the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 15 Sep 2026
Beyond the flood map: an open-source framework for forecasting cascading impacts of natural hazards on people and essential services
Abstract. Early warning and anticipatory action can help reduce negative consequences from weather and climate extremes. To inform preemptive decisions and anticipatory actions, actors in disaster risk reduction such as humanitarian organisations are increasingly relying on impact forecasts, as these can provide more concrete and actionable information in comparison to hazard-based forecasts. However, impact forecasts commonly focus on direct impacts, leaving indirect impacts such as service accessibility, which are relevant for emergency response planning, too often unaccounted for. To address this gap, we develop an impact forecasting framework to predict disruption in service access resulting from weather extremes. We illustrate the framework by predicting healthcare access disruptions for two high-impact flood events that impacted Somalia and Sudan in 2023 and 2024. Our model is able to provide actionable and spatially-explicit information such as forecast maps of where people may experience loss of service accessibility, and our results show that the pipeline is able to capture observed impacts up to a week in advance. Additionally, including indirect impacts such as healthcare access disruption significantly increases the estimated scale of a disaster. Comparison of modelling results with ground reports, satellite observations, and survey data reveals that observed post-disaster service accessibility disruption rates can be reproduced. Limitations remain in the precision of flood forecast input data and incompleteness of exposure data. An uncertainty and sensitivity analysis reveals large regional discrepancies in the sensitivity to the model parameters, and that uncertainty in exposure and accessibility thresholds can exceed the hazard forecast uncertainty. Our results illustrate how a model built on open-source data can provide decision-relevant and people-centric impact forecasts, that can be used to inform anticipatory action. We expect our findings to help improving impact forecast models and anticipatory action frameworks aimed at reducing human suffering.
- Preprint
(21038 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 27 Oct 2026)
- RC1: 'Comment on egusphere-2026-5363', Anonymous Referee #1, 22 Sep 2026 reply
-
RC2: 'Comment on egusphere-2026-5363', Anonymous Referee #2, 29 Sep 2026
reply
General comments
I enjoyed reading the article by Severino et al. which presents an open-source framework for forecasting cascading impacts from riverine floods, with a focus on impacts on healthcare access using case studies in East Africa.
The topic is of great interest, given the increasing trend to focus early warning systems on disaster impacts, rather than just hazard magnitude. Overall the article is well written, though the grammar and English wording would benefit from a review by an English native speaker. Models and data used are mostly open source, which is appreciated. I think that the proposed framework has strong potential to improve the state of the art in flood impact forecasting, provided that some improvements are made on the current version. Main concerns about this work are detailed below, followed by a list of more specific comments.
First, the original contribution of the article, that is the estimation of cascading impacts from riverine flooding, is somewhat diluted within other analyses that are less prioritary for this work. I’m not sure why the authors put so much focus on validating GloFAS forecasts versus satellite based estimates, which is certainly of interest for GloFAS developers but less for this work. The meteorological and hydrological model uncertainty plays a big role in the evaluation and these can be excluded if the evaluation is made directly using satellite derived inundation extent as input and compared it with observed data for the events.
Second, there are a number of assumptions made that have a considerable influence on the model results and weaken the approach. While I appreciate the inclusion of a sensitivity analysis to assess the different assumptions made, some choices should be made more subjectively. For example, by default I think that the median (50%) of the ensemble should be taken, rather than the 90th percentile. Anyways, this would not be needed in case the satellite based inundation map is taken as input.
Third point is related to the model resolution, where for flood impact assessment is very important to use the highest available. Why did the authors downscaled the resolution of the inundation maps and of the population grids? Both data are available from their respective source at 90m resolution, hence it would be important to run the assessment at the same resolution, and given the limited size of the two case studies it should not be a problem from the computational side.
Fourth, the choice of the figures and their content. Visually they are all nice looking, but from the practical aspects of presenting results, some of them (Figure 4, 5, 7, A1, B1, C1, D1, H1, I1, J1) are a bit too crowded with information, especially color shades, and difficult to understand in all their aspects. Some thinking should be made to try to improve the presentation of results.
Specific comments
L95: Perhaps here you could motivate the regridding from 3 to 20 arc seconds
L104: 2 km resolution is rather coarse for impact analysis, given the high resolution needed to represent flood scenarios. How this assumption can increase the uncertainty of the model output?
L107-109: The choice of a 50 cm threshold would benefit from some supporting reasoning. Is it also used to bias correct the model output? I find it quite high, as even just 30-40 cm seems enough to cause disruption to the population and services.
L151-153: FLOPROS is a dataset of flood protection standards. Please clarify the meaning of “Uncertainty associated with flood adaptation is approximated using the FLOPROS dataset”
L187: International Organization for Migration IOM appears twice
Figure 2: The figure caption should be expanded with more details. Also, there should be a reference to each figure in the text (before the actual figure).
Figure 3: The caption should clearly state that these time series are model output. Are these figures based on reanalysis or forecasts?
L233-236: The part of the triggers should be clarified. I have the impression that these triggers were set a posteriori, looking at when affected population experienced sharp increases. However no forecast data was mentioned so far, so the process leading to “response funds payout based on predefined triggers to be released earlier” is not clear at this point.
Section 3.1: Section names read as if they were article titles. I suggest turning them to more standard and concise versions.
Figure 6: “ Fourth row shows the exceedance probabilities”: there are only 3 rows, please correct the caption. Also, the 5% threshold is not shown, the caption must be changed to reflect it. I suggest explicitly naming the 2 columns at the top to make it clearer. On the conceptual side, I find the panels with the probability of return period exceedance not easy to read and of little value for the presented work, considering that the FLOPROS dataset is already used to include the information on the level of flood protection.
L331: F1-score, precision and recall are widely used performance scores. However, a brief description including how to interpret them should be added in the methods section.
Figure 7: I find the overlaying of 3 different flood extents not easy to read. Perhaps I’d split each of the two left panels in one panel with only modeled flood extent (forecast and reanalysis) and another panel for satellite flood extent.
L366-367: from figure 8, indirectly affected people seem similar to those directly affected. I don’t see the ratio 500k to 1M mentioned in the text.
L398: for or four? Double check
L421-422: This sentence reads like a general model result. It should be clarified that this result is specific of the (2) events shown. Also, I found it a bit misleading as this reflects purely the performance of the chosen hydrological forecasting model, rather than that of the proposed framework especially for indirect impact assessment.
L433: Suggested “...in Sudan by 40%”
Citation: https://doi.org/10.5194/egusphere-2026-5363-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 217 | 112 | 33 | 362 | 33 | 31 |
- HTML: 217
- PDF: 112
- XML: 33
- Total: 362
- BibTeX: 33
- EndNote: 31
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This paper details an approach to estimate indirect impacts from flooding: specifically loss of access to health centres due to inundation. Clearly the work is relevant to improving specific information for advance protection of communities as well as response. As the authors show however, integration of hazard, exposure and vulnerability also compounds uncertainties, such that in the studied contexts (Sudan and Somalia) even aside from the difficulty forecasting flood hazard, there is substantial uncertainty even in locating health centres (as an aside it suggests that the most fruitful path to reducing uncertainty may be generating up-to-date information on the location of exposed elements - which in principle is simpler than the uncertainty associated with predicting convective rainfall).
The paper is well written and structured and I have no significant concerns to address before publication. My only major comment is on the interpretation of the P(number of people indirectly affected >threshold) being above P(# directly affected>threshold) which in turn being above P(discharge>return period). i.e. figure 6 and associated discussion. It is implied that translation to indirect impacts enhance trigger decisions (e.g. title of section 3.1.3). However I am not sure this is the case. Fig 6 shows that the probabilities of indirect /direct impacts reach much higher levels than discharge return period forecasts, with the latter "too weak to trigger anticipatory action" (l. 316). Yet the critical value of a forecast to triggering a decision is in its discrimination. That is: high probabilities before 'events', but also low probabilities (ideally, mostly when events do not occur).
In this case it's my understanding that this critical discriminating information is coming entirely from the GloFAS forecast (and tracking it back, from the rainfall forecast ultimately). Everything else in the workflow provides more detail. In the case of comparing probabilities of discharge level and probabilities of impacts, is it not just a non-linear scaling? i.e. instances of higher likelihood of discharge will presumably give higher likelihood of impacts. Put another way, the fact that higher probabilities found after translating to impacts is arbitrary and depends on threshold used to determine # people (indeed "at least 1 person affected" in a flood prone region is quite a low bar to compare with a N-year return period event).
All of that is to say that I believe the impact layer will not actually enhance the trigger decision itself by improving discrimination of the model to advise on event occurrence between forecasting instances. Indeed this cannot be checked directly since the authors only consider one season per location - a more systematic look at generating and evaluating an impact 'hindcast' would be need for this (which I appreciate is out of scope but might be worth considering in future).
Having said that I do believe there is clear value in this method for decision-making. Although I am not convinced it enhances the trigger decision itself, I believe the value lies in the enhancement of the information available for planning, e.g. mapping populations likely to be cut off from services given a specific forecast.
Before publication I'd welcome the authors take on this point, and where relevant to incorporate it into the discussion.
Otherwise, a few minor points below:
l21 First paragraph of introduction is long, can split it somewhere
l66 This final paragraph is a little repetitive of what came before: could instead be used to signpost the structure of the paper
l80 "spatially-explicitly" can be rephrased more simply without a novel hyphenation
l87 please note which GloFAS version? v4 had serious problems in Somalia in 2023
l88 please give a quick explainer of what Dottori provide (is it satellite or model based - presumably inundation?). Also more explicit detail of the 'model' of Riedel would be appreciated, how do the combine stream flow with flood hazard maps?
l95 worth mentioning here what lead time is used
l100 is 'residential level' a defined term in OSM? And, do you filter out so that these are not included, or you this is the lowest level included? Also 'filter out' doesn't need a hyphen).
l104 Would welcome an overall comment on spatial resolution - populations scaled to 2km yet the flood data is 20 arc-seconds. At what resolution are the layers when they're actually interacting with each other, and what is the impact of losing the population resolution by up-scaling?
l108 Can the threshold function be clarified, presumably a binary filter where anything over 50cm is considered affected, otherwise not?
l116 'inter-dependent' is said several times in a few lines. Consider a simpler rephrasing, especially since the dependencies are explicitly detailed, hence introducing them as 'inter-dependent' is redundant.
l136 "These impact forecasts are..." I didn't quite follow this step, consider being more explicit.
l152 please describe what FLOPROS is
l153 "medium-range" is a fuzzy definition but more generally describes forecast lead time beyond one week. Most forecasters would probably consider 5 days belonging to the weather timescale.
l212 "170 thousand" check journal style, likely better as 170,000
Figure 2 please add subfigure labels and corresponding caption description. Currently also assumes readers can tell the difference between Sudan and Somalia on a map! Also the top left plot shows the two countries floating in the same box - this could be interpreted as representing an exact geographical position, which it is not. Suggest separate boxes for these. Finally, the inundation data on the map (source, date?) is not identified in the caption.
l227 Simulated "directly affected" is compared with OCHA-reported "total affected" as a way of verifying the realism of the model. I'm not sure this is a reasonable comparison point however - simulations consider directly affected if they are standing in at least 50cm of water: is this comparable to the way OCHA is counting? Or is OCHA inclusive of people who are indirectly affected as well?
Figure 3 x axis labels might be clearer to show just 1st of months and leave out the year (1-Jul, 1-Aug etc)
l228 696,000 being 'slightly below' "about a million" is a favourable reading of the data. By eye the total affected number is closer to 1.2m which makes the simulation "almost twice" the observed estimate.
Fig 4, 5 can be made bigger, perhaps put in a two panel figure
Fig 6b,d includes "additional indirect impacts" in the legend but no data on the plot, please check.
Section 3.2 The result of large underestimate of actual impacts for Somalia chimes with experience of poor GloFAS performance as noted in the paper. In fact this was specifically due to GloFAS v4 which entirely failed to flooding on the Shabelle at peak flood time. Post-event analysis by the GloFAS team showed this was not a 'forecast bust' (i.e. the heavy rainfall in the basin was correctly anticipated) but was somehow related to the soil simulation in the model - indeed even forced by real rainfall observations it would fail to produce a flood peak for 2023 Deyr and even the record-breaking 1997 Deyr. Unfortunately I do not know of a reference where this has been documented, and as far as I am aware the problem is still unfixed. No particular suggestion for an edit here as there's nothing citable, but just an FYI.
Figure 9 might be visually clarified: currently the different bars are distinguished both by labels (Access loss, access uninterrupted, no base access) as well as encoded in colour. To make it clearer just one of these might be chosen. Also, including region names along with the sub-figure labels would be helpful.
l425 "impact forecasts are able to provide a clearer predictive signal" see main comment above. Is it really 'clearer' in the sense of true discrimination (which you have not technically shown, having only included one flooding season per region) or, is it just "a larger probability" which scales with the strength of the discharge forecast signal. The latter would end up issuing high probabilities frequently, which does not in fact clarify the signal.