the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 02 Oct 2025
Forecasting people exposed to tropical cyclone flooding in Southeast Africa: Lessons learned from recent events
Abstract. East Africa has recently experienced a series of devastating tropical cyclone landfalls characterised by hundreds of fatalities, millions of displaced people and substantial economic damage. Forecasting the impact of these tropical cyclones can, in theory, better motivate anticipatory action compared to only forecasting the hazard. This paper describes an approach to forecasting the number of people directly exposed to flooding from tropical cyclones and documents experience gained communicating these forecasts to practitioners via emergency bulletins.
Forecasting flood exposure requires a complex cascade of meteorological, hydrological, hydraulic and population models. Interpretation of forecasts was difficult, even for the scientific experts developing the systems, due to uncertainties brought in at each stage of the modelling cascade. Thus, producing interpretable forecast messaging was challenging and often required extensive discussion between forecasters with expertise on different elements of the system. This paper uses practical experience gained from several tropical cyclone events to highlight essential requirements for interpreting and disseminating tropical cyclone flood impact forecasts. We also analyse how forecasts evolved with lead-time and compare them to observed flooding in the case of Tropical Cyclone Freddy. Overall, we aim to synthesise our experience into actionable learning that might inform future use of forecasting in humanitarian response.
Exposure estimates were most sensitive to storm track location, even when exposure was aggregated to districts. Uncertainty from track location remained substantial even in the days before landfall, meaning a recipient of these forecasts needs to understand and interpret the distribution of exposure. For the second landfall of Tropical Cyclone Freddy, nationwide exposure estimates were remarkably similar between remotely sensed flood extents and the best estimate from the forecast system. However, this overall similarity results from the averaging of substantial uncertainty at the district scale.
- Preprint
(2789 KB) - Metadata XML
-
Supplement
(14924 KB) - BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2025-3473', Anonymous Referee #1, 06 Nov 2025
-
RC2: 'Comment on egusphere-2025-3473', Anonymous Referee #2, 04 Jan 2026
The study drew on experiences from multiple cyclone events in Southeast Africa to forecast the number of people exposed to these flood events, and highlight challenges in interpreting and communicating forecasts due to substantial uncertainty across complex hydrologic and hydraulic modeling processes. Overall, this study is meaningful for more informed flood risk management, at least for the study area. However, I still have several comments and suggestions as follows.
1) The reasons why Southeast Africa was selected as the study area should be elaborated in more detail. The lessons learned from the cyclone flooding events may be only applicable to the specific regions of Southeast Africa, and generalized guidelines beyond the study area are suggested to be discussed.
2) It is suggested to add a list of acronyms mentioned in the manuscript. The full term of the acronym is only presented the first time it appears, e.g., GloFAS. Some acronyms seem to be unnecessary, e.g., JRC.
3) Line 139: Why were different versions of GloFAS used for different events? Would this lead to another uncertainty source about streamflow estimates?
4) Line 172: Is “WorldPop population data (population counts)” a section title?
5) Table 1: Five or six events were investigated? Six or seven landfalls were investigated? Please indicate the format of the date, e.g., DD/MM/YYYY.
6) Figures 2-3: What is the unit for population? It is suggested to improve the quality of most of the figures in the manuscript, e.g., keep the labels clear and the font size consistent, and the sub-figure title is overlapped with the vertical label in Figure 4.
7) Lines 230-231: The statement that “these outliers are typically driven by ensemble members that cause flooding over areas of high population, rather than being ensemble members with exceptionally intense rainfalls” should be supported by some evidence.
8) Lines 246-247: Considering that the ensemble mean would be sensitive to outliers, why not use ensemble median for the further analysis?
9) Figure 4: Why is the lead time not the same among the seven landfalls, 5~-1 days and 4~-1 days? What do different colors stand for in the lower plots? Maybe a typo: “Gombe Malwai” or “Gombe Malawi”?
10) Figure 5: Figures 5a-5d are for the total precipitation >=200mm, while Figures 5e-5f are for the total precipitation >=300mm?
11) Line 278 and Figure 5c: 30th February is not true.
12) Figure 6: It would be helpful to present the population and the percentage of urbanization for each district?
13) Table 3: The equation used for calculating the “Diff %” may not be the same, e.g., those for “Other (<1000 exposed)”. It would make more sense if the max exposure from observed extents is used as the denominator.
14) Lines 404-405: Given the various uncertainty sources that were not considered in the case study, the findings may be subject to change after more comprehensive analyses are conducted. Also, the uncertainty in model evaluation should not be ignored. The authors can refer to the article below for more information about the limitations of some commonly used evaluation metrics in flood models.
Reference: “Beyond a fixed number: Investigating uncertainty in popular evaluation metrics of ensemble flood modeling using bootstrapping analysis” (https://doi.org/10.1111/jfr3.12982)
15) Lines 429-430: It would be better to quantify how large or how small the district is for the corresponding conclusion.Citation: https://doi.org/10.5194/egusphere-2025-3473-RC2
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 348 | 110 | 25 | 483 | 48 | 22 | 19 |
- HTML: 348
- PDF: 110
- XML: 25
- Total: 483
- Supplement: 48
- BibTeX: 22
- EndNote: 19
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Summary
The paper presents a methodology for forecasting the number of people directly exposed to flooding from tropical cyclones in Southeast Africa, aiming to improve anticipatory humanitarian action. It combines meteorological (ECMWF), hydrological (GloFAS), hydraulic (Fathom flood maps), and population (WorldPop) models to create forecasts of flood exposure. The study highlights both the potential and limitations of forecasting flood exposure for tropical cyclones, as well as the experience gained in communicating these forecasts. As such, it is a valuable contribution to the field of disaster risk management. However, the introduction and discussion sections should better position the study within existing literature. Furthermore, important choices made in the model chain should be clarified. Therefore, I recommend a major revision before publication.
Major comments
Introduction: The introduction should include references to other studies on operational forecasting of tropical cyclone (TC) flood hazard and/or exposure, including those that dynamically run hydrodynamic or surrogate models. This would help better position the current study within the context of existing literature
L198–206: More explanation is needed to understand how the preprocessed flood inundation maps are combined with the GloFAS discharge data. Is a single return period map selected for the entire domain or per sub-basin? How are water levels interpolated between pre-calculated return period-based maps? How are GloFAS discharge forecasts linked to the Fathom flood maps? Is a flood depth threshold used to determine "people flooded"? How is flood protection (e.g., levees or raised houses/thresholds) considered in this process?
Minor comments
L68: "resolution well below 100m is necessary" — This paragraph focuses strongly on spatial resolution but does not mention vertical accuracy, which is equally important. I suggest discussing the required vertical accuracy in this paragraph as well.
L172: "Worldpop population data". This should probably be a new subsection.
Figure 3: The text and legend below the figures are very small and difficult to read. Consider increasing the font size.
L223: "We stop one day after landfall .." Is this standard SOP or specific to the case of Freddy? For TC Idai for example, for example, heavy rainfall and flooding continued for several days after landfall.
L240: When does a highly uncertain forecast become "misinformative"? Here you state that "the forecasts tend not to be misinformative" but in the discussion around L419 that "Forecasters should consider the possibility that exposure forecasts could be misinformative". Can you elaborate on this?
Figure 4: To clarify, do the colors in Figure 4 refer to the same district across each forecast lead time? If so, it would be helpful to state this explicitly in the figure caption and the text around L247. It would also be interesting to assess whether the top five districts remain stable across lead times or if there are large changes in ranking or top-5 districts.
L329: "In general, the remote sensing is likely to have missed substantial pluvial and flash flooding on smaller rivers." — Please add a reference to support this statement.
L456: "A like-for-like comparison of the inundation modelling and remote sensing was highly problematic in this context because the inundation event is not simulated dynamically". In the results section (around L358) the main reason for not comparing the modelled and observed flood extents is based on limitations of the satellite data, while here in the discussion it is based on limitations of the model. It's probably both, but it would be good provide a balanced and consistent argument in both sections.
L473: Please consider adding literature from other studies that have done dynamic operational flood forecasting for TCs, e.g. the COSMOS system in the USA.
L475: "This may i) prove to be computationally expensive if 30m resolution is maintained," Please consider adding literature from other studies that use surrogate models or subgrid approaches to balance model speed and accuracy. And how important is a 30m resolution model in the context of other uncertainties in the flood model chain such as the use of a hazard map lookup system, the uncertainty in the DTM, or the uncertainty in the hydrological model simulations?