacts. Therefore, the claim that the framework successfully captures "impact uncertainty" is currently supported mainly by the Ulysses case study. The authors should either provide additional validation or make the conclusions about impact prediction more cautious.
3. Different ensemble sizes
The three forecast centres have very different numbers of ensemble members (ECMWF 51, NCEP 16, UKMO 4), but members within each centre are given equal weight. The authors should explain whether this approach could favour or disadvantage particular centres and whether the effect was tested.
4. Seven-day recalibration
The two-parameter recalibration function is fitted using information from three thresholds (100, 150, and 200 mm). Because the 200 mm threshold has very few observations, the more frequent lower thresholds may dominate the fit. The authors should discuss whether this could explain the poor BSS at 200 mm.
5. Damage functions based on very few observations
Some damage functions are fitted using only three rainfall-damage observations. Since the quadratic model has two parameters, such a fit has almost no remaining degrees of freedom, meaning a high R² may not necessarily indicate a reliable relationship. The authors should report how many functions were fitted with only three observations and discuss the uncertainty associated with these estimates.
6. Uncertainty in forecast performance
The CRPS values for the different forecast combinations are very similar. The manuscript identifies the three-centre combination as the best, but it is unclear whether the differences are statistically significant. Confidence intervals or another uncertainty measure for CRPS and BSS would help establish whether the ranking is robust. In addition, individual lead days sometimes favour different forecast combinations. The authors should explain why the pooled performance across all lead days was selected as the main criterion rather than consistency across individual lead days.
7. Small sample sizes in Figure 4
Some points in Figure 4, particularly for the 200 mm threshold and W7–W8 periods, appear to be based on very few observations. The number of forecast-observation pairs behind each point should be indicated so readers can judge the reliability of these results.
8. Typhoon Ulysses case study
The Ulysses example is useful, but it is only one event. Furthermore, the "reference impacts" are based on hydrological/hydraulic model simulations rather than independently observed damage data. Therefore, the results demonstrate that the framework can produce impact distributions, but they do not yet demonstrate that these distributions are skilful predictions of actual impacts. The authors should therefore use more cautious wording when discussing RQ3. If possible, adding another historical flood event would strengthen the analysis.
Minor and Technical Issues
Add the monthly climatology CRPS benchmark (3.06 mm day⁻¹) as a row in Table 2 so it can be directly compared with the other forecast combinations.

Check whether the use of "et al." in the reference list for Vitart et al. (2025) and Zhan et al. (2025) follows the journal's reference style.

Verify that the reported average centre weights (0.34/0.47/0.19 for ECMWF/NCEP/UKMO) are correctly calculated from the lead-day weights in Table A1.

Overall recommendation
The manuscript is well motivated, clearly structured, and potentially valuable for impact-based flood forecasting. The methodology is generally sound, and the authors provide a useful discussion of its limitations. I recommend publication after revision, with particular attention to the statistical significance of forecast improvements, the limited validation of impact predictions, the treatment of ensemble sizes and rare events, and the uncertainty associated with the damage functions.
Citation: https://doi.org/10.5194/egusphere-2026-3482-RC2
- AC2: 'Reply on RC2', Yuta Kurihara, 03 Oct 2026
  
  We thank the referee for the constructive comments. We address each point below.
  Main issues
  RC2-1. Practical benefit of combining forecast centres. The improvement from combining ECMWF, NCEP and UKMO is relatively small, and BSS remains negative at 20 mm. Discuss whether the improvement is statistically meaningful and practically significant.
  AR. We agree that the improvement is small. We will use a block bootstrap over verification years to estimate confidence intervals for CRPS and BSS differences between centre combinations. We will also use the relative-economic-value analysis to test whether the small statistical improvement has practical value. We will state clearly that daily BSS at 20 mm remains negative for all combinations.
  CM. Bootstrap confidence intervals will be added to Table 2 and Sect. 3.1; the practical significance and RQ1 wording will be revised.
  RC2-2. Lack of end-to-end validation. The complete system is not validated against observed flood impacts; the claim of capturing impact uncertainty rests on the Ulysses case study.
  AR. We agree. For Typhoon Ulysses, we will compare the forecast impact distributions with NDRRMC-reported impacts and compare the forecast water-level distribution at Buntun Bridge with the observed hydrograph. RQ3 will be reworded to avoid implying general predictive skill from a single event.
  CM. A new comparison table will be added to Sect. 3.4, and the Abstract, Sect. 4.1 and Conclusion (3) will be revised.
  RC2-3. Different ensemble sizes. Members within each centre are given equal weight although ensemble sizes differ (51/16/4). Could this favour particular centres, and was it tested?
  AR. The BMA formulation already prevents larger ensembles from receiving more total weight simply because they have more members. Members are averaged within each centre by 1/Mc, and the centre weight wc is then estimated by MAP. We will clarify this point in the manuscript.
  CM. A sentence will be added after Eq. (9) explaining that centre weights are independent of ensemble size.
  RC2-4. Seven-day recalibration. The recalibration function is fitted using three thresholds; the more frequent lower thresholds may dominate the fit and explain the poor BSS at 200 mm.
  AR. The 100, 150 and 200 mm thresholds are pooled because only 18 seven-day totals exceeded 200 mm during 2015–2025, making a separate 200 mm fit unstable. We agree that pooling may affect the upper threshold. We will therefore add a threshold-specific sensitivity test and report the result.
  CM. A sensitivity test and discussion will be added to Sects. 3.3 and 4.2.
  RC2-5. Damage functions based on very few observations. Some functions are fitted with only three points; report how many and discuss the uncertainty.
  AR. We agree that the current wording is unclear. L210 does not correctly describe the fitting rule. We will correct it, report the number of above-threshold simulation points used for each municipality and category, and quantify uncertainty for functions based on few points.
  CM. L210 will be corrected; fitting-point counts will be added to Table B1/supplementary material; uncertainty for low-n fits will be discussed in Appendix B and Sect. 4.5.
  RC2-6. Uncertainty in forecast performance. CRPS values are very similar; are the differences significant? Why was pooled performance across lead days used rather than consistency across individual lead days?
  AR. The bootstrap analysis in RC2-1 will test whether the CRPS/BSS differences are distinguishable from zero. We used pooled performance because the subsequent seven-day accumulation uses all lead days, while rankings by individual lead day are less stable. We will explain this choice and report the lead-day robustness.
  CM. The justification will be added to Sect. 2.3 and the lead-day robustness to Sect. 3.1.
  RC2-7. Small sample sizes in Figure 4. Indicate the number of forecast–observation pairs behind each point.
  AR. Agreed. We will add the number of forecast–observation pairs to Fig. 4 and the event counts by threshold/window group to the caption.
  CM. Fig. 4 and its caption will be revised.
  RC2-8. Typhoon Ulysses case study. Only one event; reference impacts come from model simulations rather than observed damage; use more cautious wording for RQ3; if possible add another event.
  AR. We agree. We will compare the Ulysses estimates with NDRRMC-reported impacts and clarify that the RRI reference in Fig. 7 is a calibrated simulation, not observed loss. We will also soften the wording of RQ3 and related sections. We will examine whether a second event, such as Typhoon Mangkhut/Ompong (2018), can be added; if suitable reference data are unavailable, we will state this limitation.
  CM. The wording in the Abstract, RQ3, Sects. 3.4, 4.1, 4.5 and the Conclusions will be revised; a second event will be added if feasible.
  Minor and technical issues
  Table 2 climatology row. We will add the monthly-climatology benchmark (CRPS 3.06 mm day⁻¹) and CRPSS.
  Reference-list style. We will correct the “et al.” entries to follow the journal style.
  Average centre weights. We rechecked the values: 0.34, 0.47 and 0.19 for ECMWF, NCEP and UKMO, consistent with Table A1. We will clarify that these are full-sample estimates and briefly explain why BMA weights need not follow the single-centre skill ranking.
  
  Citation: https://doi.org/10.5194/egusphere-2026-3482-AC2
RC3:
'Comment on egusphere-2026-3482', Anonymous Referee #3, 23 Sep 2026
Review of “Uncertainty of Rainfall Forecasts for Impact-Based Flood Warning in the Cagayan River Basin, Philippines” by Kurihara and Miyamoto. The paper uses statistical post-processing to three individual ensemble forecasts from the S2S archive to propagate rainfall uncertainty into an impact model. The paper is well written and the statistics applied to it are standard procedures. However, the paper fails to show the usefulness of the system in a real-decision making scenario. More work need to be done to compare the framework to existing systems, and explain how it would be applied in a real situation. Further, the skill of the system is very low, and it does not capture extreme events, therefore any real application is very questionable. I recommend that the paper undergoes a major revision to add more analysis to the real usefulness (cost-loss scenarios) and provide a more robust discussion on the actual usefulness of the framework.

Major comments
In the introduction you mention how rainfall information needs to be expressed, but you do not mention co-design of this information with the stakeholders. This is very important to include to make sure that the information is fit for purpose.

In the introduction you mention that no hydrological-hydraulic model was use, why is that? Is this because there is no suitable information to setup such model? As I understand, hydrological models exist for the Luzon river basin. To translate rainfall into damage caused by flooding, a secondary step of using a hydrological model would be advisable, especially of the output is to be used in impact studies. Please explain this further.

The RQ3 as mentioned in the paper assumes that the rainfall-based forecast can be used as a proxy for flood damage. This is questionable since there are many mechanisms that will translate heavy rainfall into a flood, such as timing and intensity. I am not convinced that translating a 7-day rainfall will give you a detailed estimate of the flood risk.

The use of impact-based early-warning is discussed in section 3.4, but it is all very theoretical. Figure 7 is difficult to understand in terms of scale and action needed. Would the decision be to evacuate thousands of people? How would buildings be protected? Is there anything that can be one to protect crops? For actionable information detailed maps are needed to show which areas are affected rather than basin averages.

The case study is useful, but it also shows the flaw of the method, since the information is not available in time for action, the correct signal comes after the event, which is of no use for preparedness at a local scale

The discussion on decision value in section 4.4 is good, but I would suggest complement this with a proper cost-loss scenarios,. This would help the decision makers to prioritize the actions rather than have to make decision directly from the forecasts.

The authors correctly lists all the limitations of the framework, but by doing so the proposed methodology does not seem to be very useful in decision making. There are too many factors that are not considered, there rendering this framework only useful if nothing else exists. Going back to my second point, what are the pre-requisites in this catchment? Is there nothing in place to give warning? If there is, why is existing methodology not included in the framework?

Minor comments
When listing the models used in the study use hyphens instead of plus signs.

You have extracted information from the S2S-database, but you are using the information on a two-week lead time. Technically speaking this is referred to as medium range (up to 15 days), and usually the information after day 15 up to day 45 is referred to as sub-seasonal and longer ranges than that as seasonal. I would clarify that in the text.

Section 2.3. The CRPS is a well-known validation score for ensemble forecasts, but I would suggest using the CRPS skill score instead, as you do for BSS. The value of CRPS itself is difficult to interpret but comparing it with a benchmark forecasts makes the interpretation easier.

L165 7-day averages of rainfall is quite a long averaging, which can mean that you count separate rainfall events which can occur several days apart as one event. This means that you run the risk of over-estimating the danger in flood risk.

Table 2. The BSS shows that any forecast beyond day 4 has no or very little skill since the score is negative or very close to zero, so it should not be used. Please elaborate on this.

L289-291. Thea authors claim that the information can be used for “preparedness decisions”, but this si still very theoretical and I do not see how this information would be used by civil protection authorities.

L355-359. I am not following this line of argument. Human lives will always be prioritized before economical or infrastructure damage, regardless of any framework. What is missing here is a proper risk assessment as to how many lives are potentially lost and where and when actions need to be taken.
Citation: https://doi.org/10.5194/egusphere-2026-3482-RC3
- AC3: 'Reply on RC3', Yuta Kurihara, 03 Oct 2026
  
  We thank the referee for the thoughtful review and for emphasizing the need to demonstrate practical usefulness more clearly. We agree that the manuscript should better explain how the framework could support real-world preparedness decisions and how it relates to existing warning services. The study is methodological in scope, and we will clarify this while adding the analyses and practical interpretation requested below.
  General comment
  RC. The paper fails to show the usefulness of the system in a real decision-making scenario; more work is needed to compare the framework with existing systems and explain how it would be applied; skill is low and extreme events are not captured, so real application is questionable. A major revision adding cost-loss analysis and a more robust discussion of usefulness is recommended.
  AR. We thank the referee for this important comment. We agree that the practical role of the framework and its connection to real decision-making need to be demonstrated more clearly. In the revised manuscript, we will add (i) a relative-economic-value analysis, (ii) municipality-level maps and a probability–action table, (iii) a comparison with the existing flood-warning arrangements in the basin, and (iv) water-level and inundation outputs to make the rainfall-to-impact chain more explicit. We will also revise the discussion and conclusions to state more clearly the lead-time skill horizon and the limitations in representing rare extreme events.
  Major comments
  RC3-1. Co-design with stakeholders. The introduction does not mention co-design of the information with stakeholders.
  AR. We agree that co-design with stakeholders is essential for operational impact-based warning. This study focuses on the methodological development and evaluation of the uncertainty-propagation framework, and stakeholder co-design was not included at this stage. We will clarify this scope and state that the probability thresholds, action mapping, and presentation format should be developed jointly with PAGASA and local DRRM offices in future operational implementation.
  CM. The Introduction and Sect. 4.5 will be revised to clarify that the thresholds, action mapping, and display format are illustrative and require stakeholder co-design.
  RC3-2. Why no hydrological–hydraulic model? Hydrological models exist for Luzon basins; translating rainfall to damage should go through a hydrological model.
  AR. We thank the referee for raising this important point. A hydrological–hydraulic model is included through the RRI simulations used to derive the rainfall–damage relationships, but this role was not sufficiently clear in the manuscript. The RRI simulations cover rainfall scenarios of 50–800 mm and therefore embed the basin's hydrological response. What the current framework does not do is run RRI for every forecast sample, which would require thousands of basin-scale simulations for each forecast issue. In the revision, we will make this hydrological step explicit and add RRI-derived water-level and inundation relationships. We will also note that a surrogate model could enable direct hydrological simulation of many forecast samples in future work.
  CM. L40 and Sects. 2.1 and 2.7 will be revised to make the RRI-based hydrological step explicit; water-level and inundation relationships will be added, and surrogate modelling will be discussed as a future extension in Sect. 4.5.
  RC3-3. Seven-day rainfall as a proxy for flood damage. Many mechanisms (timing, intensity) translate rainfall into a flood; a seven-day total is unlikely to give a detailed estimate of flood risk.
  AR. We agree that a seven-day rainfall total cannot represent all processes controlling flood response, particularly the timing and intensity of rainfall within the seven-day period. The seven-day duration was selected because it is consistent with the prolonged flood response of the Cagayan River Basin and with the design rainfall duration adopted for the basin based on the 2009 Ondoy–Pepeng event (Kurihara et al., 2026, J. Hydrol. Reg. Stud.). The hydrological response is represented through the RRI simulations used to derive the impact relationships, although the temporal rainfall pattern within each seven-day total is fixed. To quantify this limitation, we will add RRI sensitivity simulations using different temporal rainfall patterns for the same seven-day totals and compare the resulting variation in water level and damage with the forecast uncertainty. The added water-level and inundation outputs will also make the intermediate hydrological response more explicit.
  CM. The justification of the seven-day duration will be added to Sect. 2.5, and a rainfall-pattern sensitivity analysis will be added to quantify its effect on water level and damage; the limitation will be discussed in Sect. 4.5.
  RC3-4. Impact-based warning is too theoretical; Fig. 7 is hard to read in terms of scale and action; detailed maps are needed rather than basin averages.
  AR. We agree that the practical interpretation of the current figures should be made clearer. The impact estimates are already calculated for 86 municipalities, although this spatial resolution was not sufficiently evident in the presentation. We will therefore add municipality-level maps of exceedance probability and median impact and revise Fig. 7 for clearer interpretation. We will also add an illustrative probability–action table linking probability and impact levels with staged preparedness actions and responsible agencies. The action thresholds will be presented as illustrative, because operational thresholds should ultimately be defined through stakeholder co-design.
  CM. Municipality-level maps and a probability–action table will be added, Fig. 7 will be revised, and the municipality-level resolution will be clarified in Sect. 2.7.
  RC3-5. The correct signal comes after the event and is of no use for preparedness.
  AR. We appreciate this concern and agree that the timing of a useful signal needs to be demonstrated more clearly. Figure 6 currently emphasises the median, which rose from 79 mm on 2 November to 101 mm on 5 November and 182 mm on 9 November. However, the predictive distribution contains additional information that is not visible from the median alone. We will therefore add the probabilities of exceeding 100 and 150 mm per seven days for each issue date, together with the lead time at which these probabilities first exceeded selected probability levels. For the 5 November issue, six days before the flood, the upper part of the seven-day distribution already extended well above 150 mm. We will clarify that such an early probabilistic signal is intended to support low-cost and reversible preparedness actions, while stronger signals closer to the event can support more resource-intensive actions. We will also state clearly that useful daily forecast skill is mainly limited to approximately lead days 5–6.
  CM. Exceedance probabilities and their lead times will be added to Sect. 3.4, and the distinction between early preparedness actions and later, more resource-intensive actions will be clarified in Sect. 4.3.
  RC3-6. Cost-loss scenarios. Complement Sect. 4.4 with proper cost-loss scenarios.
  AR. We agree and will add a relative-economic-value analysis using the 2015–2025 hindcast. For the 100 and 150 mm per seven-day thresholds, we will evaluate cost–loss ratios from 0.01 to 1 by rolling-window group and compare the raw accumulated BMA probabilities, the recalibrated probabilities, and the monthly climatology. This will show the decision contexts in which the forecast provides value beyond climatology and how that value changes with lead time. We will relate the results to the probability–action table, while noting that action-specific costs and losses would require stakeholder co-design and are beyond the scope of the present study.
  CM. A new relative-economic-value analysis and figure will be added to Sect. 3, and Sect. 4.4 will be revised accordingly.
  RC3-7. Too many limitations; the framework seems useful only if nothing else exists; what warning exists in the catchment and why is it not included?
  AR. We agree that the relationship with the existing warning system should be explained more clearly. PAGASA currently provides basin flood forecasting and warning for the Cagayan River using observed rainfall, river water levels, and dam-release information, mainly at lead times of hours to about a day. The proposed framework is intended to complement this service by providing probabilistic, municipality-level impact information several days earlier, particularly for preparedness actions that can be taken before short-range hydrological warnings become available. We will add a description of the existing warning arrangements and a comparison of their inputs, lead times, outputs, and intended uses, together with an example of how the two types of information could be used sequentially.
  CM. A description and comparison of the existing warning system and the proposed framework will be added, and Sect. 4.5 will be restructured to clarify the complementary role of the framework before discussing its limitations.
  Minor comments
  1. Plus signs. Agreed. We will replace plus signs with hyphens for model combinations throughout.
  2. Terminology of forecast range. Agreed. We will use “medium-range ensemble forecasts obtained from the S2S database” throughout the manuscript.
  3. CRPSS. Agreed. We will add CRPSS relative to the monthly climatology to Table 2 and Fig. 2 and report it together with CRPS. For the three-centre combination, the pooled CRPSS over lead days 1–14 is approximately 0.016.
  4. Seven-day averaging may merge separate events. We agree that the within-window rainfall distribution may affect the hydrological response. This will be addressed through the rainfall-pattern sensitivity analysis described under RC3-3 and discussed in Sects. 2.5 and 4.5.
  5. BSS beyond lead day 4 is near zero or negative, so those forecasts should not be used. We agree that forecast skill decreases substantially with lead time and that this limitation should be made clearer. The BSS shown in Table 2 and Fig. 2 evaluates daily rainfall exceeding 20 mm and is used for forecast-source screening, whereas the preparedness product is the recalibrated seven-day exceedance probability. For this product, BSS remains positive at 100 mm for W1–W5 and non-negative at 150 mm for W1–W5. At longer leads, the reduced skill is reflected in probabilities closer to climatology and wider predictive intervals. We will clarify this distinction and use the relative-economic-value analysis to identify the lead times and types of preparedness actions for which the forecasts retain practical value.
  6. L289–291, use for preparedness decisions is theoretical. We agree that the current wording may imply a level of operational implementation that has not yet been demonstrated. We will revise the text to state that the framework provides impact information at a decision-relevant spatial scale, while the specific use of this information in operational preparedness requires stakeholder co-design. The probability–action table and relative-economic-value analysis will be presented as illustrations of potential use rather than as an established operational procedure.
  7. L355–359. We appreciate this comment and agree that the current wording could be misinterpreted. Our intention is to show that the four impact categories support different types of preparedness actions and involve different responsible agencies. We will revise the text accordingly. We will also clarify that “affected population” refers to the population within the simulated inundation area and is not an estimate of casualties; fatality risk assessment is outside the scope of this study.
  CM for minor comments. Items 1–3 and 7 are editorial and will be applied throughout; items 4–6 are covered by the changes listed under RC3-3, RC3-4, RC3-5 and RC3-6.
  
  Citation: https://doi.org/10.5194/egusphere-2026-3482-AC3

Yuta Kurihara and Mamoru Miyamoto

Viewed

Total article views: 305 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
151	109	45	305	20	21

HTML: 151
PDF: 109
XML: 45
Total: 305
BibTeX: 20
EndNote: 21

Views and downloads (calculated since 26 Jun 2026)

Month	HTML	PDF	XML	Total
Jul 2026	19	19	6	44
Aug 2026	23	18	4	45
Sep 2026	72	46	21	139
Oct 2026	37	26	14	77

Cumulative views and downloads (calculated since 26 Jun 2026)

Month	HTML	PDF	XML	Total
Jul 2026	19	19	6	44
Aug 2026	23	18	4	45
Sep 2026	72	46	21	139
Oct 2026	37	26	14	77

Viewed (geographical distribution)

Total article views: 274 (including HTML, PDF, and XML) Thereof 274 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 16 Oct 2026

Short summary

Showing uncertainty in rainfall and damage is important for better decision-making. We developed a method that turns longer-range rainfall forecasts into estimates of possible impacts for each municipality in the Cagayan River Basin, Philippines. By comparing forecasts with observations from 2015 to 2025, we adjusted forecast reliability, then demonstrated the method for Typhoon Ulysses. The results can help stakeholders move more quickly toward flood preparedness.


Total:	0
HTML:	0
PDF:	0
XML:	0